Paper deep dive
Robotic Foundation Models for Industrial Control: A Comprehensive Survey and Readiness Assessment Framework
David Kube, Simon Hadwiger, Tobias Meisen
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 94%
Last extracted: 3/13/2026, 12:17:09 AM
Summary
This paper provides a comprehensive survey and readiness assessment framework for Robotic Foundation Models (RFMs) in industrial control. It categorizes robotic control methods, defines RFMs, and evaluates 324 manipulation-capable models against 149 industrial criteria, concluding that current industrial maturity is limited and uneven.
Entities (4)
Relation Signals (2)
Vision-Language-Action Models → isa → Robotic Foundation Models
confidence 95% · VLAs form a prominent and distinct sub-category within the broader class of RFMs
Robotic Foundation Models → usedin → Industrial Robotics
confidence 90% · This survey gives an extensive overview over the RFM-landscape and analyses... how industrial domains and use cases shape the requirements of RFMs
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Robotic foundation models (RFMs) are emerging as a promising route towards flexible, instruction- and demonstration-driven robot control, however, a critical investigation of their industrial applicability is still lacking. This survey gives an extensive overview over the RFM-landscape and analyses, driven by concrete implications, how industrial domains and use cases shape the requirements of RFMs, with particular focus on collaborative robot platforms, heterogeneous sensing and actuation, edge-computing constraints, and safety-critical operation. We synthesise industrial deployment perspectives into eleven interdependent implications and operationalise them into an assessment framework comprising a catalogue of 149 concrete criteria, spanning both model capabilities and ecosystem requirements. Using this framework, we evaluate 324 manipulation-capable RFMs via 48,276 criterion-level decisions obtained via a conservative LLM-assisted evaluation pipeline, validated against expert judgements. The results indicate that industrial maturity is limited and uneven: even the highest-rated models satisfy only a fraction of criteria and typically exhibit narrow implication-specific peaks rather than integrated coverage. We conclude that progress towards industry-grade RFMs depends less on isolated benchmark successes than on systematic incorporation of safety, real-time feasibility, robust perception, interaction, and cost-effective system integration into auditable deployment stacks.
Tags
Links
- Source: https://arxiv.org/abs/2603.06749v1
- Canonical: https://arxiv.org/abs/2603.06749v1
Trouble viewing inline? Open PDF directly →
Full Text
347,644 characters extracted from source content.
Expand or collapse full text
Robotic Foundation Models for Industrial Control: A Comprehensive Survey and Readiness Assessment Framework David Kube 1,2* , Simon Hadwiger 1,2 and Tobias Meisen 2 1* Factory Automation, Siemens AG, Gleiwitzer Straße 555, Nuremberg, 90475, Germany. 2 Institute for Technologies and Management of Digital Transformation, Bergische Universit ̈at Wuppertal, Gaußstraße 20, Wuppertal, 42119, Germany. *Corresponding author(s). E-mail(s): david.kube@siemens.com; Contributing authors: simon.hadwiger@siemens.com; meisen@uni-wuppertal.de; Abstract Robotic foundation models (RFMs) are emerging as a promising route towards flexible, instruction- and demonstration-driven robot control, however, a critical investigation of their industrial applica- bility is still lacking. This survey gives an extensive overview over the RFM-landscape and analyses, driven by concrete implications, how industrial domains and use cases shape the requirements of RFMs, with particular focus on collaborative robot platforms, heterogeneous sensing and actua- tion, edge-computing constraints, and safety-critical operation. We synthesise industrial deployment perspectives into eleven interdependent implications and operationalise them into an assessment framework comprising a catalogue of 149 concrete criteria, spanning both model capabilities and ecosystem requirements. Using this framework, we evaluate 324 manipulation-capable RFMs via 48,276 criterion-level decisions obtained via a conservative LLM-assisted evaluation pipeline, validated against expert judgements. The results indicate that industrial maturity is limited and uneven: even the highest-rated models satisfy only a fraction of criteria and typically exhibit narrow implication- specific peaks rather than integrated coverage. We conclude that progress towards industry-grade RFMs depends less on isolated benchmark successes than on systematic incorporation of safety, real- time feasibility, robust perception, interaction, and cost-effective system integration into auditable deployment stacks. Keywords: Robotic Foundation Models, Industrial Robotics, Industrial Maturity Assessment, Robot Manipulation, Industrial AI, Collaborative Robotics 1 Introduction Over the past decades, industrial robotics is undergoing a transition from isolated, highly opti- mised automation towards flexible and increas- ingly collaborative deployment paradigms. In par- ticular, the growing adoption of Collaborative Robots (cobots) reflects a shift towards automa- tion solutions that can be commissioned rapidly, redeployed across frequently changing processes, and operated in closer proximity to human work- ers. While these platforms offer high mechanical precision and repeatability alongside compara- tively restrained power consumption, their indus- trial utility is increasingly limited by the availabil- ity of robust, adaptive and intuitive control meth- ods rather than by actuation performance alone. This motivates a renewed focus on learning-based 1 arXiv:2603.06749v1 [cs.RO] 6 Mar 2026 robot intelligence that can cope with variability in tasks, environments, actuation and sensing, while respecting industrial constraints such as safety, reliability, low-latency operation and cost-effective integration. Recent progress in large-scale learn- ing has led to the emergence of Foundation Models (FMs), which promise to reduce manual engi- neering effort by transferring knowledge across tasks – a capability particularly attractive in high- mix, low-volume industrial settings. Within this development, Robotic Foundation Models (RFMs) aim to supply generalist competence for embodied decision-making by providing a flexible platform for transfers across tasks, interfaces and embodi- ments. In this survey, we focus on models capable of generating low-level actions for direct robot actuation – RFMs for control or integrated RFMs as per our definition in Section 2 – because this interface is central for real-world deployment and exposes the strongest coupling to indus- trial constraints (e.g. latency, safety gating, and hardware-specific action spaces). While paradigms for building such models vary widely – including diffusion-based approaches [1, 2], dreaming-based methods [3, 4], and dual-system designs [5–7] – we deliberately chose not to review architectural taxonomy, as related works already cover those topics; for architectural-centred surveys, we thus refer to Section 4. Instead, we use the industrial lens to connect the rapid RFMs’ methodological progress to the practical constraints that deter- mine whether such systems can be deployed in industrial environments. To structure this discus- sion and to position RFMs within the broader evolution of robot control, we propose a compre- hensible categorisation of robotic control methods, summarised by the hierarchy in Figure 1: 1. Predefined/Rule-based: Employ fixed tra- jectories or simple triggers, lacking dynamic adaptation. Transition: Upgraded by integrat- ing sensor-driven, dynamic input. 2. Reactive/Parametrised:Enablelimited adaptation via sensors and parameter adjust- ment. Transition: Move beyond predefined behaviours by introducing learning capabilities. 3. Specialised/Narrow: Solutions tailored for a specific problem on a given hardware setup. Transition: Achieve generic, multi-purpose competence. 1. Predefined/ Rule-Based 2. Reactive/ Parametrised 3. Specialised/ Narrow 4. Foundational 5. Generalist proprio-/exterioception learned behaviour multi-purpose foundation zero/ few-shot & online adaptation Fig. 1: Robotic control-method hierarchy: Each level increases flexibility, intelligence and gener- alisation capabilities, while reducing manual engi- neering effort and required expertise for operation and adaptation. 4. Foundational: Broadly pretrained systems capable of efficient fine-tuning with minimal dedicated engineering. Transition: Eliminate need for further manual tuning and enable online adaptation. 5. Generalist: Deliver zero-shot or in-context generalisation to new tasks and configurations, requiring only high-level instructions or online demonstration, akin to human operators in industry. This hierarchy highlights the core developmental trajectory relevant to industrial deployment: mov- ing from rigid, task-specific automation towards adaptable, instruction- and demonstration-driven operation with decreasing commissioning effort. The introduction of FMs in robotics expressively marked the progression from specialised control systems (3.) to foundational robotics (4.), pro- ducing cross-embodiment and multi-task control systems. Yet, as our subsequent analysis shows, this step does not automatically yield industry- grade autonomy: reaching the next level towards truly generalist systems (5.) requires capabili- ties such as robust online adaptation, dependable instruction following, and safe operation under uncertainty – capabilities that are particularly important in dynamic, cost-driven industrial envi- ronments. The pace at which the field is moving makes a dedicated, industrially grounded review a necessity. As shown in Figure 2, publications rel- 2 202020212022202320242025 Year 0 100 200 300 400 500 600 #Publications Publications/Year RFMs Related High-Level Data/Bench Surveys Fig. 2: Publication increase. Visualising the amount of publications gathered per category. evant to RFMs have increased sharply in recent years, with the most pronounced rise occurring in 2025. This rapid growth risks fragmentation and inflated claims without structured industrial eval- uation. A similar growth trend is visible in related literature addressing enabling components, appli- cations, and evaluation. While several surveys already exist (see Section 4), none provides a systematic, implication-grounded assessment of industrial deployment readiness of RFMs. At the same time, the rapid expansion of the literature implies that infrequent snapshots quickly become outdated, motivating a review that both consoli- dates the current RFM landscape and provides an assessment structure that remains useful as new models appear. Accordingly, this survey is structured to progres- sively connect foundational concepts to industrial requirements and, subsequently, to a systematic assessment of current RFMs (cf. Figure 3): We first define RFMs and position them within the robotics control landscape (Section 2). We then describe our review methodology and corpora composition (Section 3), followed by related works (Section 4) and further background to motivate industrial relevance: the rise of cobots and their deployment domains (Section 5) as well as the practical constraints of hardware, sensing, and edge computing that shape feasible RFM deploy- ment (Section 6). 1. Introduction 2. Robotic Foundation Models (RFMs) 4. Related Works 9. Summary of Findings 3.1 Literature Acquisition Procedure 3.2 RFM-Related Corpora 3.3 Industrial Implication Corpus 3.4 Further Corpora & Manual Searches 5.1 Landscape of Industrial Robotics 5.2 Productive Cobot Applications 5.3 Drivers Toward Cobots 6.1 Cobot Capabilities & Power Consumption 5.4 Toward Intelligent & Collaborative Automation 6.2 Sensing & Manipulation Hardware 6.3 Industrial Edge PCs 7.1 General Industrial Implications 7.2 Attributes of Industry-Grade RFMs 7.3 Evaluating Industrial Maturity: A Criteria Catalogue 8.1 Evaluation Procedure 8.2 Highest Rated RFMs 8.3 Industrial Applicability Overview 6. Cobot & Compute Hardware 7. Industrial Implications 8. RFM Overview & Evaluation 3. Review Methodology & Corpus Composition 5. Rise of Cobots in Industry Fig. 3: Table of Contents. Highlighting prelimi- nary sections in red, background in purple and the main review in green. Building on these foundations, we distil indus- trial deployment viewpoints into general impli- cations and operationalise them into an explicit assessment framework (Section 7). Finally, we apply this framework to the manipulation-capable RFM landscape, providing both an overview and a criteria-based industrial maturity evaluation 3 (Section 8), before concluding with a synthe- sis and discussion of our findings (Section 9). A key preview of our results is that, despite impressive benchmark-level progress and increas- ing breadth of contributions, industrial maturity remains limited and uneven across the field: cur- rent models tend to address a small number of enabling dimensions at a time, rather than offer- ing integrated coverage of the deployment-critical requirements that dominate real industrial set- tings. This motivates our primary goal: to move the discussion from isolated benchmark successes towards a traceable, implication-grounded view of what it would mean for RFMs to be credible candidates for industrial deployment. Our main contributions are: 1. A clear categorisation of robotic control approachesthatsituatesRFMswithin the broader progression from rule-based automation to generalist robotic intelligence (Figure 1). 2. A concrete definition and clustering of RFMs (Section 2). 3. A comprehensive overview of the current RFM literature and state of the art, with a focus on manipulation-capable systems (Table D5). 4. An industrially motivated perspective on the rise of cobots, clarifying why their adop- tion amplifies the need for more general, instructable and adaptable control methods (Sections 5 and 6). 5. A synthesis of domain-agnostic industrial deployment implications distilled from indus- trial literature into eleven interdependent view- points (Section 7.1). 6. An extensive, implication-grounded criteria catalogue comprising 149 items to assess indus- trial maturity of RFMs (including ecosys- tem requirements), which we believe has not previously been developed in this form for RFM-centred industrial readiness assessment (Section 7.3). 7. A large-scale evaluation of the current RFM landscape against this catalogue, elucidat- ing strengths, weaknesses, and research gaps that must be addressed to advance towards industry-grade RFMs (Sections 8 and 9). 2 Robotic Foundation Models FMs represent a significant milestone in artificial intelligence, forming the basis for highly gener- alisable systems that can be efficiently adapted to diverse downstream tasks. These models are trained on extensive, heterogeneous datasets, often employing self-supervised learning methods to extract patterns and relationships well beyond their original training objectives [8–10]. This breadth of data enables FMs to exhibit emergent behaviour – capabilities and skills that substan- tially exceed the sum of their training data and initial design intentions. As Marsh [11] describes, emergent behaviour is “greater than the sum of the parts”. In practice, this capability is shown in Large Language Models (LLMs), trained sim- ply to predict the next word in a sentence, which demonstrate abilities in arithmetic, programming, and reasoning that were never explicitly targeted during training, but arise from their scale and dataset diversity [8]. Thus, as seen in domains like Natural Language Processing (NLP) where input-output formats and modalities can often be standardised, these models deliver robust, adapt- able performance across a variety of tasks that previous methods were unable to solve [8–10]. A fundamental interpretability challenge in FMs centres around their dualistic nature: on the one hand, they may function as single, general models whose mechanisms are widely applicable; on the other hand, their ability to adapt to profoundly different tasks may stem from an inherent collec- tion of specialised expert models, each tailored to a specific context [8]. To efficiently interpret the behaviour of FMs and enable both their adap- tation and the creation of new FMs for novel domains, it is vital to understand the mecha- nisms by which a FM acquires, represents, and processes knowledge. However, when extending the FM paradigm to robotics, additional domain- specific complexities emerge: The robotic context introduces high-dimensional, closed-loop decision processes, in which a robot’s actions directly and recurrently influence the environment’s state, thus its sensory feedback and subsequent behaviour – a dynamic fundamentally distinct from static, offline data scenarios prevalent in language and vision tasks [8]. Moreover, robotic platforms dis- play vast heterogeneity in physical embodiments, configurations, and operational contexts. For this 4 reason, classical control solutions are typically tai- lored to specific robots, environments, and tasks. These systems are characterised by determin- ism, yet inflexibility and have historically lacked generalisability across tasks or hardware, high- lighting the need for control methods that are inherently task-agnostic, cross-embodiment, and open-ended. [9] Foundational methods – capable of leveraging highly diverse robotic datasets – can fill this gap, when satisfying criteria impli- cated by industrial use cases and needs. We argue that, analogous to how FMs in NLP seamlessly handle varied languages or dialects, emergence in robotic-specific FMs could result in the capa- bility to operate across different robot embod- iments, modalities, and environments. In effect, such RFMs could be modelled to interpret every robot platform and configuration simply as yet another dialect or language, enabling generalisable control, reasoning, and planning capacity irrespec- tive of the current task or embodiment. Definitions of RFMs vary among researchers: Hu et al. [12] distinguish between ”single-purpose” RFMs (focused on one modular capability, e.g. perception, planning, or control) and ”general- purpose” RFMs – models that can span per- ception, planning, control, and even non-robotic tasks. However, we argue that only the latter, models capable of integrating perception with actionable outputs (such as control, trajectory or plan generation), qualify as true RFMs. Meth- ods restricted to perception or control alone lack the necessary generalism, functioning more as encoders, decoders or specialised solutions rather than holistic foundational models. Furthermore, recent perspectives sometimes conflate Vision- Language-Action Models (VLAs) and RFMs, sug- gesting those terms are equivalent [13]. However, similar to Xiang et al. [14] defining VLAs as a class of RFMs, we also consider VLAs a distinct RFM sub-category. In particular, VLAs employ multi- modal inputs, incorporating language and vision, to provide human-like instructability and gen- eralisation, generating actions for robot control. Zhou et al. [15] note that VLAs unify vision, lan- guage, and action by aligning all modalities into a common format, such as tokens, thus allowing end-to-end training and prediction. Xu et al. [16] describe VLAs as typically built from three mod- ules: perception, reasoning/planning, and action. While this categorisation may be true for certain candidates, we think that models not explicitly modelling or defining a strict separation of those concerns also qualify as VLAs. Beyond modalities, Zhong et al. [17] highlight that VLAs belong to both digital- and embodied-Artificial Intelligence (AI), thus need to bridge digital intelligence with physical interaction, demanding robust adaptabil- ity to real-world complexity and robot hardware constraints. Synthesising these interpretations, we define Robotic Foundation Models (RFMs) as highly adaptable, data-driven models specialised for addressing robotic tasks, distinguished by a generalist core: The ability to efficiently adapt to varied tasks, settings, embodiments and hard- ware configurations with low dedicated engineer- ing or retraining effort – e.g. through high-level instructions, flexibility in action representation/ interpretation or providing a generalist foundation for task-specific tuning. A model that does not exhibit such a generalist core, e.g. in the sense of cross-task or cross-embodiment adaptability, does not qualify as an RFM under our definition. To satisfy their multi-purpose nature, RFMs must encompass: • Multi-Modality: Capability to ingest diverse proprio- and exterioceptive signals, such as sen- sor data, images and Natural Language (NL). • Output flexibility: Producing actions, plans, trajectories or similar outputs suitable for a spectrum of robots and task domains. • Adaptability: Ability to adapt or generalise to tasks, environments or embodiments with little manual intervention. Given the above definition, it becomes clear that VLAs form a prominent and distinct sub-category within the broader class of RFMs [14], as VLAs fulfil RFMs’ key criteria: they harness multimodal inputs in the form of natural language instruc- tions and visual perception, which grants them the flexibility to be intuitively prompted by human operators across a wide array of tasks, while pro- ducing low-level actions for direct robot actuation as output. However, RFMs are not constrained to producing low-level actions, hence they can be cat- egorised according to the principal functions they perform within robotic systems, such as (but not limited to): 5 • RFMs for Control: These models directly generate low-level action outputs for robot actu- ation, such as motor commands, joint velocities or force/torque signals, across diverse embodi- ments and environments. • RFMs for Planning: These models focus on high-level reasoning, task decomposition, (re)planning or explainability generation. They may operate by integrating multimodal percep- tion and context, producing sequential plans or sub-goals that downstream controllers execute. • Integrated RFMs: Advanced models jointly combining planning and control capabilities. These models most closely align with the vision of general-purpose robotic intelligence, capa- ble of internal reasoning, failure-explanation or -recovery and long-horizon (re)planning com- bined with direct robot control. A similar taxonomy (control & planning) has been proposed by Ma et al. [13] for VLAs, how- ever, we contend that particularly planners align more closely with the overarching RFM paradigm, as their outputs are predominantly of a high- level nature. Other than through NL leveraged in VLAs, adaptability can equally arise through alternative modalities, methodologies or instruc- tioning mechanisms – such as using goal images, state representations, or exhibiting a general high adaptability to novel tasks, e.g. through demon- stration. Therefore, subcategories like Vision- Action Models (VAs), such as DP [1] or Dreami- tate [4], devoid of NL understanding, still qualify as RFMs if they possess a multi-purpose nature, such as robust foundations for adaptation to var- ious tasks/ embodiments or multi-task/ domain integration. RFMs also encompass a myriad of specialised subclasses, targeting specific application areas or domains such as manipulation [18–20], loco- motion/navigation [21–23] or aerial [24] and underwater [25] robotics. Others target combina- tions thereof, such as mobile-manipulation [26– 29]. Such specialisation to embodiments or task- categories often stems from predefined action spaces or fine-tuning data composition. In the context of industrial deployment, manipulation stands out as the most pertinent capability, under- pinning essential functions like (dis)assembly, pick-and-place operations, sorting, and dexterous material handling. Arguably, manipulation con- stitutes one of the central challenges to resolve in general robotics, laying the groundwork for mastering other domains such as locomotion or navigation, which themselves can be interpreted as a form of self-manipulation. Ultimately, how- ever, RFMs should pursue the goal to be both application- and embodiment-independent, under- scoring their foundational nature. In summary, within the scope of this survey, a model qualifies as an RFM only if it exhibits a generalist core in the sense of cross-task or cross-embodiment adaptability, while providing multimodal input integration and flexible out- put generation. Models lacking these properties are considered specialised or narrow approaches rather than RFMs. 3 Review Methodology & Corpora Composition Total: 1263 Main (81.2%) Implications (9.9%) Further (8.9%) Fig. 4: Overview of all utilised corpora and num- ber of considered publications, as presented within this section. This section provides a detailed description of the procedures used to assemble, filter, and cat- egorise the literature corpora that underpin this survey. Our methodology is designed to ensure systematic coverage, transparency, volume man- agement and reproducibility throughout all phases 6 of literature acquisition and analysis. Figure 4 presents an overview of all main corpora – com- prising the principal RFM-related corpora (main), the industrial implication corpus, and further sup- plementary sets – together with the proportion of all considered publications. In the following sub- sections, we describe the automated acquisition pipeline, filtering strategies, and post-processing steps for each corpus, as well as the criteria guid- ing their composition and specific role within the wider context of this review. 3.1 Literature Acquisition Procedure To ensure both comprehensive coverage and reproducibility of our literature acquisition, we developed a modular Python pipeline for auto- mated article retrieval and filtering as depicted in Figure 5. This approach was designed to maximise retrieval breadth across relevant literature while minimising blind spots, standardising the litera- ture research procedure throughout all stages of corpora generation, and reducing manual inter- vention. The main tool in our pipeline (DB-Search) starts with following user-provided information: • One or multiple search queries formulated in a novel, structured Academic Search Query Syn- tax (ASQS), tailored to thoroughly capture all relevant literature domains. • The publication date range to be applied for article retrieval (e.g. from 2020 to the present). • Selection of target Databases (DBs) used for querying, from the list of implemented DBs (arXiv, GoogleScholar, OpenAlex, Scopus and SemanticScholar). • Specification of retrieval preference methodol- ogy (latest vs. most relevant results). • Upper bounds for the number of articles to retrieve per DB-query combination, serving as a cut-off to manage volume (e.g. 100 results per DB-query pair). The pipeline then automatically translates each ASQS query into the respective syntax of each target DB, relaxing unsupported features where necessary (such as proximity operators or fuzzy matching). All selected DBs are then automati- cally scanned for each specified query – resulting Query Proposal Manual Refinement Manual Refinement ASQS Translation Q1QN ... Merge & Duplicate Removal Metadata Retrieval & Check Duplicate Removal Raw Corpus Stage 1 ... Stage N Filtered Base Corpus Final ASQS Queries Proposed ASQS Queries Final Corpora ASQS- Syntax Filter Process DB-Search Retrieval Intent Fig. 5: General Literature Acquisition Pipeline Procedure: Manual user involvement is required only for the Manual Refinement steps (highlighted in pink); all other processes execute automatically based on user-input. Only the Query Proposal and Filter Process modules leverage LLMs, while the DB-Search is fully deterministic. in the usage of all possible DB-query permuta- tions. During the search process, all metadata returned by the used APIs for the resulting arti- cles are merged. Thereafter, identified duplicates, based on unique identifiers (such as DOIs), are ini- tially removed for all articles containing sufficient information to permit duplicate removal. Subse- quently, the system retrieves missing metadata from auxiliary APIs (e.g. SemanticScholar, arXiv, 7 Crossref) to ensure completeness. Metadata valid- ity is programmatically checked – e.g. by verifying consistency between titles and abstracts. Articles lacking essential information (not uniquely iden- tifiable) or presenting errors (e.g. title-abstract mismatch) are excluded from the results. Another round of duplicate removal is performed after these checks. The resulting collection of validated articles is stored in serialised files (JSON or YAML), and, if desired, also exported to PDF format with struc- tured details including, but not limited to: title, URL, authors, abstract, DOI, originating query, and unique identifier assigned by the retrieval- pipeline. Comprehensive reports documenting the entire retrieval process and its results are auto- matically generated. These include: • Overall Retrieval Report: – Actual (ASQS-derived) queries employed per database. – Time span and database selection. – Maximum and actual number of retrieved results. – Pipeline execution timestamps. • Per-DB-Query Combination: – Unique query ID, timestamp, ASQS/base and translated query. – Applied time and sorting restrictions. – Actual and maximum possible number of results returned. After the main retrieval, an LLM-based filter- ing tool (Filter Process) using GPT-4o is lever- aged on-demand for semantic relevance scoring. This tool splits the corpus for parallel process- ing and individually passes each article’s title and abstract, alongside the user-intended relevance criteria, through the LLM. It outputs a binary relevance label, a floating-point relevance score in the interval [0, 1] and a concise NL justification of its decision. Sequential or layered filtering can be adopted, using progressively stricter relevance cri- teria to optimise classification accuracy within the resulting corpora. The pipeline is further supplemented by another component (Query Proposal) for initial query recommendation generation, utilising two fur- ther LLM-instances of GPT-4o. Given the ASQS syntax-definition and user intent, they jointly craft and propose a suite of search queries. This assists in establishing a robust baseline for explo- ration and iterative query refinement. Combining these tools, we developed all princi- pal literature corpora used throughout this work and presented within the following subsections. The pipeline’s results were then complemented by manual post-processing, including enrichment (with articles from references or iterative hand- searches), correction, and the application of addi- tional manual filters to ensure quality, complete- ness, and relevance. 3.2 RFM-Related Corpora The RFM-related corpus constitutes the principal literature base for this survey, aiming to target all major research directions regarding RFMs and their associated domains. This includes not only Σ 1025 Manual Refinement Manual Refinement Query (Q) Proposal DB-Search Filter Process 48Qs 11Qs 6,497 1,408 5,089 383 22,755 341 M 321 R 231 L 77 D 55 S 2022+ All DBs Max. 400 Sort: relevance N=2 Fig. 6: Literature Acquisition Pipeline: Pro- cess of gathering the RFM-related corpora (see Figure 7). Results yielded by manual cross-search are included for simplicity. RFM themselves, but also adjacent contributions such as training protocols, universal extensions, concrete applications and evaluation studies as well as broader considerations on RFMs (e.g. safety, robustness, or adversarial challenges). As we focus on RFMs capable of direct low-level con- trol – as per our definition from Section 2: RFMs for control & integrated RFMs – we separate 8 those approaches from LLM-based, high-level or narrow (too specialised) approaches, denoted as ”High-Level”. In addition, we cluster literature concerning datasets, benchmarks and simula- tors, especially those highly relevant to RFM research. To give an overview of related works (see Section 4), survey and review publications relating to RFMs or general embodied AI form the last category to be distinguished within the main corpus. To get an initial overview of the literature- landscape and obtain both a comprehensive and up-to-date corpus, we performed two iterations of large-scale literature retrievals using our pro- posed standardised Python pipeline (Section 3.1), each configured with identical parameters and subsequently merged to maximise article cover- age. Figure 6 shows the concrete retrieval process combining both searches, for which the following settings were applied: • Queries: Eleven extensive ASQS-queries, defined to ensure broad coverage across the complete scope of the main corpus (as shown in Table B2 of the Appendix). • Publication Year Range: From 2022 to date of search (see below). • Databases: arXiv, GoogleScholar, OpenAlex, Scopus and SemanticScholar. • Results Per Query: A retrieval limit of 400 articles per database (x5) per query (x11), lead- ing to a maximal possible result set of 22, 000 articles. We intentionally combined a large set of non- restrictive and targeted queries with a high cut-off value, to ensure a comprehensive result list, while managing volume with targeted filters afterwards. Search 1: Executed on the 4th of July 2025, returning 10, 728 valid article entries, of which 5, 035 were determined unique after deduplica- tion. Purpose: Initial overview. Search 2: Conducted on the 30th of October 2025, returned 12, 027 valid articles, yielding 4, 999 unique entries after improved duplicate removal. Purpose: Currentness. Notably, despite identical parameters, the ini- tial corpus increased significantly in this short period, illustrating the rapid pace of the field and increased database coverage. Additionally, the duplicate removal algorithm was improved between search 1 & 2 leading to fewer unique articles in the latter. These two result sets were merged, followed by a cautious duplicate removal protocol that erred on the side of retaining ambiguous entries, resulting in a base corpus of 6, 497 as unique classified articles. This base was then filtered and refined through a multi-stage process: 1. Robotics-Relevant Filtering: A coarse LLM-based filter was first applied to remove articles not relevant to robotics. Of the 6, 497 publications, 4, 834 were retained as relevant, while 1, 663 were disregarded. 2. RFM-Relevant Filtering: The remaining 4, 834 articles were subjected to a stricter LLM filter, targeting relevance for RFMs, VLAs, and additional foundational approaches or relevant publications as per our targeted specification for this corpus. This yielded 1, 408 relevant articles. For both filter stages, the not-relevant entries were manually sampled and cross-checked. Within this process no false negative cases were identified. The resulting set of 1, 408 records then became the subject of manual in-depth review, further filtering, enrichment (e.g. inclusion of key refer- ences and works found through manual chain- searching), and categorisation. Duplicates not resolved algorithmically (e.g. due to identifier inconsistencies) were discarded manually. The procedure resulted in the following final, cate- gorised main corpus (see also Figure 7): • Total articles: 1, 025 unique and highly rele- vant works. • Models: 341 control or integrated RFMs, of which 324 focus on manipulation, mobile manipulation, or are generally applicable to multiple robotic domains. The remaining 17 target other specific embodiments or domains (drones, surgical robots, locomotion, navigation or underwater). • Related Works: 321 articles, e.g. training pro- tocols, extensions (not further categorised for this review). • High-Level: 231 articles, covering LLM-based (multi-agent) systems, as well as foundational but more narrowly scoped models or RFMs for planning. 9 • Data and Benchmarks: 77 articles, including dataset, simulators, benchmark suites, and data generation procedures. • Surveys: 55 literature survey and review papers, comprising 24 on general or embodied AI, and 31 that include RFMs (of which 11 focus specifically on RFMs). Across the 55, only 2 focus on industrially relevant topics. For more detail, see Section 4. Total: 1025 RFMs (33.3%) Related (31.3%) High-Level (22.5%) Data/Bench (7.5%) Surveys (5.4%) Fig. 7: Main RFM-Related Corpus Composition This categorised corpus provides the reference set for all subsequent, review-related analyses, ensur- ing broad and systematic coverage of RFM and closely associated fields. 3.3 Industrial Implication Corpus To systematically capture the requirements, chal- lenges, and context-specific attributes of real- world robotic applications in industrial domains, we gathered an industrial implication corpus, util- ising the same modular literature research pipeline described in Section 3.1. The intention underly- ing this corpus was dual: to gather (1) studies of concrete industrial robotics applications and use cases that implicitly reveal relevant industrial requirements, and (2) works providing explicit discussions or analyses of robotics requirements, constraints, and challenges unique to industrial environments. Manual Refinement Manual Refinement Query (Q) Proposal DB-Search Filter Process 49Qs 7Qs 892 296 596 171 348 2015+ All DBs Max. 50 Sort: relevance 125 N=1 Fig. 8: Literature Acquisition Pipeline: Process of gathering the industrial implication-related cor- pus. For this purpose, a comprehensive search depicted in Figure 8 was conducted according to the follow- ing parameters: • Queries: Seven broad ASQS queries (as shown in Table B3 of the Appendix), to ensure cover- age across a variety of industrial fields, including applications, case studies, analyses, and review papers. • Publication Year Range: From 2015 up to July 2025. • Databases: arXiv, GoogleScholar, OpenAlex, Scopus and SemanticScholar. • Results Per Query: A retrieval limit of 50 articles, yielding a maximum possible result set of 1, 750 publications. The pipeline yielded 1, 240 valid entries, of which 892 were determined to be unique after deduplica- tion. To reduce the volume and focus on relevant material, a coarse LLM-based relevance filter, con- figured with the same intent as the original search, was applied. As a result, the corpus was narrowed to 296 relevant articles and 596 classified as unre- lated. Thereafter, a manual in-depth review was performed on 125 of the relevant articles that were assigned the highest LLM-based relevance score (relevance ≥ 0.9). We chose to focus on this subset of most relevant articles, since the vol- ume seemed sufficient for the process of gathering 10 industrial implications. Lesser-scored entries (rel- evance < 0.9) were not included in the detailed review. The insights distilled from this core set have been iteratively synthesised, clustered, and mapped to industrial implications for robotic sys- tems. An overview and in-depth discussion of the extracted implications and their interpreta- tion within the RFM paradigm are provided in Section 7. 3.4 Further Corpora & Manual Searches In addition to the two main corpora presented in the previous sections, additional automated and manual searches about industrial robots, cobots, compute hardware, sensors, manipulators, effi- ciency considerations, and further related fields as well as norms/ standards have been performed, resulting in a total of 113 auxiliary publications. Those have been considered and used throughout the process of crafting and writing of this work. 4 Related Works Within our extensive literature review, as described in Section 3.2, we identified a total of 55 relevant survey and review articles. Figure 9 Total: 55 Surveys +RFMs (56.4%) GEA (43.6%) NON-RFM-focus (64.5%) NON-IND (95.0%) IND (5.0%) RFM-focus (35.5%) NON-IND (100.0%) NON-IND (95.8%) IND (4.2%) Fig. 9: Identified Related Works (Survey cor- pus) composition: GEA=general embodied AI, +RFMs=including RFMs, IND=industrial-focus. shows the distribution of this survey corpus across general embodied AI (GEA), those includ- ing RFMs (+RFMs), and with explicit industrial focus (IND). A comprehensive list of all gathered surveys is provided in the Appendix (Table A1). Here, however, we concentrate on works that are most relevant to our objectives: surveys focused on RFMs, as well as those dedicated to industrial sce- narios – even if their main scope does not extend to RFMs. Most surveys to date have concen- trated on VLAs (as distinct sub-category of RFMs as discussed in Section 2) and their associated aspects, including architectures, challenges, learn- ing paradigms and other core considerations as illustrated in Table 1. There are several specialised surveys addressing particular facets of the VLA landscape, such as action tokenisation [17], dual- system architectures [31] or post-training strate- gies [14]. Some also discuss related model classes, like LLMs, Vision Language Models (VLMs) or Vision Foundation Models (VFMs) [12, 13, 15, 17]. Focusing on the RFM perspective, Table 1 com- pares the 11 surveys (20% of the 55 total) which specifically target the RFM field. Importantly, none of these works present an explicit industrial focus and almost all centre on VLAs, which forms only one distinct subset of RFMs. Merely two of those diverge from a VLA-centric approach: On the one hand, Zhou et al. [15] investigate language- conditioned manipulation, therefore bearing close resemblance to the VLA category. On the other hand, Hu et al. [12] discuss the overarching RFM paradigm, but primarily give an overview over cur- rent methodologies. Consequently, although there is a myriad of recent surveys around VLAs, comprehensive investigation of the broader RFM concept remains largely unaddressed, with vir- tually no attention given to further RFM sub- categories – such as Vision-Action models or other control-oriented and integrated RFMs, as defined in Section 2. Considering industrial relevant-reviews, Figure 9 and Table 2 reveal that only 2 out of the 55 analysed surveys (3.6%) possess a clear industrial application focus. Furthermore, neither empha- sises the implications of RFMs within industrial domains: Fan et al. [35] explicitly focuses on VLMs in Human-Robot Collaboration (HRC) for smart manufacturing – thus only a narrow subset of what we consider important for industrial robotics 11 Table 1: RFM: Overview of most relevant survey/ review papers with RFM focus. Reporting the focus area, the review’s main considerations (Scope), the types of targeted model families (Models), the timespan of reviewed papers (Span), whether the survey is considered concise or extensive (Broad), and the survey’s publication year. RefFocus AreaScopeModelsSpanBroadYear [13] VLAs for embodied AI Architectures, challenges, components, control, datasets, objectives, resources, simulators, tasks, task planning, taxonomy Encoder, LLM, VLA, VLM 14-24✓2025 [17] Action tokenisation in VLAs Action token categorisation: (affordance, code, goal state, language, latent, raw action, reasoning, trajectory), datasets, FM evolution LFM, VFM, VLA, VLM 17-25✓2025 [16] Challenges and roadmap in VLAs Applications, challenges: (dataset and evaluation, execution, generalization, representation, safety), evolution, roadmap, milestones, modules, opportunities VLA22-25✓2025 [15] Language- conditioned robot manipulation Auxiliary tasks, challenges, datasets, evaluation, environments, learning paradigms, limitations, neuro-symbolic, planning, reasoning, reward shaping, semantics extraction, task representation FM, LLM, VLA, VLM 18-24✓2025 [30] Large VLM- based VLAs Architectures (hierarchical, monolithic), benchmarks, characteristics, datasets, task planning, learning paradigms, memory mechanisms, multi-agents, operational strengths, taxonomy VLA23-25✓2025 [31] Dual-System VLAs Architectures, definition, design-elements, empirical evaluation, training strategies VLA24-25✗2025 [14] Post-training strategies for VLAs Challenges, embodiment awareness, model development, multi-component integration, perception, post-training methods, task comprehension, taxonomy VLA23-25✓2025 [32] VLA overview and taxonomy Applications, architectures, benchmarks, challenges, classification, datasets, hardware, paradigms, simulators, taxonomy VLA21-25✓2025 [33] VLA design and guidelines Design choices (architectures, backbones, cross-embodiment), guidelines, performance comparison, RoboVLMs VLA22-24✗2025 [34] VLA advancement synthesis Agentic AI, architectures, challenges, concepts, control, cross-embodiment, domains, evolution, foundations, planning, real-time, training strategies VLA22-25✓2025 [12] FMs for general- purpose robotics Benchmarks, challenges (generalisation, data scarcity, safety, requirements, task specification), control, datasets, FM application, manipulation, perception, RFM characteristics, task planning, taxonomy LLM, RFM, VFM, VLM 22-24✓2024 Table 2: IND: Overview of most relevant survey/ review papers with industrial focus. RefFocus AreaScopeModelsSpanBroadYear [35] VLMs in HRC for smart manufacturing Architectures, human-robot skill transfer, limitations, manipulation, navigation, pretraining, task planning LLM, VLM20-24✓2025 [36] Embodied intelligence in smart manufacturing Applications, architecture, challenges, characteristics, capabilities, limitations N/AN/A✗2024 applications (see Section 7) – while also focus- ing solely on higher-level VLMs, rather than on RFMs. Ren et al. [36] conceptually address general embodied intelligence, evaluating capabilities and limitations in industrial contexts, but they do not consider any specific models or what the industrial use case may imply for a given type of model. Synthesising these findings, we expose a pro- nounced gap in the literature: given our corpus, there is currently no survey dedicated to the industrial application of RFMs or implications of industrial use cases on RFMs, not even within the restrictively narrower context of VLAs. Industri- ally focused surveys are predominantly either too general or limited to specific facets (e.g. VLMs in HRC). By contrast, our work diverges from the prevailing focus on VLAs’ architectural or functional paradigms. Instead, besides discussing taxonomy and giving an extensive overview over the RFM landscape, we provide a detailed anal- ysis of the attributes and capabilities demanded of RFMs by industrial use cases, elucidating how 12 such requirements should shape or be integrated into future RFM design and how industrial matu- rity can be evaluated in given approaches – a perspective unmet in current literature. 5 Rise of Cobots in Industry Recent advances in RFMs are predominantly demonstrated alongside cobots, yet these devel- opments are frequently confined to laboratory, table-top or kitchen settings. To highlight the broader industrial significance of this trend, it is essential to first consider the evolution of indus- trial robotics and the transition toward more collaborative robotic systems. 5.1 Landscape of Industrial Robotics: Tradition & Transformation Industrial robots have long constituted a fun- damental component of automated manufactur- ing, typically defined as automatically controlled, reprogrammable multipurpose manipulators, pro- grammable in three or more axes and intended for use in industrial automation applications [37, 38]. Core to their design is an incomplete-machine architecture, requiring the integration of addi- tional tooling such as grippers or welding torches to fulfil specified tasks. Their flexibility stems from complex series structures – kinematic chains with multiple rotational degrees of freedom – which provide mobility and adaptability across a broad range of production processes. [38, 39] However, traditional industrial robots are primarily pro- grammed for pre-defined, automated function and demand specialised expertise for complex repro- gramming. Due to their high speeds and signifi- cant payload capacities, they typically operate in designated spaces separated from human workers for safety reasons and are thus commonly deployed for heavy-duty, repetitive or hazardous industrial tasks. [40–42] To solve varying tasks within automation, several industrial robotic configurations have emerged, ranging from highly specialised parallel struc- tures to general-purpose articulated arms. They typically make use of revolute (R) or prismatic (P) joints and can be classified into six cate- gories [41, 43, 44], which are shown in order of increasing similarity to cobots within Table 3. Table 3: Main industrial robot configuration categories. [41, 45–48], [49, p. 72, 222, 452, 1393] TypeConfigComments Delta (P) Parallel • Several symmetric kinematic chains • High speed & acceleration • Redundancy increases accuracy Cartesian (P) Serial • Perpendicular prismatic joints • Smooth linear movement • Rectangular workspace Cylindrical (RPP) Serial • Rotational base combined with two prismatic joints • Cylindrical workspace SCARA (RRP) Serial • Both revolute joints along the same axis • High accuracy & repeatability • Planar positioning Spherical Arm (RRP/R) Serial • Revolute axes intersect at a common center point • Simplifies collision & reachability analysis • Spherical workspace Articulated (R) Serial • Revolute joints’ axes typically offset • Similar to human arm • Flexible 3D workspace Despite their varied forms and operational strengths, these traditional industrial robot types inherently lack the flexibility, adaptability, safe direct HRC and Human-Robot Interaction (HRI) – attributes now increasingly demanded by mod- ern manufacturing, especially Small- and Medium- sized Enterprises (SMEs)[50–53]. Cobots have emerged as a dedicated response to these limita- tions. Initially introduced as intrinsically passive devices designed for operator guidance and con- strained movement [54, 55], cobots have evolved into a distinct subcategory of industrial artic- ulated robots engineered to operate safely and efficiently alongside humans. Contemporary def- initions emphasise their lower payloads, reduced speeds, rounded geometries and sensor suites for real-time contact or human detection, allowing safe intentional HRC [39, 56–58]. Importantly, cobots are tailored for flexible deployment, intu- itive HRI and programming as well as rapid reconfiguration – minimising downtime, setup costs and enable frequent application changes. This makes them particularly attractive for SMEs requiring high product variability and small batch sizes. [39, 56–59] The principal differentiator from 13 traditional industrial robots lies in their func- tional ethos: while conventional robots are gen- erally engineered to replace human involvement – automating entire processes within isolated environments – cobots are expressly intended to enhance human capability by enabling joint exe- cution of complex, variable, or ergonomically chal- lenging tasks. Rather than aiming for maximal speed or payload, cobots prioritise safe coop- eration, ease of integration and efficiency gains through shared workflows. [40] 5.2 Productive Cobot Applications & Domains Due to their increased flexibility and safety com- pared to traditional industrial robots, modern cobots have exceeded their origins in laboratory and demonstrator settings to become integral elements across a wide spectrum of real-world industrial and commercial domains, from manu- facturing floors to healthcare and logistics: Manufacturing and Assembly: Cobots are fre- quently deployed in the manufacturing sector, par- ticularly in tasks that require consistent precision alongside adaptability, such as assembly, pick- and-place operations, machine tending, quality inspection, palletising, packaging, general mate- rial handling and welding or gluing [50, 60, 61]. Concrete examples include the use of Universal Robots’ UR10e cobots at PSA and Fiat Chrysler automotive assembly plants for activities such as door, hood and soft-top assembly, alongside riveting and mounting [60]. Further applications include repetitive pick-and-place tasks in plant nurseries, automated packaging in board manu- facturing or assistance in soldering, drilling and arc welding processes in metal working environ- ments. [50, 60, 62]. Quality Control and Inspection: By leverag- ing advanced sensor and computer vision systems, cobots perform real-time inspection of part dimen- sions, assembly integrity and defect detection. In electronics manufacturing, for example, UR3s are efficiently utilised for quality inspection of auto- motive touchscreen panels [60]. These inspection roles leverage the cobots’ repeatability, adaptable end-effectors and their capacity to be easily recon- figured for frequent product changes. [50, 60, 61] Material Handling, Logistics, and Mobile Manipulation: Industrial environments fre- quently integrate cobot arms with Autonomous Mobile Robot (AMR) platforms, resulting in mobile manipulators that can transport goods between workstations, feed machines and thus enable flexible and precise in-factory logistics. Concrete use cases include the movement of assembly components using MiR200 AMRs in appliance manufacturing [60], as well as the man- agement of heavy, repetitive, or ergonomically challenging transport tasks, thus enabling more agile and just-in-time production processes. [42, 59, 60]. Healthcare: Cobots support patient handling, rehabilitation, medication dispensing and precise surgical assistance. Applications include robotic surgery, manipulation of medical tools and remote ultrasound or endoscope positioning [50, 52, 63]. Agriculture: Tasks include precise planting, har- vesting and sorting in harsh or unpredictable environments. During periods of high-volume workloads, continuous 24/7 operation can reduce reliance on manual seasonal labour for highly repetitive tasks. [60, 63] Retail and Service: Cobots assist with shelf stocking, object retrieval and customer engage- ment, thereby improving service quality and oper- ational efficiency in commerce and hospitality sectors [52, 63]. Military, Security and Inspection: To increase throughput and operator safety, cobots equipped with specialised sensors, are employed in luggage scanning, dangerous substance detec- tion, bomb disposal, facility inspection, cleaning and delicate manipulation tasks in environments such as spacecraft maintenance, hazardous mate- rial operations, airport and public security con- texts. [50, 63, 64] In particular, cobots have been employed to inspect and clean the Space Rider orbiter’s surfaces during extended missions [52], for bomb defusing during military operations using specialised tools [63], or for assistance of military personnel during border inspection [50]. 5.3 Drivers of the Industrial Shift Toward Cobots As implied by the adoption of cobots across diverse sectors outlined in Section 5.2, there has been a notable shift within industrial and commercial domains from exclusive reliance on 14 manual labour or traditional industrial robots towards increasing adoption of cobots. This shift is driven by a broad array of factors – tech- nological, organisational and economical – that collectively highlight the growing necessity for adaptable, user-friendly and interactive automa- tion solutions within diverse industrial domains. Traditional industrial robots, with their high pay- load capacities, precision and speed, have long been essential for large-scale manufacturers focus- ing on high-volume, repetitive processes. However, these systems, which are complex, expensive, and difficult to reconfigure, often prove unsuitable, especially for SMEs whose operational require- ments centre around low-volume, high-mix pro- duction. [59, 65, 66] Cobots directly address these needs through their more flexible design, compact footprint and comparable ease of pro- gramming, enabling straightforward redeployment across varying tasks and locations within man- ufacturing environments [59, 60, 67]. Notably, cobots claim to offer user-friendly interfaces and intuitive teach modes, thereby reducing the need for specialised engineering expertise and minimis- ing integration time [61, 67]. Aside from increas- ing operational efficiency, cobots are intended to directly assist human workers by taking over repetitive, physically demanding, or hazardous tasks, thus enhancing workplace safety and allow- ing human operators to focus on assignments requiring higher cognitive skills and problem- solving abilities [40, 58, 59]. In contrast to con- ventional industrial robotics, their design enables direct HRC, allowing the combination of human creativity with robotic reliability and precision in shared workspaces [58, 67]. Several safety features – including real-time sensor integration, force/- torque monitoring and intelligent behaviour adap- tation – are intended to enable cobots to operate closely with people, reducing the need for imprac- tical and expensive barriers or isolated cages for a respective use cases [39, 60, 61, 63]. In addition to optimising available production space, this also enables workflow modernisation and a reduction in ancillary costs. The increasing demand for cobots is reflected in both industrial practice and research trends. While the total volume of publications on tradi- tional industrial robots remains higher, Khedr et al. [56] showed in their survey, that the growth rate of cobot-focused research has accelerated rapidly in recent years – over 5-fold within a sin- gle decade – underscoring their expanding role in advanced manufacturing and their rising rele- vance for future automation solutions [56]. This is also supported by the absolute number of global installations: Annual cobot deployments have increased from 11,000 units in 2017 to 55,000 units in 2022, reflecting not only ongoing innovation but also expanding industrial uptake alongside continued strong deployment of con- ventional robots [41, 68]. Correspondingly, the International Federation of Robotics has noted a consistent increase in the proportion of cobots rel- ative to traditional industrial robots [68]. Beyond hardware, advanced technologies such as com- puter vision and adaptive control algorithms have extended cobot capabilities, paving the way for more sophisticated and autonomous interaction within dynamic industrial settings [60, 69]. This ongoing evolution aligns with the principles of Industry 4.0, where digitalisation, interconnectiv- ity and smart automation require flexible systems competent of both autonomous operation and collaborative behaviour. Given sufficiently capa- ble control methods, cobots could be increasingly integrated as key components of smart facto- ries, to connecting with other industrial Internet of Things (IoT) devices and AI-driven systems, enable data-driven optimisation, remote operation and rapid reconfiguration for changing market demands [59, 61, 68, 69]. 5.4 Towards Intelligent & Collaborative Automation Despite these advances, there remain specific chal- lenges and requirements for cobot deployment in industrial environments. While cobots are effec- tive at supplementing human labour and under- taking routine or ergonomic tasks, their ability to autonomously handle complex, cognitive, or dex- terous processes is still limited [40]. Issues such as real-time hazard detection, continuous safe oper- ation, as well as trustworthy HRI and HRC call for further development of sophisticated control strategies and intelligent interfaces. This is partic- ularly pertinent regarding trust and transparency during both initial deployment and ongoing opera- tion, which are regarded as critical success factors for the introduction of cobots [40, 58]. In this con- text, RFMs – as introduced in Section 2 – have 15 the potential to address many of these emerging challenges. RFMs, when integrated in industrial cobot platforms, can provide powerful tools for intuitive instruction, autonomous decision-making and real-time adaptation by leveraging natural language inputs, multisensor fusion and advanced reasoning capabilities. Both, cobots’ productive use cases and their current limitations form a perfect fit for RFM integration, making the assess- ment of emerging RFMs’ industrial applicability crucial. Future research must focus on ensuring that these models are not designed solely for con- trolled laboratory settings, but are robust and effective within the complex, variable and safety- critical landscapes of actual industrial operation. In summary, the ongoing move from traditional industrial robots to cobots reflects a broader reori- entation of manufacturing towards collaborative, intelligent and flexible automation. This transition underscores two key aspects: (a) the industrial rel- evance and necessity of RFMs, which have high potential to directly shape the capabilities, acces- sibility and trustworthiness of next-generation col- laborative robotics; but conversely (b) cobots also create the deployment pressure that exposes the limitations and weak points of current RFMs for productive use. 6 Cobot & Compute Hardware To meaningfully assess the maturity and applica- bility of RFMs in industry, it is essential to first understand the capabilities and typical platforms for deployment. As discussed in Section 5, contem- porary cobots still inherently lack the ability to autonomously handle complex or cognitive tasks – gaps that RFMs are well positioned to address. However, evaluating RFMs without reference to the practical capabilities, such as reach, payload, accuracy and utilised perception or manipulation hardware of commonly used cobots, we would risk incomplete analysis. Additionally, considera- tions such as power consumption and inference hardware, which are often overlooked in scien- tific studies, are critical for translating laboratory results to industrial applications. Clear insight into typical industrial computing platforms and their real-world performance is thus necessary. Therefore, this section provides an overview of the key hardware and system components required for the deployment of cobots in industrial settings. We examine the quantitative capabilities and power consumption profiles of contemporary cobots in Section 6.1, the characteristics and suitability of industrial edge computing platforms necessary for real-time control and inference in Section 6.3, and the essential sensing and manipulation hard- ware required for real-world robotic applications in Section 6.2. Through a combination of sur- vey data and representative hardware comparison, we elucidate technical prerequisites for integrating RFMs in cobotic environments, considering robot capability, computational infrastructure, sensor integration and operational efficiency. 6.1 Cobot Capabilities & Power Consumption To highlight aforementioned prerequisites for RFM integration and to assess cobots suitabil- ity across various industrial tasks, a quantitative understanding of contemporary cobots’s capabili- ties is crucial. To this end, an extensive survey of cobots by Taesi et al. [50] provides comprehensive data, which we have synthesised in Tables 4 and 5, reporting minimum, maximum, mean, and median values across key specifications. In the follow- ing, only the articulated anthropomorphic cobots of [50] are considered, consistent with the typical definition of cobots as presented in Section 5.1. Table 4: Composed cobot configuration and capability statistics. [50] Values are rounded to two decimal places. n = 180DoFPayloadReachAccuracy (kg)(m)(m) Min40.32800.01 Max717027901 Mean 6.09 8.56 970.120.09 Median 65.59100.05 Table 4 aggregates statistics from 180 articulated cobot models, focusing on Degree of Freedoms (DOFs), payload, reach, and accuracy. The major- ity of cobots feature a mean of 6.09 DOFs, underscoring the dominance of six-axis articulated robots in collaborative applications. Typical pay- load capacity spans a median of 5.5kg (mean: 8.56kg), clearly indicating that most commer- cially available cobots are optimised for handling 16 moderate loads. The average reach is approxi- mately 910m (mean: 970m) and positioning accuracy is typically very precise, ranging from a median value of 0.05m to a mean of 0.09m. Notably, there is a significant spread in both payload and reach, with extreme cases such as the Comau Aura, which offers a payload up to 170kg and a reach of 2790m. While the presence of high-payload outliers slightly biases the mean values upwards, these statistics capture the core capabilities of most cobots deployed in industrial contexts and confirm the suitability of cobots for applications prioritising moderate payloads, high repeatability and operational flexibility across a diverse range of industrial fields as specified in Section 5.2. Table 5: Composed peak cobot TCP velocity and power consumption statistics. [50] Showing only values for typical cobots with a payload between 0.5 and 20kg as in the wattage comparison of Taesi et al. [50]. Values are rounded to two decimal places. n = 63Payload TCP Velocity Power Consumption (kg)(m/s)(kW) Min0.50.20.06 Max2060.9 Mean6.691.680.31 Median51.10.25 Table 5 provides insight into the dynamic perfor- mance and power requirements of typical cobots rated for payloads between 0.5 and 20 kg – as those more accurately represent the average cobot identified in table 4. The maximum Tool Center Point (TCP) velocity for these models ranges from 0.2 to 6 m/s, typically between 1.1 and 1.68 m/s. Power consumption is remarkably restrained com- pared to conventional industrial robots: typically between 0.25 and 0.31 kW, with a maximum of 0.9 kW for heavier-duty models within this class of cobots as shown in table 5. This is in strong contrast to traditional industrial robots, whose average power usage has been estimated at around 3 kW by Barnett et al. [70], with peak draws mea- sured as high as 12 kW (KUKA KR60HA) [71] or over 15 kW (KUKA KR210 R2700) [72] and sus- tained values near 6 kW under typical acceleration or velocity [71, 72]. The lower energy require- ments of cobots yield significant operational cost advantages, facilitate use in environments with lower power infrastructure and enable deployment on mobile or battery-powered platforms. Impor- tantly, these comparatively low power require- ments have direct effect on the hardware selec- tion and computational strategy for cobots con- trol, thus RFMs: Successful real-time deployment requires models that operate efficiently within the constrained energy budgets typical of cobotic sys- tems, particularly for mobile or battery-powered applications. Achieving low-latency inference on hardware that adheres to strict industrial power limits underscores the importance of model effi- ciency. 6.2 Sensing & Manipulation Hardware Sensors are essential to the functional operation of cobots, serving as the primary means for envi- ronmental perception and self-monitoring. Their central role covers key industrial aspects such as HRC, HRI, autonomous navigation, manipulation and real-time task execution [52, 73]. A variety of sensor types are commonly deployed in both research and industrial cobotic settings [40, 42, 50, 52, 57, 61, 62, 65, 73–75]: • Visual: RGB (single/stereo), depth, Charge- Coupled Device (CCD) and 2D/3D vision sen- sors • Range: Light Detection and Ranging (LiDAR), Radio Detection and Ranging (RADAR), Sound Navigation and Ranging (Sonar), ultrasonic, Time-of-flight (TOF)/Photonic Mixing Device (PMD) and laser scanners • Dynamics:InertialMeasurementUnits (IMUs), gyroscopes, accelerometers, force/- torquesensors,tactileandpressure sensors • Position and Identification: Encoders, RadioFrequencyIdentification(RFID), infrared, Global Positioning System (GPS), odometry and magnetic sensors • Specialised: Acoustic, thermal and similar sensors 17 Advanced applications, such as RFMs, can bene- fit from sensor fusion, combining modalities such as vision with proximity, inertial, or force/torque measurements to increase perception robustness and accuracy [52]. Configurations can involve redundant or complementary sensing to support safety, compliance, and reliable manipulation – for example, using visual and tactile feedback for pre- cise object handling, or adding force sensors to detect unwanted contact in collaborative settings. End-effectors constitute the physical interface for manipulation tasks and are tailored to spe- cific application requirements. As outlined in Section 5.1, the integration of these interfaces is essential for enabling an industrial robot/ cobot to perform its designated tasks. Concrete exam- ples include grippers with configurable fingers and pads, suction cups, dual-stage end-effectors incor- porating both gripping and suction capabilities or custom tooling such as drills and polishers for advanced operations [62, 65]. Many end-effectors embed their own dedicated sensors [42] – such as force feedback for manipulation tasks or embed- ded vision modules for object detection and pose estimation. Notably, the modularity of modern cobot end-effectors allows quick adaptation to varying industrial tasks, which increases reusabil- ity and operational uptime. Nonetheless, attaining this level of modularity assumes similar flexibility in action-space of the control algorithm. Overall, the selection and integration of sen- sors and end-effectors play a pivotal role in the industrial applicability of cobots and must be con- sidered for deploying robust and efficient RFMs in real-world scenarios. Specifically, the sheer amount of sensing and manipulation hardware results in an even greater number of observation- /action-space permutations, further highlighting the need for multi-modality and output flexibility in modern control algorithms – capabilities central to RFMs (cf. Section 2). 6.3 Industrial Edge PCs Given the low-latency requirements of cobotic applications, computational platforms for real- time AI-control and -perception are often realised as edge devices – hardware installed as close to the robot and its data sources as feasible, often directly integrated into the cobot’s controller or platform. They are characterised by processing real-time data locally rather than relying on cloud or external datacentres. [80] This proximity is vital for latency-sensitive robotic applications, espe- cially for mobile or remote environments, where network traffic must be kept to a minimum [81]. Alternative paradigms such as fog or cloud com- puting can provide scalable resources but come at the cost of additional infrastructure or third-party service requirements. Fog computing, as defined by Yi et al. [82], encompasses distributed and col- laborative architectures involving heterogeneous devices close to the edge, but still depends on non- local hardware and stable networking. Classical cloud computing, while enabling on-demand com- putational resources, is generally associated with higher latency, lower throughput and an increased risk of service interruption – limitations that are particularly problematic for time-critical robotic control. [82–84] According to Ismail et al. [85], moving computation closer to the edge grows increasingly important for industrial deployments, due to a drop in compute costs and the oper- ational advantage of minimising recurring net- working expenses [85]. Accordingly, we provide an overview of currently available, representa- tive edge devices to clarify relevant hardware requirements for industrial RFM deployment and to establish a foundation for real-world inference evaluation and power consumption considerations. For the quantitative hardware analysis presented in Table 6, we selected NVIDIA as the repre- sentative manufacturer for several reasons. Their Jetson devices are among the most widely adopted platforms for edge AI in robotics research and industrial applications, offering a standardised and scalable portfolio (from the compact Nano to the new high-end Thor series). Most importantly, relevant performance metrics, pricing, and other specifications are consistently documented and publicly available for the entire product family. This transparency enables an accurate, fair, and reproducible comparison of hardware options and their suitability for running modern robotic con- trol methods, predominantly based on AI models. While Table 6 concentrates on the Jetson port- folio, a range of alternative manufacturers – such as Siemens, AMD (Xilinx), Intel, Advantech or Beckhoff among others [86] – also provide relevant edge devices. However, for many of these alterna- tives, comparable and up-to-date public data on 18 Table 6: Comparison of NVIDIA edge devices’ AI performance in INT8 precision (dense) [76–78]. Prices for all Jetson devices correspond to single-unit purchases from the same NVIDIA-specified retailer [79], collected on the same date. Operational costs are estimated for 24/7 usage with 0.25$/kWh. All values are rounded to two decimal places. Su = Super, DK = Developer Kit, IND = Industrial Jetson Device Price ($)↓ TOPS (INT8)↑ Memory (GB)↑ Power (W)↓ GOPS /$↑ TOPS /W↑ Memory (MB)/$↑ $ /year↓ TOPS /y$↑ Orin Nano 4GB22934425148.471.3617.4754.750.62 Orin Nano 8GB24967825269.082.6832.1354.751.22 Orin Nano SuDK24967825269.082.6832.1354.751.22 Orin NX 8GB479117840244.262.9316.787.61.34 Orin NX 16GB6991571640224.613.9322.8987.61.79 Orin AGX 32GB9992003240200.2532.0387.62.28 Orin AGX 64GB17992756460152.864.5835.58131.42.09 Orin AGX DK19992756460137.574.5832.02131.42.09 Orin AGX IND23492486475105.583.3127.25164.251.51 Xavier NX5022181541.831.415.9432.850.64 Xavier NX 16GB57921162036.271.0527.6343.80.48 Xavier AGX99932323032.031.0732.0365.70.49 Xavier AGX 64GB139932643022.8710.745.7565.70.49 Xavier AGX IND144930324020.70.7522.0887.60.34 Thor T50003199517128130161.613.9840.01284.71.82 Thor AGX DK3499517128130147.763.9836.58284.71.82 Min2292141520.70.7515.9432.850.34 Max3499517128130269.08545.75284.72.28 Mean1292.31163.1342.2549.06138.422.7729.26107.451.27 Median999923240148.112.832.0287.61.28 RTX 5090 a 2600838N.A.575322.311.46N.A.1259.250.67 a Pricing estimated due to fluctuation and local availability at time of writing. As Jetson Devices use low-power DDR memory (LPDDR) for the complete system and the RTX cards are supplied with dedicated graphics memory (GDDR), those values are not considered. prices, performance (in AI-relevant metrics) and efficiency is often limited or unavailable. We addi- tionally report the specifications of a desktop-class Graphics Processing Unit (GPU), the RTX 5090, to provide context for typical platforms used in academic benchmarking of foundation and control models. While the RTX 5090 is not an edge device, but a standalone GPU intended for workstations and servers, it serves as a reference point for the upper bound of single-device computational performance in recent literature. Comparing the RTX 5090 to integrated Jetson devices highlights the trade-offs between desktop/workstation-grade and edge-class hardware. Examining the compara- tive statistics in Table 6 reveals several significant trends relevant to both industrial and research applications. The Orin Nano series stands out for its exceptionally low entry cost and high perfor- mance relative to purchase price, making it more suitable for cost-sensitive or large-scale deploy- ments. On the other hand, the entry-level Xavier NX – while more expensive – provides the low- est overall AI compute performance in the lineup. However, its 8GB variant is noteworthy for having the lowest maximum power consumption across all compared models, potentially favoring applica- tions where minimal energy draw is of essence. The mid-tier, such as the Orin AGX 32GB, provide a balance between increased computational capa- bility and moderate power requirements. Notably, the Orin AGX 32GB emerges as the most efficient device in terms of compute power per operational 19 cost, offering a strong compromise for applications prioritising both capability and efficiency. At the high end of the spectrum, devices like the Jetson Thor series or workstation-class desktop GPUs exhibit strong increases in computational through- put. However, this heightened performance is directly linked to substantial rises in both energy consumption and annual operational costs – a factor that can restrict their scalability and ren- der them impractical for mobile, battery-powered, or energy-constrained robotics platforms. Particu- larly, the RTX 5090 can consume more than twice the power of a median cobot as shown in Table 5. This contrast towards edge devices’ power draw, underscores that desktop GPUs are designed for short, intensive workloads or for environments where energy is abundant, whereas edge devices are optimized for consistent, energy-efficient use in real-world, often resource-constrained settings. Thus, a trade-off emerges: deploying larger and more capable RFMs is constrained by both com- pute and energy efficiency considerations, under- scoring the importance of aligning model design with operational cost and industrial practicality. Additionally, industrial-class devices (IND) fre- quently employ features such as ECC-memory for increased reliability, which can raise both initial and operational expenses or lower computational power compared to standard research hardware. 7 General & RFM-specific Industrial Implications Industrial deployment settings impose constraints and expectations that systematically differ from laboratory environments: they combine hetero- geneous hardware, tight safety and compliance obligations, high reliability targets, and contin- uous operational pressures. This section there- fore distils the recurring industrial viewpoints extracted from our industrial implication cor- pus (Section 3.3) – aligned with our discus- sions in Sections 5 and 6 – into an assessment- oriented structure that can be used to anal- yse how well current and future RFMs com- ply with real-world industrial demands. We pro- ceed in three steps: First, Section 7.1 synthe- sises the corpus into eleven general industrial implications that capture the dominant, domain- agnostic requirements reported across industrial robotics applications and that are supported by our cobot- and hardware-related considerations. Second, Section 7.2 translates these implications into a structured set of attributes that enable a more fine-grained discussion of where indus- trial constraints can manifest in practice within RFMs and their surrounding ecosystem. Third, Section 7.3 consolidates these considerations into a comprehensive criteria catalogue intended to support a consistent, transparent evaluation of industrial maturity across existing RFMs. 7.1 General Industrial Implications To derive general, domain-agnostic implica- tions from the industrial implication corpus (Section 3.3), we conducted an iterative qualita- tive synthesis across the 125 manually reviewed works. Concretely, we (1.) extracted statements that described recurring industrial requirements, constraints, and deployment challenges for robotic systems (both explicitly and as implied by described industrial case studies), (2.) grouped these excerpts into thematic clusters, and (3.) iter- atively refined and consolidated the most promi- nent clusters until they formed a compact set of implications that was sufficiently expressive to capture the dominant industrial viewpoints while remaining actionable for later assessment of RFMs. Throughout this process, we noticed no explicitly contradictory themes, instead found the implications to be interdependent rather than mutually exclusive: industrial authors frequently discuss the same practical challenge (e.g. human presence) through different lenses such as safety, interaction design, robustness, or perception, and our synthesis preserves these overlaps by separat- ing what is demanded from why it is demanded. The resulting set comprises eleven implications, each reflecting a recurring industrial expecta- tion about deployed robotic applications and, by extension, about learning-based robot control and foundation-model-based approaches. In the fol- lowing, we motivate each detected implication by summarising the corresponding clusters from the literature corpus. Implication 1: Adaptability & Flexibility Across industrial domains, authors consistently describe variability as a defining property of real deployments: processes change, product variants 20 proliferate, and boundary conditions shift, which in turn increases re-parameterisation effort and uncertainty [87–89]. Multiple works contrast this with classical industrial automation, emphasising that traditional solutions are often designed for repetitive, tightly structured workflows and tend to be brittle under even minor process changes [36, 90, 91]. As a consequence, adaptability is framed not merely as a convenience but as a key enabler for future “fast and low-cost” production and for increasing personalisation [36, 92, 93]. The corpus further highlights that industrial environ- ments are frequently dynamic, partially known, or unstructured, with unpredictable events and random disruptions being expected rather than exceptional [51, 89, 94]. This is intensified when humans are present in the workspace [51, 95]. Several authors therefore argue for robots and con- trol methods that can adjust behaviour online, be reconfigurable across products and processes, and reduce costly reprogramming cycles [88, 95–97]. In addition, sector-specific discussions (e.g. logistic- s/warehousing, outdoor industry) underscore that environmental variation (lighting, weather, clut- ter) can materially affect operation and must be handled robustly [98, 99]. Additionally, as outlined in Sections 5 and 6.2, cobots are seeing growing adoption in industry due to their flexible design and ease of reconfiguration. However, such adapta- tions must also be reflected within the applied con- trol system. Only methods that are intrinsically agnostic to changing tasks, embodiments or sensor setups and quality would enable cobots to realise their full intended potential. Taken together, the literature motivates adaptability and flexibility as a primary industrial implication: systems are expected to cope with frequent task, process, and environment variation while avoiding extensive downtime and engineering overhead [91, 96, 100]. Implication 2: Safety & Compliance Safety is consistently portrayed as a gating factor for industrial deployment, particularly in shared workspaces, physical interaction, and other high- consequence settings [51, 101, 102]. This is of espe- cial relevance owing to the expanding implemen- tation of cobots, as described in Section 5: Despite being intended to significantly reduce the need for physical separation between workers and robots, deploying cobots in shared workspaces requires an equally safety-aware control logic. Such safety awareness spans both general considerations that injuries must be prevented in human-shared envi- ronments [51, 103] and more explicit arguments that intelligent or adaptive robots must be tested, certified, and equipped with appropriate safety features prior to deployment [96]. Critically, mul- tiple works connect safety to standards and certi- fication frameworks as a prerequisite for industrial acceptability. The literature cites the role of ISO standards for safe installation and operation of robotic systems (e.g. ISO10218) and for collab- orative operation and contact-related constraints (e.g. ISO/TS15066), as well as further regional standards [104–106]. Beyond formal compliance, several authors emphasise that safety must be maintained under dynamic and unpredictable con- ditions, including human behavioural variability and sensing uncertainty [34, 36, 107]. Industrial discussions also note that safety mechanisms can reduce efficiency if they trigger overly conserva- tive stops or interruptions, motivating approaches that integrate safety-aware reasoning and sensing to preserve productivity without compromising protection [68, 108, 109]. Overall, the corpus moti- vates safety and compliance as a distinct impli- cation: industrially viable systems must satisfy relevant safety norms and preserve safe operation under realistic uncertainty, including failures and Out-Of-Distribution (OOD) inputs [34, 94, 107]. Implication 3: HRI & HRC In line with the drivers for the rising adoption of cobots discussed in Section 5.3, a significant segment of the literature likewise envisions future industrial robotics as inherently human-centric, where robots and humans work in proximity or physical collaboration to combine human flexibil- ity and cognitive skills with robotic repeatabil- ity and precision [110–112]. The literature uses established interaction taxonomies (e.g. differen- tiating coexistence, cooperation, and collabora- tion; and distinguishing remote, proximity, and physical interaction) to emphasise that indus- trial scenarios span multiple interaction regimes and thus demand correspondingly diverse inter- face and coordination solutions [93, 103, 113]. Works addressing HRC further highlight that direct physical collaboration is valued for its flex- ibility, but requires the robot to perceive con- tact and respond appropriately [92, 114]. Beyond 21 physical coordination, authors repeatedly empha- sise that communication quality and interface usability are decisive for acceptance and pro- ductivity. This includes calls for more intuitive, comfortable, and natural interaction, including speech and other modalities, to reduce reliance on expert programming and to better support operators in real time [36, 115, 116]. Several works discuss that current industrial control and programming workflows remain labour-intensive, motivating approaches that allow operators to instruct, supervise, and correct systems more directly [88, 97, 117]. The literature also points out that collaborative scenarios require reactive meth- ods that can handle uncertainty and unexpected behaviour to maintain fluid workflow between humans and robots [103, 118]. Hence, HRI and HRC emerge as a separate implication: indus- trial systems are expected to support human- centred collaboration with effective, natural com- munication and continuous operator involvement, particularly for complex, variable, and partially automatable tasks [35, 51, 104]. Implication 4: Robustness & Reliability Industrial environments are repeatedly charac- terised as uncertain, noisy, and prone to distur- bances and component imperfections. Authors dis- cuss both internal uncertainties (e.g. unmodelled dynamics, friction effects, sampling delays) and external uncertainties (e.g. measurement errors and environmental noise) as persistent challenges for deployed systems [100, 119]. This leads to calls for robust methodologies that can integrate per- ception, planning, and control under uncertainty, especially in complex systems such as warehouse automation or autonomous mobility [51, 120]. Reliability is also linked to monitoring and timely status updates in order to support reactive plan- ning and practical operation [108]. Importantly, the corpus extends robustness beyond stochas- ticity to include resilience against faulty sen- sors and OOD inputs, noting that invalid sen- sor readings can induce unpredictable outputs in learning-based systems and may be particu- larly dangerous in safety-critical contexts [107]. Further, some works explicitly consider adversar- ial and cybersecurity threats as integral require- ments for industrial robots [121, 122]. Accordingly, robustness and reliability form a distinct implica- tion: deployed systems are expected to maintain stable operation under uncertainty, disturbances, sensor degradation and security threats, while pro- viding operationally meaningful status and fault awareness [34, 102, 107, 121]. Implication 5: Precision & Accuracy While flexibility and generalisation are empha- sised, the industrial corpus simultaneously stresses that robots derive value from high precision, repeatability, and quality-critical execution [110, 123]. Several works frame industrial competitive- ness in terms of efficiency and accuracy, reflecting that many tasks require both throughput and reliable outcome quality [124, 125]. Moreover, accuracy is not restricted to actuation: sensing and perception accuracy is treated as foundational for safe and correct operation [121, 126]. Shen et al. [126] further note that current systems can exhibit insufficient grasping and positioning accu- racy in demanding settings, motivating improved sensing and algorithms. In the context of vision- /language-guided robot control, Fan et al. [35] explicitly highlight the gap between household- style demonstrations and industrial requirements, including the need for higher precision and improved motion planning accuracy while retain- ing flexibility and generalisation. Therefore, pre- cision and accuracy form an implication in their own right: industrial deployment demands con- sistent, quality-aligned execution and perception accuracy, often under noisy or varying sensing and environmental conditions, while preserving the required throughput to enable efficient work- flows [35, 100, 124, 125]. Implication 6: Real-Time Performance Real-time capability is presented as a perva- sive constraint affecting sensing, decision-making, interaction and control. The corpus references execution-time considerations (e.g. action-level and task-level timing) [51] and highlights low-level requirements such as real-time communication and motion control in integrated robotic sys- tems [53]. For industrial autonomy more broadly, authors emphasise real-time sensing and process- ing needs, including the acquisition of real-time environmental information and the integration of multi-modal signals to enable both sudden responses and higher-level reasoning [36, 52, 127]. Several works connect real-time feasibility to com- putational and resource constraints, especially for 22 edge deployment, and note that data-driven meth- ods are often limited by computational complexity and large training/inference requirements [115, 128, 129]. Indeed, as illustrated in Table 6, even the most capable contemporary edge-devices lag significantly behind modern workstation GPUs in computational power, as their design priori- tises reasonable energy consumption. Neverthe- less, real-time performance must still be achieved on these resource-constrained platforms. Besides computational concerns, real-time planning in dynamic scenes is explicitly identified as an unre- solved challenge in vision-language(-action) dis- cussions, while recent work also points to the pos- sibility of smaller models enabling real-time infer- ence in constrained settings [34, 35]. As a result, real-time performance emerges as an implication spanning the full system: industrial robots must perceive, decide, communicate, and act within tight latency bounds despite environmental com- plexity and limited compute budgets [119, 130]. Implication 7: Cost-Effectiveness & Inte- gration Capabilities A recurring industrial framing is that automa- tion must be economically justified, including for low-volume production and for SMEs. Authors argue that solutions must reduce manual effort, errors, and operational costs, and that high sys- tem cost remains a barrier to wider adoption of advanced autonomy [108, 116, 124]. Multi- ple works explicitly discuss that frequent prod- uct changes increase reprogramming effort and costs, motivating approaches that reduce com- missioning overhead and ease reconfiguration [88, 96]. Beyond acquisition cost, Triantafyllou et al. [120] stress integration as a practical obstacle: deploying state-of-the-art research components in industrial systems requires coherent integra- tion methodologies, yet community-wide stan- dards and guidelines are described as lacking. This concern is also reflected in specific domains such as AGVs/AMRs, where robust integration into factory architectures is highlighted as necessary for smart manufacturing success [131]. Further, authors link industrial adoption to scalability and cost-effectiveness of intelligent control, as well as to the ability to operate with cost-sensitive hardware and sensing configurations [132, 133], aligning with our observation that an efficient deployment of control methods on low-power edge devices is a necessity, as modern high-tier GPUs can consume more than twice the power of a median cobot itself (cf. Section 6.3). Hence, cost- effectiveness and integration capabilities emerge as an implication: industrial solutions must be economically viable and practically integrable into existing production environments, including under heterogeneous, resource-limited hardware and organisational constraints [35, 68, 132]. Implication 8: Explainability & Trust Authors repeatedly link the adoption of AI- driven robotics to transparency, interpretability, and operator trust – particularly in settings where failures have severe consequences. Alt et al. [134] argue that explainability mechanisms, together with suitable user interfaces for presenting expla- nations, are key enablers for practical uptake of AI in industry. For safety- and risk-sensitive deployments, the limited adoption of perception models is explicitly attributed to the difficulty of understanding and trusting model outputs under potential failure consequences, motivating explainable approaches beyond simple classifica- tion settings [102]. Trust is also discussed as a determinant of acceptance in collaborative work, with concerns that fear of failure can reduce pro- ductivity and that appropriate trust levels are important success factors [58, 103, 125]. Related literature additionally emphasises psychological and ergonomic aspects – operators should feel comfortable and safe, and interfaces should pro- vide visibility into system status and feedback about actions and responses [106, 112]. Empiri- cal observations further suggest that limitations in sensing and flexibility can negatively affect perceived safety and trust in industrial collabora- tion [116]. Therefore, explainability and trust form an implication: industrial systems are expected not only to perform well, but also to provide comprehensible accounts of their behaviour and status to support acceptance, oversight, and safe collaboration [117, 134]. Implication 9: Sensor Fusion & Perception Authors frequently describe perception as a cen- tral capability limiting automation scope, particu- larly in shared industrial workspaces and variable environments. Seleem et al. [52] explicitly state that no single sensing modality is sufficient to handle varying complex situations in real time, 23 motivating multi-modal sensing and fusion. This includes combining internal and external sensors for detecting humans and obstacles [52, 131] and using complementary sensing such as vision and tactile feedback to extend perception and improve safe collaboration [92, 107]. The corpus further connects sensor fusion to robustness: occlusions, dust, lighting variation, and sensor degradation can impair single-sensor solutions, and multi- ple works describe fusion or multi-sensor setups as mitigation strategies [105, 135]. At the same time, perception is framed as a prerequisite for real-time HRC, requiring holistic scene under- standing that includes object recognition, environ- ment parsing, and human recognition [93, 107]. Authors also highlight that advanced industrial tasks often require customised toolings and addi- tional sensors, implying that control systems must accommodate heterogeneous and evolving obser- vation spaces [69, 136]. Hence, when implemented on cobots, industry requires control methods that can accommodate the diverse and dynamic modalities arising from the sensor configurations commonly used in conjunction with cobots (cf. Section 6.2). Consequently, sensor fusion and perception emerge as an implication: industrial deployment demands multi-modal, robust percep- tion pipelines that can support safe, reliable oper- ation and collaboration under real-world sensing limitations [36, 99]. Implication 10: Standardised Benchmark- ing & Evaluation Multiple works emphasise the need for replicable, standardised performance assessment to support adoption and comparability across systems. For example, standard test methods are described as necessary so that reported measurements can be reproduced with limited cost and effort [131]. At the same time, Yang et al. [42] identify a lack of unified standards and metrics for complex, component-rich industrial robotic systems, which complicates evaluation and hampers dissemina- tion. Further, Sarathkumar et al. [129] argue more generally for benchmark and protocol develop- ment for intelligent control algorithms, including best practices that improve interoperability and reliability. In manufacturing-focused discussions, there is an explicit expectation for training and evaluation systems that can replay manufactur- ing tasks precisely and provide suitable interfaces, reflecting dissatisfaction with benchmarks that do not reflect industrial task realities [94]. Comple- mentary perspectives highlight the role of virtual- isation and digital twins as an enabler for testing and monitoring industrial systems [68, 104]. In addition, in our own prior work we argue for using multiple, distinct benchmarks to better charac- terise generalisation and distribution shifts rather than relying on a single testbed [137]. Overall, the literature motivates standardised benchmark- ing and evaluation as an implication: credible industrial adoption requires reproducible assess- ments that meaningfully reflect industrial tasks, conditions, and system complexity [94, 129]. Implication 11: Data Requirements & Usage Finally, a consistent theme is that industrial robotics is data-constrained and that learning- based methods must address both data scarcity and deployment gaps. Several works explicitly note limited labelled data in industrial domains and the difficulty of building representative datasets [115, 133]. Related observations include that small datasets limit the strength of conclu- sions for specialised industrial use cases, under- scoring how narrow and expensive data collection can be in practice [102]. In parallel, authors cau- tion that many advanced approaches demand large volumes of training data and substan- tial computational resources, which can be mis- aligned with agile manufacturing contexts [91, 129]. Calder ́on-cordova et al. [100] also discuss the importance of bridging simulated and real- world conditions, noting limitations of simulation fidelity and the resulting transfer gap for vision- based inputs. Digital-twin perspectives reinforce that data infrastructure is foundational for indus- trial modelling and services [87]. For future embodied intelligence and mass personalisation, Ren et al. [36] argue for continuous adaptation over time, enabled by automatic data collection and continual learning. In the specific context of vision-language(-action) approaches, the liter- ature highlights the need for large-scale multi- modal data and sim-to-real adaptation strategies to improve transfer and robustness, alongside calls to reduce data and compute requirements for industrial deployment [34, 35]. Finally, data- efficient methods, such as few-shot learning, are explicitly motivated as a practical necessity in 24 industrial settings [138]. Thus, data requirements and usage form a distinct implication: industri- ally relevant learning-based robotics must contend with limited, costly, and heterogeneous data while supporting transfer and adaptation under realistic deployment gaps [34, 100, 133, 138]. Synthesis&relevanceforindustrial assessment of RFMs In summary, the implication corpus does not merely enumerate desirable properties, rather it repeatedly frames them as adoption bottlenecks and as reasons why laboratory-grade learning- based approaches often fail to translate into production environments [91, 94, 116]. The eleven proposed implications above therefore serve as the organising structure for the remainder of this section: they are used to discuss which attributes are required for RFMs to be credible candidates for industrial deployment (Section 7.2), and motivate a later, explicit assessment framework for industrial readiness that is grounded in the recurring requirements identified in the reviewed literature (Section 7.3). 7.2 Attributes of Industry-Grade RFMs To operationalise the eleven industrial implica- tions from Section 7.1 for the RFM paradigm, we performed a second synthesis step that translates the largely qualitative, application-facing observa- tions of the implication corpus into a structured set of model- and deployment-relevant viewpoints. The motivation for this step is that the implica- tions themselves describe what industry expects (e.g. safe collaboration, robustness under uncer- tainty, or real-time feasibility), but do not yet specify where in an RFM-based system such expectations could manifest. We therefore analysed each implication individu- ally through a set of derived attributes that cap- ture the principal degrees of freedom of industrial RFM deployment. This attribute set was designed to be sufficiently broad to cover both implica- tions regarding embodied-operation (e.g. hard- ware, sensing, actions) and organisation (e.g. doc- umentation, compliance), while remaining gran- ular enough to express typical failure modes and deployment bottlenecks discussed across the industrial robotics literature (Section 7.1). In total, we consolidated 47 attributes as shown in Table 7.Using these attributes, we con- ducted an iterative mapping procedure in which we revisited each of the eleven implications and asked, for each attribute, what the implication would demand from an RFM-based robotic sys- tem when that system is viewed through the lens of the respective attribute. This yields 11× 47 = 517 implication-attribute combinations, forming a conceptual matrix in which each cell captures the relevant industrial concerns and constraints for that attribute under the given implication, and their specific interpretation in terms of RFM development and deployment. This matrix-based perspective is particularly important because the implications are interde- pendent: for industrial deployments, safety is not separable from perception, and real-time feasibil- ity is not separable from interaction or robustness. Similarly, issues such as authorisation, adversarial robustness, and fault handling naturally span mul- tiple implications (e.g. safety, robustness, trust, and integration). By enforcing a systematic traver- sal across all implication-attribute pairs, we aimed to reduce the risk of under-representing industrial requirements that might otherwise be subsumed under a single headline implication. The primary outcome of this step is a consoli- dated, implication-grounded view of what it would mean for RFMs to be industry-grade. Rather than proposing a single, monolithic list of require- ments, the attribute set enables a structured discussion of how industrial expectations man- ifest at different layers – ranging from sensing and action interfaces, over adaptation behaviour, to governance-related aspects such as compliance and documentation. At the same time, the full matrix is too extensive to be reproduced verbatim in this work with- out disproportionate volume. We therefore use the attribute set as a compact, transparent inter- mediate representation: it makes explicit which viewpoints were considered during the synthesis of the proposed criteria catalogue (Section 7.3), and it provides traceable coverage of the industrial concerns extracted from the implication corpus (Section 7.1). In the succeeding subsection, these structured considerations are subsequently dis- tilled into an explicit assessment framework for industrial readiness, enabling a consistent eval- uation of existing RFMs against the industrial 25 Table 7: Consolidated list of the 47 attributes considered for each implication. (1) Hardware(13) Security(25) Actions/ Outputs(37) Precision & Accuracy (2) Tasks(14) Detection, Recognition & Understanding (26) Skills(38) Trust (3) Environment(15) Multimodal Inputs/ Outputs (27) Observation/ Input(39) Goal Specification & Interpretation (4) Objects/ Materials(16) Integration(28) External Influences/ Disturbances (40) Solution Strategy (5) Instructability/ Commanding (17) Safety(29) Reaction(41) Offline Adaptation (Finetuning/ Retraining) (6) Distractors/ Distractions(18) Compliance(30) Validation/ Benchmark/ Evaluation (42) Energy (7) Adaptability/ Generalisation/ Flexibility (19) Supervision(31) Realtime Performance & Inference Frequency (43) Data (8) Online Adaptation/ Learning (20) Fault Tolerance & Recovery/ Safeguards (32) Noise(44) Documentation/ Open-Source (9) Internal Interfaces/ Representations (21) Reasoning/ Intention(33) Autonomy(45) (Self-)Awareness/ Assessment & Monitoring (10) ID/ OOD Robustness(22) Model Feedback to User(34) Consistency & Stability(46) Memory (11) Adversarial Robustness(23) External Guidance/ Correction to Model (35) Explainability/ Insights(47) Anomaly/ Fault/ Error Detection & Recognition (12) Authorisation(24) Interaction/ Collaboration (36) Repeatability expectations highlighted throughout the litera- ture. 7.3 Evaluating Industrial Maturity: A Criteria Catalogue The eleven implications in Section 7.1 describe what industrial settings recurrently demand from deployed robotic systems. However, for RFMs, these expectations must be made con- crete at the level of model capabilities, interfaces, and the surrounding deployment stack. Build- ing on the 517 implication-attribute combina- tions from Section 7.2, we therefore translated the implication-grounded requirements into an explicit criteria catalogue that can be used to reason about, compare, and ultimately rate the industrial readiness of existing and future RFMs. For each implication-attribute combination, we collected and refined the key questions that an industrial deployment would have to answer (e.g. what happens under distribution shifts, how fail- ures are detected and handled, how human input is interpreted and verified, and which artefacts are required for audit and integration). We iteratively consolidated these questions into a catalogue of 149 criteria (C) – shown in the Appendix, Table C4 – spanning all eleven general industrial implications (I). While comprehensive, we do not consider this catalogue final: it is intended as a liv- ing foundation that can be extended as additional industrial constraints or regulatory expectations emerge. Criteria scope: RFMs & their deployment ecosystem Although most criteria target RFMs directly, industrial readiness is not determined by the model alone. Industrial deployments require an ecosystem – robot software stack, interfaces, safety infrastructure, monitoring, logging, and documen- tation – that must satisfy similar constraints. Consequently, the catalogue includes criteria that (a) are primarily RFM-centric (e.g. adaptive behaviour under changing observations/actions), (b) are primarily ecosystem-centric (e.g. the pres- ence and maintenance of accessible documenta- tion on training data composition and transfer behaviour), or (c) require joint support by model and ecosystem (e.g. a model producing struc- tured signals that can be persistently logged and leveraged by the surrounding system). Intentional redundancy across implications Because the industrial implications are inter- dependent (cf. Section 7.2), certain criteria 26 re-appear deliberately across different implica- tions, albeit with different emphasis. We there- fore avoided merging superficially similar criteria across implications when doing so would remove an aspect that is essential for assessing matu- rity with respect to a particular implication. For example, hot-swapping appears under adaptabil- ity (I1-C2), integration/cost (I7-C4), and percep- tion (I9-C12), but each instance stresses a differ- ent industrial concern: autonomous adjustment of observation/action spaces, recognition and han- dling of various interface changes during commis- sioning, or modularity of the observation space alone and its documentation. Retaining these cri- teria within their respective implications preserves the ability to diagnose why an RFM is imma- ture in a specific industrial dimension rather than merely concluding that it ”lacks hot-swapping support”. Key paradigms reflected in the catalogue Many of the catalogue’s central motifs are rarely discussed jointly in prior work and often not yet implemented in current RFMs, however, most are already being investigated as distinct research directions scattered across the RFM, VLA/LLM or general robotics control landscapes. We there- fore use the following paradigms as anchor points to illustrate that a substantial subset of the cat- alogue’s requirements is directly supported by emerging literature, and that our contribution lies in consolidating these partially disconnected strands and enriching them from an industrial viewpoint to craft a single, implication-grounded assessment framework. Central to criteria on multi-hardware support, hot-swapping, and dimensionality adaptation is the treatment of changing observation and action interfaces. Recent work has begun to explicitly adapt VLAs to new embodiments by learning unified action representations and steering pre- trained policies towards platform-specific action distributions during downstream adaptation [139] or by enabling architectures to handle differ- ing sized action and observation spaces [140]. These approaches align closely with our empha- sis that industrial deployment requires transfer across morphologies, whereas a natural extension emerges in runtime resilience to interface changes and sensor/actuator substitutions. Changing action- and observation spaces nat- urally motivate concerns about handling OOD scenarios. Challenging industrial environments are indeed characterised by disturbances, clutter, and changes that are not exhaustively covered during training, requiring robustness against distribu- tion shifts, anomalies, and deliberate manipula- tion. Accordingly, our catalogue places repeated emphasis on online detection, confidence sig- nalling, and fallback behaviour. Complementary to this, emerging evidence suggests that current VLA systems may inherit security weaknesses known from LLMs and VLMs, amplifying their consequences. Both textual attacks that can per- sist over long horizons and vision-side attacks (e.g. through disrupting patch exposure) have been shown to substantially degrade performance or even obtain absolute control authority of VLAs, highlighting that adversarial robustness cannot be treated as an optional add-on for physically embodied systems. [141, 142] In parallel, anomaly prediction has been explored as a proactive mech- anism, for instance by forecasting action outcomes and anticipating whether they lead to anoma- lous states [143]. This combination motivates why our catalogue repeatedly requires threat/attack detection, anomaly handling, and explicit recovery pathways: in industrial settings, models must dif- ferentiate between benign novelty and adversarial or safety-critical deviations. These mechanisms also depend on the system’s ability to quantify and communicate uncertainty in a usable manner. We therefore include multiple confidence-related criteria (spanning operational readiness, safety margin, real-time compliance, perceptual certainty, and velocity/ precision & accuracy (v/P&A) fulfilment), reflecting the view that trust requires calibrated estimates rather than raw success rates. This is supported by work that explore confidence calibration in VLAs for uncertainty quantification [144], as well as methods that align uncertainty to trigger help- seeking behaviour with statistical guarantees for LLM-based planning settings [145]. Such capabil- ities enable explicit halting, logging, clarification requests, or escalation to supervisors when the sys- tem’s own predicted reliability degrades. Confidence also naturally links to interactive clar- ification and feedback, repeatedly mentioned in our catalogue, which are essential for industrial 27 HRI/HRC where instructions can be ambigu- ous, incomplete, or safety-constrained. Recent approaches explicitly reason about linguistic ambiguity grounded in the scene and gener- ate targeted follow-up questions, thereby reduc- ing downstream execution failures caused by NL command-misinterpretation [146]. Conversely, other works demonstrate that NL can serve as an intuitive channel for fine-grained trajectory mod- ification and incremental corrections, suggesting practical routes towards mid-execution trajectory adjustability and operator-in-the-loop refinement as demanded by several criteria [147, 148]. More broadly, collaboration frameworks that incorpo- rate expert interventions during deployment indi- cate that interactive supervision can simultane- ously improve immediate reliability and collect data for future refinement [143, 149]. A further recurring motif within our catalogue is the need for explicit supervisory or verification layers that monitor execution quality and safety independently of the core policy. This direction is reflected by Zhou et al. [150] that leverage VLMs for reactive and proactive failure detec- tion, and methods designed for general failure detection across multiple tasks rather than nar- row, task-specific detectors [151], which is essential for task-agnostic policies. Related efforts target the gap between semantic failure understanding and actionable correction by explicitly bridging high-level reflection and low-level action adjust- ment [152]. Such results substantiate the repeated insistence on decoupled safety gating, redun- dant safety mechanisms, fault categorisation, and explainable recovery behaviour: industrial assur- ance typically requires independent checks, pre- dictable intervention pathways, and evidence of consistent handling across task variation. Finally, several works reinforce why the catalogue treats memory, data governance, and documen- tation as recurring concerns. Methods that add lightweight memory modules to VLAs highlight that exploiting historical context can materially improve long-horizon performance [153], while other frameworks explicitly summarise prior expe- rience to support failure reasoning and replan- ning [154]. At the same time, privacy-preserving training mechanisms for VLAs [155] imply that industrial adoption may be constrained by data protection requirements, not by model capabil- ity alone. In addition, dataset construction efforts that deliberately incorporate failures and recov- eries imply that training data must cover various edge- or failure-cases to enable robust deployment behaviour [156]. These developments collectively support the inclusion of criteria on persistent logging, memory mechanisms, data composition variation and transparency, privacy compliance, and learning from faults and incidents. Taken together, these strands illustrate that many catalogue requirements correspond to active research directions – the industrial immaturity is less the absence of isolated technical proposals than the lack of their systematic combination into a deployable, auditable, and safety-governed RFM ecosystem. The criteria catalogue therefore con- solidates these partially disconnected lines into a unified framework that is directly grounded in the implication corpus and tailored to the operational constraints of industrial deployment, serving as a structured bridge between industrial expec- tations and RFM assessment: it translates the implication-level requirements into concrete items that can be used to analyse where current RFMs are industrial mature, where they are incomplete, and which research directions are likely neces- sary for credible industrial uptake. In the survey section that follows, we use this framework as a reference lens to discuss existing RFMs and to highlight gaps between current capabilities and the implication-derived expectations. 8 RFM Overview & Evaluation This section provides a structured overview of the current landscape of manipulation-capable RFMs and assesses their industrial readiness using the implication-grounded criteria catalogue introduced in Section 7.3. For transparency and overview, we report the full corpus of all 324 sur- veyed models in the Appendix (Table D5). We evaluate each model against the 149 catalogue criteria, enabling both an overall maturity score per paper and implication-wise maturity profiles. We describe the resulting evaluation methodology and procedure in Section 8.1, including the design decisions required to scale the assessment to the full corpus, corresponding to 324×149=48,276 criterion-level decisions. We additionally report an agreement and con- sistency check of our evaluation pipeline against 28 expert judgements on a small ground-truth sub- set in Section 8.1. Finally, we summarise the strongest-performing RFMs (Section 8.2) and provide a broader industrial applicability analy- sis over the full corpus and the Top-50 subset (Section 8.3). The combination of those results is used to characterise the current industrial matu- rity of RFMs, highlight which industrial implica- tions are most and least addressed, and finally pro- vide an evidence-based baseline for the synthesis and outlook in Section 9. 8.1 Evaluation Procedure Toassessindustrialreadinessacrossthe manipulation-capable RFM landscape, we eval- uated all 324 manipulation-focused or generalist RFMs contained in the models sub-corpus of our main RFM-related corpus (out of 341 total model publications; cf. Section 3.2). Each paper was assessed against the 149 criteria of our implication-grounded catalogue (cf. Section 7.3, Appendix D5), yielding 324 × 149 = 48, 276 criterion-level decisions. A fully manual expert annotation of this scale is infeasible within a fixed time frame and would furthermore restrict the analysis to a small, subjectively filtered subset, limiting the validity of corpus-level conclusions. We therefore adopted an automated, LLM-based evaluation pipeline to (a) enable a complete- corpus assessment, (b) reduce reviewer-dependent variability via a fixed and mostly repeatable pro- cedure, and (c) obtain both a global maturity score per paper (fraction of fulfilled criteria) and implication-wise maturity profiles (fraction of fulfilled criteria per implication). Our evaluation was implemented as an objec- tive, structured pipeline that leverages an LLM (GPT-5.1 in our case) to extract evidence from the full text of a paper and to decide whether a single criterion is fulfilled by the proposed model and/or its deployment ecosystem (as defined by the criterion; cf. Section 7.3). A key design goal was to minimise cross-contamination between papers and between criteria. Accordingly, each single evaluation run was performed under two strict isolation constraints: (a) the LLM received exactly one criterion at a time (and was not informed about any other criteria), and (b) the retrieval context contained exactly one paper (including figures), provided through a dedicated vector store that was created for that evaluation run only. After the decision for that criterion- paper pair, the vector store was deleted, the session was closed, and the next criterion-paper pair was evaluated in a fresh run. This enforced independence across the 48, 276 evaluations and prevented leakage of information or implicit calibration across different models or criteria. Importantly, all papers were processed using private, Siemens-internal infrastructure. In particular, no papers were uploaded to public servers, and the evaluation was executed on pri- vate systems where the processed documents are not used for any training/ learning. This ensured that the large-scale assessment did not create any disclosure or copyright-related risks. For every paper-criterion pair, the LLM was required to produce a structured output with the following fields: a binary judgement of criterion fulfilment, a flag indicating whether the property was directly stated, a confidence score in the range [0, 1] indicating explicitness and evidence strength, a boolean evidence-availability signal, textual extracted quotes (or an explicit state- ment that no relevant information was found), and a brief rationale connecting evidence to the decision. Importantly, we instructed the LLM to decide conservatively: if a criterion was not supported clearly and in-scope by the paper, it was to be marked as not fulfilled. This makes the resulting maturity estimates intentionally cautious, particularly for criteria that are only partially addressed, or for capabilities that are not sufficiently documented in the publication. The complete corpus run required over 5 days under rate limits, further underscoring why a purely manual evaluation is practically unattainable at this scale. Expert Agreement Comparison Because the pipeline is used to draw corpus- level conclusions, we evaluated its reliability against expert judgements on a small ground- truth subset. Specifically, we created three blind expert-rated catalogue evaluations: two models were annotated by an expert prior to running the automated pipeline, and one additional model was annotated after the automated run had com- pleted (the best-scoring model according to the LLM results). These three papers serve as a check on whether the procedure generates results with 29 sufficient accuracy to draw corpus-level conclu- sions. Table 8 shows absolute values, specifying how many criteria per paper-implication pair have been marked as fulfilled by expert and LLM. The table reveals that the LLM-based procedure closely reproduces the implication-level patterns of the expert assessments for these representative cases. In particular, the qualitative conclusion is preserved across all three models: Gem- ini Robotics 1.5 [157] attains substantially higher industrial maturity than OpenVLA [18], while π 0 [26] remains limited across most implications. Acrossthethreeexpert-annotatedpapers, Table 9 reports strong criterion-level agreement (N = 447): the LLM reaches an accuracy of 0.966 and an F 1 score of 0.762 for the positive (ful- filled) class, with substantial agreement beyond chance (Cohen’s κ = 0.744). Crucially for the maturity estimator, the false-positive rate is low (FPR = 1.9%), indicating that the LLM rarely over-attributes industrial capabilities that the expert does not confirm. The false-negative rate is higher (FNR = 22.6%), consistent with our conservative decision policy. Consequently, the derived maturity values should be interpreted as a cautious, lower-bound estimate rather than an optimistic readiness score. Restricting the analysis to those cases where the LLM reports informa- tion found as true, (Table 9, InfoFound ) does not increase the F 1 score and slightly reduces κ. This supports our intention that the information-found flag primarily captures evidence availability in the text, rather than serving as a reliable filter for correctness. At the per-paper level (Table 10), agreement between expert and LLM is high for both Gemini Robotics 1.5 and OpenVLA, especially supported by strong κ and J + (Jaccard similarity for fulfilled cases). For π 0 , the expert positive rate at only 2.7% is notably low, with the correspond- ing F 1 , κ and J + substantially lower than for the other two papers. This underlines that even a small number of disagreements can dominate positive-class metrics when positives are rare, however, also explains the higher false negative rates reported in Table 9. Therefore, per-paper metrics should be read with care in very low- positivity cases. Nevertheless, lower agreement cases as reported for π 0 do not necessarily reflect substantial judg- ment errors, rather often report interpretation differences: e.g. Black et al. [26] include a brief evaluation of language-following capabilities and report an increased task success rate when pro- viding intermediate language commands to the policy. The expert acknowledged this as sufficient for I1-C11 (”Mid-Execution Feedback Respon- siveness”), whereas the LLM argued that only an increase in success rate does not necessarily imply the requested ability of course corrections. In fact, we observed a few patterns within our criteria that may inflate reported disagreement, while not necessarily introducing hard errors. As stated for the above example, some criteria may leave some room for interpretation, even among experts: regarding the ”Low-Data Adaptation & Few-Shot Requirements” criterion (I11-C1), OpenVLA reports requiring on average 10-150 episodes for downstream task adaptation [18] – the LLM interpreted this as few-shot and marked the criterion fulfilled, whereas our expert judgement considered up to 150 episodes too demanding for an industrial-ready ”low-data adaptation” interpretation. This implies that some criteria may need stricter boundaries to ensure higher evaluator agreement. In contrast, other criteria are inherently stringent (e.g. formu- lations that imply guarantees such as ”always”), which leads the LLM to often mark such criteria as not fulfilled, even when the capability is basi- cally implemented, but the paper does not state a near 100% guarantee of adherence. Taken together, the agreement comparison indicates that the proposed pipeline provides sufficiently reliable, conservative estimates for our goal of scaling industrial maturity assessment to hundreds of RFMs. The low over-claiming rate supports its use for comparing models and identifying implication-level blind spots in the current state of the art, while the elevated miss rate and the interpretative latitude of some crite- ria motivate caution when analysing individual, fine-grained criterion outcomes. Accordingly, in the following, we treat criterion-level results as indicative and focus primarily on implication-level trends and cross-corpus patterns. 8.2 Highest Rated RFMs Across the 324 evaluated RFMs (cf. Table D5), the overall highest-rated models, as depicted in Table 11, reach total maturity scores of only 30 Table 8: Comparison between expert- and LLM-based evaluation results per implication for three representative RFMs. Gemini Robotics 1.5 [157]OpenVLA [18]π 0 [26] ImplicationExpertLLMExpertLLMExpertLLM I1983411 I2110000 I3110000 I4111100 I5000000 I6000000 I7000000 I8110000 I9120010 I10235511 I11210111 P 181891143 Table 9: LLM-expert agreement on criterion fulfillment. Reporting the number of N evaluations considered, the overall Accuracy (Acy), F 1 -score, Cohen’s kappa agreement (κ) and false positive/ negative rates (FPR/FNR). SettingNAcyF1κFPR/FNR All447 0.966 0.762 0.744 1.9%/22.6% InfoFound 304 0.951 0.762 0.734 2.9%/22.6% Table 10: Per-paper agreement. Pos% reports the expert positive rate and J + the Jaccard similarity index, computed only for the positive class to ignore the dominating negatives. PaperPos%AcyF1κJ + Gemini 1.5 [157] 0.121 0.960 0.833 0.810 0.714 OpenVLA [18]0.060 0.973 0.800 0.786 0.667 π 0 [26]0.027 0.966 0.286 0.269 0.167 0.11 − 0.12, i.e. they satisfy roughly one out of ten catalogue criteria on average. Although possibly lower-bound, this pronounced ceiling effect is informative in itself: even the strongest current approaches cover only a narrow subset of the industrial requirements captured by our implication-grounded framework, and thus no sin- gle model can yet be regarded as comprehensively industry-grade. A commonality among the five best models is that they do not merely present a monolithic visuo- motor policy, but explicitly introduce and focus on single mechanisms that address failure modes and operational bottlenecks frequent to industrial settings: structured reasoning, human-in-the-loop guidance, self-correction, and improved percep- tion/data utilisation. Accordingly, their maturity profiles exhibit pronounced peaks on a small num- ber of implications while remaining low elsewhere (cf. Table 11 and Figure 10). This implies that recent RFM work seems to advance one or two enabling dimensions at a time, but do not yet integrate the broader industrial requirements cap- tured by our catalogue. Specifically, Gemini Robotics 1.5 [157] achieves the highest total score (0.12) and the strongest value on adaptability/flexibility (I1: 0.44). A plau- sible reason for this scoring is that it combines two complementary components (Gemin Robotics 1.5 as a VLA, and Gemini Robotics 1.5-ER as a VLM), which can be combined to an agentic framework, rather than relying on a single con- troller in isolation. From an industrial perspective, this implies that near-future deployments may 31 Table 11: Overview of the five highest rated RFMs (total) of the 324 evaluated. Reporting their proportional industrial applicability ratings for each individual implication, and a total score over all implications. Per-criterion ratings are created utilising the publication’s context, by leveraging our LLM-based pipeline as described in Section 8.1. Highest values within the five RFMs for each implication are shown in green, lowest in red. Values are rounded to two decimal places. RefModelTotal↓I1I2I3I4I5I6I7I8I9I10I11 [157] Gemini Robotics 1.5 0.120.440.060.070.070.000.000.000.070.140.230.09 [158] Collab VLA 0.11 0.280.000.200.070.080.110.000.070.210.080.09 [159] GF- VLA 0.11 0.280.000.000.140.080.000.000.070.140.230.27 [6] OneTwo VLA 0.110.33 0.000.200.140.000.110.100.070.140.000.09 [160] SC- VLA 0.110.33 0.060.000.210.230.110.000.000.000.150.09 I3 I2 I1 I11 I10 I9 I8 I7 I6 I5 I4 0.10.20.30.4 Implication (I) Fulfillment GF-VLA SC-VLA OneTwo VLA Collab VLA Gemini Robotics 1.5 Fig. 10: Proportional implication fulfillment. Depicting the five overall highest-rated RFMs (see Table 11 for detailed values). demand layered systems where high-level reason- ing, verification, and interaction are distinct from low-level action generation. As Gemini’s leading score is not driven by uniformly higher values across all implications, it suggests that hybrid or multi-component designs may only be one part of the puzzle towards industrial maturity, primarily aiding flexibility. In general, we noticed a strong consistency in positive evaluator agreement across the 5 high- est ranked RFMs: CollabVLA [158] attains the highest scores within this group for HRI/HRC (I3: 0.20) and sensor fusion/perception (I9: 0.21). Those implications highly correlate, since percep- tion of human/ operator detection is a prerequisite for HRC (cf. Section 7.1). Importantly, this con- centration directly matches the paper’s explicit objective of moving from a closed-loop visuo- motor policy towards a collaborative agent by incorporating human guidance into action genera- tion [158]. An equally strong agreement is obvious in data requirements and usage (I11: 0.27) for GF-VLA [159]. The authors focus on learning from human demonstrations aiming to improve gen- eralisation beyond low-level trajectory imitation, framing the efficient utilisation of demonstration data as their main goal [159]. Similar to Gemini, OneTwoVLA [6] shows high flexibility (I1: 0.33) and also employs more than one component, how- ever, explicitly distinguishes between a fast acting mode (System One) and a slower reasoning mode (System Two) and switches between them dur- ing execution [6]. SC-VLA [160]’s high values on robustness/reliability (I4: 0.21) and precision/ac- curacy (I5: 0.23) are equally well explained by the model’s main intention: Li et al. [160] integrate a fast action prediction with a slower reflec- tion mechanism to correct failed actions for more robust manipulation. Such failure recovery could naturally translate to a more robust system. 32 Table 12: Overview of the highest rated RFMs per implication of the 324 total evaluated. Per-criterion ratings are created utilising the publication’s context, by leveraging our LLM-based pipeline as described in Section 8.1. Highest values within the respective implication are highlighted in green. Values are rounded to two decimal places. The entry ”Multiple” represents the fact that several works have the similar highest score of 0.11 for I6, which are thus consolidated in one row. RefModel↑TotalI1I2I3I4I5I6I7I8I9I10I11 [158] Collab VLA 0.110.280.000.200.070.080.110.000.07 0.210.080.09 [161]FuSe0.050.060.000.000.000.000.000.000.000.210.000.27 [157] Gemini Robotics 1.5 0.120.440.060.070.070.000.000.000.070.140.230.09 [162] GraphCoT VLA 0.050.060.000.000.000.000.000.00 0.210.140.000.09 [163]HAMSTER 0.070.110.000.000.000.000.000.100.000.070.230.36 [164] HiBer NAC 0.100.220.110.000.000.080.110.100.00 0.210.150.09 [165]LCIL0.070.110.000.270.000.000.110.000.000.000.080.18 [166]NanoVLA 0.070.110.000.000.070.080.000.200.000.000.310.00 [18]OpenVLA 0.070.220.000.000.070.000.000.000.000.000.380.09 [167] PaLM- E 0.090.330.000.000.290.000.000.000.000.070.000.18 [160] SC- VLA 0.110.330.060.000.21 0.230.110.000.000.000.150.09 [143] Sirius- Fleet 0.100.390.220.000.140.000.000.000.000.000.080.09 [168] Smol VLA 0.040.060.000.000.070.000.00 0.200.000.000.150.00 [169]UniVLA0.070.110.000.000.000.080.000.000.070.000.230.36 –Multiple–0.11– Taken together, Table 11 and Table 12 illus- trate a consistent trend: the strongest models predominantly optimise for one or two identi- fied bottlenecks, and this intent is reflected by implication-specific peaks in the evaluation. How- ever, even the best per-implication results remain limited in absolute terms: the highest proportional fulfilment observed across all implications is 0.44 by Gemini Robotics 1.5 on I1, and this is the only value over all evaluated RFMs to exceed 0.4 at all (cf. Table D5). This indicates that, while individual industrial concerns are beginning to be addressed, current RFMs do not yet attempt – nor achieve – integrated coverage across the full set of industrial implications. In the following subsection, we therefore move beyond the top- ranked models to provide a broader view of how industrial maturity is distributed across the full RFM landscape, and which implications remain systematically underrepresented. 8.3 Industrial Applicability Overview To provide a corpus-level perspective beyond the highest-ranked models (Section 8.2), we anal- ysed the distribution of implication-wise fulfilment across all 324 evaluated RFMs (cf. Table D5) and the stronger Top-50 model subset, leveraging the total and per-implication industrial matu- rity scores. We report two complementary views: 33 (a) a brief temporal stability overview of which implications are addressed (Figure 11), and (b) breadth-versus-depth of coverage per implication for the full corpus and the Top-50 (Figure 12, with exact values in Table 13). Finally, we contextualise these trends by inspecting which individual crite- ria are most frequently fulfilled (Table 14) and by quantifying whether criteria are widely discussed yet rarely met (Table 15). Temporal Stability Overview Figure 11 shows the normalised, per-implication I3 I2 I1 I11 I10 I9 I8 I7 I6 I5 I4 0.51.01.52.0 Yearly Criterion Fulfillment Ratio 2020 2021 2022 2023 2024 2025 Fig. 11: Normalised per-implication criteria cov- erage. Each value represents the sum over all criteria fulfilled within the given implication & year, divided by the total number of papers pub- lished in the respective year. criteria coverage aggregated by publication year. Across the years depicted, the overall pro- file is highly consistent: the same small subset of implications dominates the addressed space, while the remaining implications remain per- sistently low. In particular, the corpus repeat- edly concentrates on adaptability/flexibility (I1), benchmarking/evaluation (I10), and data require- ments/usage (I11), whereas safety/compliance (I2), HRI/HRC (I3), real-time performance (I6), and cost-effectiveness/integration (I7) remain comparatively underrepresented. The absence of a clear, monotonic rise in any single implica- tion suggests that current RFM research is not (yet) converging towards progressively more holis- tic industrial coverage at the implication level – rather, the field appears to deepen within a relatively stable set of research emphases. Breadth vs. Depth Analysis A central question to our industrial applicability analysis is, whether the literature does not men- tion an implication at all, merely mentions it, or provides substantive, multi-criterion coverage. We therefore contrast two metrics per implication: (a) 1 + -Coverage, the share of papers meeting at least one criterion, representing breadth of impli- cation considerations, and (b) Average Criteria Coverage (ACC), the average fraction of criteria met per paper, implying consideration depth. The combined view of Figure 12 and Table 13, reveals a consistent breadth-depth mismatch across the current RFM landscape: While many papers touch a subset of implications, far fewer demonstratemulti-criterion,implementation- relevant coverage within those implications (note scale differences of Figure 12a and Figure 12b). Concretely, the full corpus exhibits strong breadth for only three implications: Adaptability/flexibility (I1: (1 + =0.756)), bench- marking/evaluation (I10: (1 + =0.753)), and data requirements/usage (I11: (1 + =0.707)). Yet even for these leading implications, depth remains low: their ACC values stay below (0.10) (I10: (0.099), I11: (0.097), I1: (0.089)). This pat- tern indicates that on average, papers satisfy only a small fraction of the criteria within the implications they most frequently address. Addi- tionally, several implications central to industrial deployment remain already weak at the 1 + - Criterion level – most notably safety/compliance (I2: (1 + =0.034)), HRI/HRC (I3: (1 + =0.052)), real-time performance (I6: (1 + =0.049)), and cost- effectiveness/integration (I7: (1 + =0.071)). The persistently low coverage for these implications aligns with the temporal stability observation in Figure 11: the research field’s emphasis appears comparatively stable over time and does not show a clear shift towards holistic industrial require- ment satisfaction at the implication level. Restricting attention to the Top-50 models strengthens the same overall picture rather than changing it. Breadth becomes near-universal for 34 I3 I2 I1 I11 I10 I9 I8 I7 I6 I5 I4 0.20.40.60.81.0 1+ Coverage Full Corpus Top 50 (a) 1 + -Coverage over evaluated RFMs: The metric spec- ifies the percentage of works that meet at least one criterion for the respective implication (I) within the given (sub)corpus. I3 I2 I1 I11 I10 I9 I8 I7 I6 I5 I4 0.050.100.150.20 Average Criteria Coverage Full Corpus Top 50 (b) ACC for evaluated RFMs: ACC specifies the average percentage of criteria each paper meets for the respective implication (I) within the given (sub)corpus. Fig. 12: Implication-specific evaluation coverage across the full set of 324 RFMs (Full Corpus) and the Top-50 subset. Table 13: Detailed comparison of 1 + -Coverage and ACC between the full evaluated corpus (324) and the Top 50: For visualisation, see Figure 12. Values are rounded to three decimal places. MetricCorpusI1I2I3I4I5I6I7I8I9I10I11 1+ Full0.7560.0340.0520.1720.1670.0490.0710.2190.2130.7530.707 Top 501.00.220.341.01.00.320.461.01.01.01.0 ACC Full0.0890.0030.0070.0160.0140.0050.0080.0170.0190.0990.097 Top 500.2380.0190.0430.0960.0860.0360.050.0810.0970.2320.227 seven implications (I1, I4, I5, I8, I9, I10, I11 all at (1 + =1.0)), suggesting that higher-performing RFMs increasingly include at least some features, experiments, or claims related to these dimen- sions. However, the remaining implications still lag even in the Top-50 subset (I2: (0.22), I3: (0.34), I6: (0.32), I7: (0.46)). Depth improves sub- stantially but remains far from comprehensive: the highest Top-50 ACC values are again concen- trated in I1 (0.238), I10 (0.232), and I11 (0.227), whereas safety/compliance (I2: 0.019) and real- time performance (I6: 0.036) remain particularly low. Thus, even among the strongest papers, industrial maturity is characterised more by selective advances than by broad, high-coverage satisfaction across implications. Table 14 sharpens this conclusion by showing which capabilities are most frequently met. The ten most commonly fulfilled criteria cluster heav- ily in the same three implications that dominate in depth (I1, I10, I11). The top entries – Multi- Task Generalisation (I1: 0.4938), Flexible Data Type Acceptance (I11: 0.4877), Repeatability & Reproducibility (I10: 0.4475), and Low-Data Adaptation & Few-Shot Requirements (I11: 0.4198) – indicate that the present RFM liter- ature prioritises generality across tasks, flexible use of available (and heterogeneous) data, and 35 Table 14: Top 10 most frequently fulfilled criteria across all papers. Reporting each criterion’s implication (I1-I11), absolute count and rate of fulfillment, explicitness among fulfilled cases (Expl(fulf )), evidence availability (Evid), and mean confidence when fulfilled (ConfFul). Values are rounded to four decimal places. CriterionImplFulfilledRate↓Expl(fulf )EvidConfFul Multi-Task GeneralisationI11600.49380.28120.99380.8605 Flexible Data Type AcceptanceI111580.48770.36710.99070.8693 Repeatability & ReproducibilityI101450.44750.24830.98150.8474 Low-Data Adaptation & Few-Shot Requirements I111360.41980.80880.99070.9102 Module & Skill-Specific BenchmarkingI101290.39810.06980.99380.8517 Generalisation to Environmental VariationI11010.31170.54460.98150.8739 Generalisation & Adaptation AssessmentI10900.27780.11111.00000.8676 Autonomous Handling of Distractors/Obstacles I1870.26850.50570.95680.8655 Performance Benchmarking & OOD Feedback I8630.19440.20630.98770.8367 Online/Mid-Execution AdaptationI1490.15120.53060.99070.8771 Table 15: Evidence-Fulfilled-Gap analysis: Reporting the gap score g c alongside its evidence availability (EvidRate) and full-corpus fulfillment rate (FulfillRate) for each implication’s criterion with the highest evidence rate. Values are rounded to four decimal places. Impl↑Criteriong c EvidRate↓FulfillRate I1Multi-Task Generalisation0.50000.99380.4938 I2Reactive, Real-Time Safety0.80860.81170.0031 I3Collaborative Adaptation0.87350.89810.0247 I4Environmental Robustness & Adaptation0.94750.99070.0432 I5Consistency Over Time & Contexts0.87960.99690.1173 I6Goal-Responsive Solution Strategies0.93520.97220.0370 I7Flexible Input/Output Adaptation0.98150.98150.0000 I8Performance Benchmarking & OOD Feedback0.79320.98770.1944 I9Multimodal Sensor Acceptance & Fusion0.96601.00000.0340 I10Generalisation & Adaptation Assessment0.72221.00000.2778 I11Low-Data Adaptation & Few-Shot Requirements0.57100.99070.4198 benchmark-driven reporting. At the same time, the explicitness values for several of these fre- quently fulfilled criteria are relatively low: e.g. Multi-Task Generalisation (0.2812) and Repeata- bility & Reproducibility (0.2483), suggesting that fulfilment is often inferred from experimental set-ups or evaluation sections rather than being framed as a directly stated requirement. The comparatively high explicitness for Low-Data Adaptation (0.8088) is consistent with this topic being frequently highlighted explicitly as a con- tribution, though – as discussed in Section 8.1 – the interpretation latitude of this criterion may also inflate fulfilment rates. In general, however, the most fulfilled criteria support what we would expect of RFMs: systems adaptive to multiple tasks that can digest various forms of data for behaviour modulation. Finally, Table 15 indicates that many criteria are discussed in the current literature, however, 36 most seem to be not, or only partially fulfilled. The gap score g c (EvidRate c − FulfillRate c ) is large for the top-evidence criterion in almost every implication, frequently exceeding (0.8). As EvidRate reflects the rate of evidence found in the given texts, these large gaps imply that many criteria-relevant aspects are widely mentioned or at least documentable, yet rarely satisfied under the (binary, conservative) fulfilment policy of our evaluation. In combination with the low ACC values in Table 13, this points to a broader tendency: the corpus increasingly acknowledges industrial constraints and desiderata, but typical papers still implement or document only lim- ited, criterion-incomplete solutions – particularly for implications tied to deployment integra- tion (I7), safety/ compliance (I2), real-time performance (I6), and HRC/HRI (I3). 9 Summary & Discussion Industrial robotics is undergoing a discernible transition from isolated, highly optimised automa- tion towards flexible, collaborative deployment paradigms in which humans and robots share workspaces and responsibilities. A driver of this shift is the uptake of cobots, particularly in high- mix, low-volume settings and in SMEs, where traditional automation is often economically or organisationally impractical due to commission- ing effort, frequent production changes, and the need for specialised expertise (cf. Section 5). While modern cobots provide the hardware foundations for such a transition – safe physical design, appro- priate payload and sizing, comparatively low oper- ating power – they do not, by themselves, resolve the central bottleneck: robust, deployment-ready intelligence that can flexibly adapt behaviour to changing tasks, environments, and multimodal input without imposing prohibitive engineering overhead. Our synthesis therefore frames RFMs as a promis- ing enabling technology for this industrial trans- formation, because they target precisely the dimensions in which classical control and even spe- cialised learning-based approaches remain limited: multi-task competence, cross-domain generalisa- tion, and instruction- or demonstration-driven/ low-effort reconfiguration (cf. Section 2). How- ever, the same factors that make industrial cobot deployments attractive – shared workspaces, con- tinuous operation, heterogeneous actuation/ sens- ing stacks and cost pressure – also amplify the demands on any learning-based control paradigm. In particular, typical cobots exhibit high mechan- ical precision and repeatability, moderate payload capability, and low median power draw in the widely deployed 0.5-20kg class (cf. Section 6.1). These characteristics imply that industrial value is frequently limited not by actuation quality but by perception and decision-making under real-world variability, and they simultaneously constrain RFM deployment to low-latency, energy-aware inference on industrial edge platforms rather than workstation-grade hardware (cf. Section 6.3). The practical prevalence of multi-sensor configura- tions, sensor fusion, and modular end-effectors further highlights that industrial intelligence must be interface-flexible: it must tolerate changing observation and action spaces that arise from tool changes, sensor substitutions, and integra- tion into existing automation infrastructure (cf. Section 6.2). To capture these deployment realities systematically, we distilled eleven general interde- pendent implications (Section 7.1) from industrial literature and translated them into a structured assessment framework (Section 7.3). The key contribution of this step is not merely a list of desirable properties, but a traceable oper- ationalisation: As foundation for our assessment framework, we started with 11 implications of what industry generally demands from robotic applications, extracted 47 deployment-relevant attributes to create an implication-attribute matrix of 11× 47 = 517 considerations, to finally retrieve an implication-grounded, assessable cri- teria catalogue of 149 concrete items spanning both model capabilities and ecosystem require- ments (cf. Section 7). Our observations clearly show that raw success rates alone are not suffi- cient for industrial application. Rather, capability is inseparable from deployability: safety depends on perception and verification, real-time feasibil- ity depends on compute constraints and system integration, and trust depends on observable and auditable behaviour rather than raw benchmark success. When applying this framework to the manipulation-capable RFM landscape, our evalu- ation results paint a consistent picture: the field is progressing rapidly, e.g. in research breadth, 37 yet remains at an early stage of industrial matu- rity (cf. Section 8). Across the 324 discovered and evaluated models and 48, 276 criterion-level decisions, even the top-ranked RFMs reach only 0.11-0.12 overall maturity – roughly one out of ten criteria met on average (cf. Section 8.2) and most models only score (comparably) high on a maximum of two to three varying impli- cations. This implies that the primary limiting factor for industrial deployment is not the absence of isolated technical ideas, but the incomplete and non-systematic integration of those ideas into comprehensive, deployable RFM stacks. The frontier models further indicate how progress is currently realised. The five strongest sys- tems exhibit implication-wise ”peaks” rather than balanced coverage, typically by introducing tar- geted mechanisms such as structured reasoning, human-in-the-loop interaction, self-correction, or improved data utilisation (cf. Section 8.2). Gem- ini Robotics 1.5, for example, leads overall and reaches the highest observed implication fulfil- ment on adaptability/flexibility (I1: 0.44), yet remains near zero on multiple deployment-critical implications, underscoring that hybrid or lay- ered architectures may advance flexibility without automatically closing safety, integration, or real- time gaps. Similarly, models that score compara- tively well in HRI/HRC and perception emphasise collaborative mechanisms and richer observation handling, but do not simultaneously satisfy the broader criteria spectrum required for sustained industrial operation. Taken together, the cur- rent RFM state of the art appears to advance by isolating and optimising individual bottle- necks, yet comprehensive ”industry-grade” inte- gration, by contrast, remains largely unaddressed. This interpretation is reinforced when expand- ing evaluation to the full corpus. Over time, the implication profile remains strikingly stable: research consistently concentrates on adaptabili- ty/flexibility (I1), benchmarking/evaluation (I10), and data requirements/usage (I11), while safety/- compliance (I2), HRI/HRC (I3), real-time per- formance (I6), and cost-effectiveness/integration (I7) remain persistently underrepresented (cf. Section 8.3). The breadth-depth analysis of our evaluation sharpens the same conclusion: In the full corpus, several implications are frequently ”touched”, yet few papers satisfy more than a small fraction of criteria within those implications. Even the Top-50 subset does not convert their breadth into comprehensive satisfaction: depth remains modest even for leading implications (I1/I10/I11), and many deployment-critical impli- cations (I2/I3/I6/I7) remain relatively weak both in breadth and depth (cf. Section 8.3). Hence, current progress is better characterised as wider acknowledgement of industrial concerns than as demonstrable, multi-criterion completion of those concerns. These findings suggest several concrete priorities for future work towards industry-grade RFMs: First, the field must move from isolated capa- bility demonstrations towards deployable, layered systems in which planning, memory, action gen- eration, verification, and safety gating are treated as central, independently testable components – an architectural direction implicitly supported by high-performing models, yet not broadly matured across the corpus (cf. Section 8.2). Second, indus- trial adoption may require bridging the per- sistent underrepresentation of safety/compliance, real-time feasibility, and integration concerns by developing evaluation protocols and artefacts that connect RFM capabilities to operational con- straints: low-latency inference on edge hardware, structured logging and monitoring, and explicit handling of changing sensor and tooling interfaces (cf. Sections 6.3 and 7.3). Third, the strong and stable emphasis on benchmarking (I10) should be leveraged to improve industrial relevance rather than to merely report success rates in a diversity of benchmarks: evaluations must reflect deploy- ment realities such as distribution shifts, failure recovery, human supervision, and long-horizon operation, and must enable diagnosing why a model fails under industrial conditions rather than only whether it succeeds on curated tasks. Finally, the large evidence-fulfilment gaps (cf. Section 8.3) indicate an opportunity and a requirement: many industrial desiderata are already recognised by the research community, yet the step from recog- nition to robust implementation, documentation, and validation remains a dominant obstacle. In conclusion, RFMs constitute a credible tech- nological pathway to unlock the full potential of cobot-centred automation: they directly tar- get the flexibility, reconfigurability, and intuitive instructioning demanded by modern industrial 38 production, especially in cost- and changeover- sensitive settings (cf. Section 5). Nevertheless, our corpus-scale evaluation demonstrates that current RFM research is still predominantly laboratory- oriented when measured against deployment- grounded industrial requirements. The presented implication synthesis and criteria catalogue pro- vide a transparent industrial readiness reference that can be used both to evaluate future RFMs and to guide their design towards auditable, safe, real-time-capable and integration-ready systems. Closing the identified gaps – particularly in safe- ty/compliance, HRI/HRC, real-time performance, and cost-effectiveness/integration capabilities – therefore represents the central research agenda for translating the rapidly advancing field of RFMs into reliable industrial autonomy. References [1] Chi, C., Feng, S., Du, Y., Xu, Z., Cousineau, E., Burchfiel, B., et al.: Diffusion Pol- icy: Visuomotor Policy Learning via Action Diffusion (2023) https://doi.org/10.48550/ arXiv.2303.04137 [2] Liang, Z., Li, Y., Yang, T., Wu, C., Mao, S., Nian, T., et al.: Discrete Diffu- sion VLA: Bringing Discrete Diffusion to Action Decoding in Vision-Language-Action Policies. arXiv (2025). https://doi.org/10. 48550/arXiv.2508.20072 [3] Zhang, W., Liu, H., Qi, Z., Wang, Y.,Yu,X.,Zhang,J.,etal.: DreamVLA:AVision-Language-Action ModelDreamedwithComprehen- sive World Knowledge. arXiv (2025). https://doi.org/10.48550/arXiv.2507.04447 [4] Liang, J., Liu, R., Ozguroglu, E., Sudhakar, S., Dave, A., Tokmakov, P., et al.: Dreami- tate: Real-World Visuomotor Policy Learn- ing via Video Generation. arXiv (2024). https://doi.org/10.48550/arXiv.2406.16862 [5] Bu, Q., Li, H., Chen, L., Cai, J., Zeng, J., Cui, H., et al.: Towards Synergistic, Generalized, and Efficient Dual-System for Robotic Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2410.08001 [6] Lin, F., Nai, R., Hu, Y., You, J., Zhao, J., Gao, Y.: OneTwoVLA: A Unified Vision- Language-Action Model with Adaptive Rea- soning. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.11917 [7] Chen, H., Liu, J., Gu, C., Liu, Z., Zhang, R., Li, X., et al.: Fast-in-Slow: A Dual-System Foundation Model Unifying Fast Manipula- tion within Slow Reasoning. arXiv (2025). https://doi.org/10.48550/arXiv.2506.01953 [8] Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., Arx, S., et al.: On the Opportunities and Risks of Foundation Models. arXiv (2022). https://doi.org/10. 48550/arXiv.2108.07258 [9] Firoozi, R., Tucker, J., Tian, S., Majumdar, A., Sun, J., Liu, W., et al.: Foundation Mod- els in Robotics: Applications, Challenges, and the Future. arXiv (2023). https://doi. org/10.48550/arXiv.2312.07843 [10] Yuan, Y.: On the Power of Foundation Mod- els (2022) https://doi.org/10.48550/arXiv. 2211.16327 [11] Marsh, G.E.: THE DEMYSTIFICATION OF EMERGENT BEHAVIOR (2013) https: //doi.org/10.48550/arXiv.0907.1117 [12] Hu, Y., Xie, Q., Jain, V., Francis, J., Patrikar, J., Keetha, N., et al.: Toward General-Purpose Robots via Foun- dation Models: A Survey and Meta- Analysis. arXiv (2024). https://doi.org/10. 48550/arXiv.2312.08782 [13] Ma, Y., Song, Z., Zhuang, Y., Hao, J., King, I.: A Survey on Vision-Language-Action Models for Embodied AI. arXiv (2025). https://doi.org/10.48550/arXiv.2405.14093 [14] Xiang, T.-Y., Jin, A.-Q., Zhou, X.-H., Gui, M.-J., Xie, X.-L., Liu, S.-Q., et al.: Paral- lels Between VLA Model Post-Training and Human Motor Learning: Progress, Chal- lenges, and Trends. arXiv (2025). https:// doi.org/10.48550/arXiv.2506.20966 [15] Zhou, H., Yao, X., Mees, O., Meng, Y., Xiao, 39 T., Bisk, Y., et al.: Bridging Language and Action: A Survey of Language-Conditioned Robot Manipulation. arXiv (2025). https:// doi.org/10.48550/arXiv.2312.10807 [16] Xu, C., Zhang, S., Liu, Y., Sun, B., Chen, W., Xu, B., et al.: An Anatomy of Vision- Language-Action Models: From Modules to Milestones and Challenges. arXiv (2025). https://doi.org/10.48550/arXiv.2512.11362 [17] Zhong, Y., Bai, F., Cai, S., Huang, X., Chen, Z., Zhang, X., et al.: A Survey on Vision- Language-Action Models: An Action Tok- enization Perspective. arXiv (2025). https: //doi.org/10.48550/arXiv.2507.01925 [18] Kim, M.J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., et al.: OpenVLA: An Open-Source Vision- Language-Action Model. arXiv (2024). https://doi.org/10.48550/arXiv.2406.09246 [19] Li, X., Liu, M., Zhang, H., Yu, C., Xu, J., Wu, H., et al.: Vision-Language Foun- dation Models as Effective Robot Imi- tators. arXiv (2024). https://doi.org/10. 48550/arXiv.2311.01378 [20] Amin, A., Aniceto, R., Balakrishna, A., Black, K., Conley, K., Connors, G., et al.: π0∗.6: A VLA That Learns From Experience (2025) https://doi.org/10. 48550/arXiv.2511.14759 [21] Xiao, W., Xue, H., Tao, T., Kalaria, D., Dolan, J.M., Shi, G.: AnyCar to Anywhere: Learning Universal Dynamics Model for Agile and Adaptive Mobility. arXiv (2024). https://doi.org/10.48550/arXiv.2409.15783 [22] Zhang, J., Li, A., Qi, Y., Li, M., Liu, J., Wang, S., et al.: Embodied Navigation Foundation Model. arXiv (2025). https:// doi.org/10.48550/arXiv.2509.12129 [23] Cheng, A.-C., Ji, Y., Yang, Z., Gongye, Z., Zou, X., Kautz, J., et al.: NaVILA: Legged Robot Vision-Language-Action Model for Navigation. arXiv (2025). https://doi.org/ 10.48550/arXiv.2412.04453 [24] Fuentes, L.F.M., Khan, M.H., Cabrera, M.A., Serpiva, V., Iarchuk, D., Mahmoud, Y., et al.: VLH: Vision-Language-Haptics Foundation Model. arXiv (2025). https:// doi.org/10.48550/arXiv.2508.01361 [25] Gu, J., Wu, Z., Si, P., Qiu, S., Feng, Y., Sun, L., et al.: USIM and U0: A Vision- Language-Action Dataset and Model for General Underwater Robots. arXiv (2025). https://doi.org/10.48550/arXiv.2510.07869 [26] Black, K., Brown, N., Driess, D., Esmail, A., Equi, M., Finn, C., et al.: $π 0$: A Vision- Language-Action Flow Model for General Robot Control. arXiv (2024). https://doi. org/10.48550/arXiv.2410.24164 [27] Black, K., Brown, N., Darpinian, J., Dhabalia, K., Driess, D., Esmail, A., et al.: $π 0.5$: A Vision-Language- Action Model with Open-World General- ization. arXiv (2025). https://doi.org/10. 48550/arXiv.2504.16054 [28] Brohan, A., Brown, N., Carbajal, J., Cheb- otar, Y., Dabis, J., Finn, C., et al.: RT-1: Robotics Transformer for Real-World Con- trol at Scale. In: Robotics: Science and Sys- tems XIX. Robotics: Science and Systems Foundation, ??? (2023). https://doi.org/10. 15607/RSS.2023.XIX.025 [29] Brohan, A., Brown, N., Carbajal, J., Cheb- otar, Y., Chen, X., Choromanski, K., et al.: RT-2: Vision-Language-Action Mod- els Transfer Web Knowledge to Robotic Control. arXiv (2023). https://doi.org/10. 48550/arXiv.2307.15818 [30] Shao, R., Li, W., Zhang, L., Zhang, R., Liu, Z., Chen, R., et al.: Large VLM-based Vision-Language-Action Models for Robotic Manipulation: A Survey. arXiv (2025). https://doi.org/10.48550/arXiv.2508.13073 [31] Cui, C., Ding, P., Song, W., Bai, S., Tong, X., Ge, Z., et al.: OpenHelix: A Short Survey, Empirical Analysis, and Open- Source Dual-System VLA Model for Robotic Manipulation. arXiv (2025). https://doi. org/10.48550/arXiv.2505.03912 40 [32] Zhang, D., Sun, J., Hu, C., Wu, X., Yuan, Z., Zhou, R., et al.: Pure Vision Language Action (VLA) Models: A Comprehensive Survey. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.19012 [33] Liu, H., Li, X., Li, P., Liu, M., Wang, D., Liu, J., et al.: Towards Generalist Robot Policies: What Matters in Building Vision-Language-Action Models. research- square (2025). https://doi.org/10.21203/rs. 3.rs-5770637/v1 [34] Sapkota, R., Cao, Y., Roumeliotis, K.I., Karkee, M.: Vision-Language-Action Mod- els: Concepts, Progress, Applications and Challenges. arXiv (2025). https://doi.org/ 10.48550/arXiv.2505.04769 [35] Fan, J., Yin, Y., Wang, T., Dong, W., Zheng, P., Wang, L.: Vision-language model-based human-robot collaboration for smart manufacturing: A state-of-the-art sur- vey. Frontiers of Engineering Management 12(1), 177–200 (2025) https://doi.org/10. 1007/s42524-025-4136-9 [36] Ren, L., Dong, J., Liu, S., Zhang, L., Wang, L.: Embodied Intelligence Toward Future Smart Manufacturing in the Era of AI Foun- dation Model. IEEE/ASME Transactions on Mechatronics, 1–11 (2024) https://doi. org/10.1109/tmech.2024.3456250 [37] EN ISO 10218-1:2011 (2011) [38] Guertler, M., Tomidei, L., Sick, N., Carmichael, M., Paul, G., Wambsganss, A., et al.: WHEN IS A ROBOT A COBOT?MOVINGBEYONDMAN- UFACTURINGANDARM-BASED COBOT MANIPULATORS. Proceedings of the Design Society 3, 3889–3898 (2023) https://doi.org/10.1017/pds.2023.390 [39] Olszewski, M.: Modern Industrial Robotics. Pomiary Automatyka Robotyka 24(1), 5–20 (2020) https://doi.org/10.14313/PAR 235/ 5 [40] Borboni, A., Reddy, K.V.V., Elamvazuthi, I., AL-Quraishi, M.S., Natarajan, E., Azhar Ali, S.S.: The Expanding Role of Artificial Intelligence in Collaborative Robots for Industrial Applications: A Sys- tematic Review of Recent Works. Machines 11(1), 111 (2023) https://doi.org/10.3390/ machines11010111 [41] Amiri, P., M ̈uller, M., Southgate, M., Theodoridis, T., Wei, G., Richards-Brown, M., et al.: A Statistical Analysis of Com- mercial Articulated Industrial Robots and Cobots. Journal of Manufacturing and Materials Processing 8(5), 216 (2024) https: //doi.org/10.3390/jmmp8050216 [42] Yang, M., Yang, E., Zante, R.C., Post, M., Liu, X.: Collaborative mobile industrial manipulator: A review of system architec- ture and applications. In: 2019 25th Interna- tional Conference on Automation and Com- puting (ICAC), p. 1–6. IEEE, Lancaster, United Kingdom (2019). https://doi.org/10. 23919/iconac.2019.8895183 [43] Robotics, I.I.F.: International Federation of Robotics. https://ifr.org/industrial-robots [44] Muru, J., Rass ̃olkin, A.: A Scoping Review of Energy Consumption in Industrial Robotics. Machines 13(7), 542 (2025) https://doi.org/10.3390/machines13070542 [45] Brinker, J., Funk, N., Ingenlath, P., Takeda, Y., Corves, B.: Comparative Study of Serial- Parallel Delta Robots With Full Orientation Capabilities. IEEE Robotics and Automa- tion Letters 2(2), 920–926 (2017) https:// doi.org/10.1109/LRA.2017.2654551 [46] Autsou, S., Rass ̃olkin, A., Vaimann, T., Kudelina, K.: Analysis of possible faults and diagnostic methods of the Cartesian industrial robot. Proceedings of the Esto- nian Academy of Sciences 71(3), 227 (2022) https://doi.org/10.3176/proc.2022.3.04 [47] Collins, F., Yim, M.: Design of a spherical robot arm with the Spiral Zipper prismatic joint. In: 2016 IEEE International Confer- ence on Robotics and Automation (ICRA), p. 2137–2143. IEEE, Stockholm, Swe- den (2016). https://doi.org/10.1109/ICRA. 41 2016.7487363 [48] Suri, S., Jain, A., Verma, N., Prasert- poj, N.: SCARA Industrial Automation Robot. In: 2018 International Conference on Power Energy, Environment and Intelli- gent Control (PEEIC), p. 173–177. IEEE, Greater Noida, India (2018). https://doi. org/10.1109/PEEIC.2018.8665440 [49] Khatib, O., Siciliano, B. (eds.): Springer Handbook of Robotics, 2nd ed. 2016 edn.SpringerHandbooks.Springer, Cham(2016).https://doi.org/10.1007/ 978-3-319-32552-1 [50] Taesi, C., Aggogeri, F., Pellegrini, N.: COBOT Applications—Recent Advances and Challenges. Robotics 12(3), 79 (2023) https://doi.org/10.3390/robotics12030079 [51] Johannsmeier, L., Haddadin, S.: A Hier- archical Human-Robot Interaction-Planning Framework for Task Allocation in Collabo- rative Industrial Assembly Processes. IEEE Robotics and Automation Letters 2(1), 41– 48 (2017) https://doi.org/10.1109/lra.2016. 2535907 [52] Saleem, Z., Gustafsson, F., Furey, E., McAfee, M., Huq, S.: A review of exter- nal sensors for human detection in a human robot collaborative environment. Journal of Intelligent Manufacturing 36(4), 2255–2279 (2025) https://doi.org/10.1007/ s10845-024-02341-2 [53] Tan, M., Liu, Z., Wang, P.: Application and Optimization of Robot Systems in Industrial Production Line. In: 2024 Second Inter- national Conference on Inventive Comput- ing and Informatics (ICICI), p. 701–706. IEEE, Bangalore, India (2024). https://doi. org/10.1109/icici62254.2024.00119 [54] Colgate,J.E.,Wannasuphoprasit,W., Peshkin, M.A.: Cobots: Robots for Col- laboration With Human Operators. In: Dynamic Systems and Control, p. 433– 439. American Society of Mechanical Engineers, Atlanta, Georgia, USA (1996). https://doi.org/10.1115/IMECE1996-0367 [55] Peshkin, M.A., Colgate, J.E., Wannasupho- prasit, W., Moore, C.A., Gillespie, R.B., Akella, P.: Cobot architecture. IEEE Trans- actions on Robotics and Automation 17(4), 377–390 (2001) https://doi.org/10.1109/70. 954751 [56] Khedr, M., Yang, E.: An overview of cobots for advanced manufacturing: Human- robot interactions and research trends. MATEC Web of Conferences 401, 12005 (2024) https://doi.org/10.1051/matecconf/ 202440112005 [57] Aziz, A.N.A., Yap, H.J., Sekaran, S.C., Ismail, A.H., Hisaburi, A.S.M., Tan, C.H.: Integration of Collaborative Robots in AerospaceManufacturing:Enhancing Precision and Safety through Advanced Sensor and Communication Technologies. In: 2024 IEEE 12th Conference on Sys- tems, Process & Control (ICSPC), p. 177–182. IEEE, Malacca, Malaysia (2024). https://doi.org/10.1109/ICSPC63060.2024. 10862208 [58] Kopp, T., Baumgartner, M., Kinkel, S.: Success factors for introducing industrial human-robot interaction in practice: An empirically driven framework. The Interna- tional Journal of Advanced Manufacturing Technology 112(3-4), 685–704 (2021) https: //doi.org/10.1007/s00170-020-06398-0 [59] Javaid, M., Haleem, A., Singh, R.P., Rab, S., Suman, R.: Significant applications of Cobots in the field of manufacturing. Cog- nitive Robotics 2, 222–233 (2022) https:// doi.org/10.1016/j.cogr.2022.10.001 [60] Kakade, S., Patle, B., Umbarkar, A.: Appli- cations of collaborative robots in agile man- ufacturing: A review. Robotic Systems and Applications 3(1), 59–83 (2023) https://doi. org/10.21595/rsa.2023.23238 [61] George, D.A.S., A.S.Hovan George: The Cobot Chronicles: Evaluating the Emer- gence, Evolution, and Impact of Collab- orative Robots in Next-Generation Man- ufacturing (2023) https://doi.org/10.5281/ ZENODO.8021406 42 [62] El Zaatari, S., Marei, M., Li, W., Usman,Z.:Cobotprogrammingfor collaborativeindustrialtasks:An overview.RoboticsandAutonomous Systems 116,162–180(2019)https: //doi.org/10.1016/j.robot.2019.03.003 [63] Sen, R., Lohar, A., Panwar, A., Singh, A., Sahu, A., Gurjer, A., et al.: COBOT (COLLABORATIVE ROBOT) IN MAN- UFACTURINGINDUSTRIESAND ASSEMBLY LINE. International Journal of Technical Research & Science 9(Spl), 27–35(2024)https://doi.org/10.30780/ specialissue-ISET-2024/047 [64] Saleem, M., Jameel, S., Hassan, A., Khan, Z.H.: Design and Fabrication of a Six DOF Gantry Robot for Weld- ing Applications. In: 2024 ASU Interna- tional Conference in Emerging Technologies for Sustainability and Intelligent Systems (ICETSIS), p. 1940–1946. IEEE, Man- ama, Bahrain (2024). https://doi.org/10. 1109/ICETSIS61505.2024.10459371 [65] Yang, X., Zhou, Z., Sørensen, J.H., Chris- tensen, C.B., ̈ Unalan, M., Zhang, X.: Automation of SME production with a Cobot system powered by learning-based vision. Robotics and Computer-Integrated Manufacturing 83, 102564 (2023) https:// doi.org/10.1016/j.rcim.2023.102564 [66] Sahan, A.S.M., Kathiravan, S., Lokesh, M., Raffik, R.: Role of Cobots over Indus- trial Robots in Industry 5.0: A Review. In: 2023 2nd International Conference on Advancements in Electrical, Electronics, Communication, Computing and Automa- tion (ICAECA), p. 1–5. IEEE, Coimbat- ore, India (2023). https://doi.org/10.1109/ ICAECA56562.2023.10201199 [67] Lefranc, G., Lopez-Juarez, I., Osorio- Compar ́an, R., Pe ̃na-Cabrera, M.: Impact of Cobots on automation. Procedia Computer Science 214, 71–78 (2022) https://doi.org/ 10.1016/j.procs.2022.11.150 [68] Grau, A., Indri, M., Lo Bello, L., Sauter, T.: Robots in Industry: The Past, Present, and Future of a Growing Collaboration With Humans. IEEE Industrial Electronics Mag- azine 15(1), 50–61 (2021) https://doi.org/ 10.1109/MIE.2020.3008136 [69] Dzedzickis,A.,Subaˇci ̄ut ̇e- ˇ Zemaitien ̇e, J., ˇ Sutinys, E., Samukait ̇e-Bubnien ̇e, U., Buˇcinskas, V.: Advanced Applications of Industrial Robotics: New Trends and Possi- bilities. Applied Sciences 12(1), 135 (2021) https://doi.org/10.3390/app12010135 [70] Barnett, N., Costenaro, D., Rohmund, I.: Direct and Indirect Impacts of Robots on Future Electricity Load. In: 2017 ACEEE Summer Study on Energy Efficiency In Industry (2017) [71] Uhlmann, E., Reinkober, S., Hollerbach, T.: Energy Efficient Usage of Industrial Robots for Machining Processes. Procedia CIRP 48, 206–211 (2016) https://doi.org/10.1016/j. procir.2016.03.241 [72] Gadaleta, M., Berselli, G., Pellicciari, M., Grassia, F.: Extensive experimental inves- tigation for the optimization of the energy consumption of a high payload indus- trial robot with open research dataset. Robotics and Computer-Integrated Manu- facturing 68, 102046 (2021) https://doi.org/ 10.1016/j.rcim.2020.102046 [73] Li, P., Liu, X.: Common Sensors in Industrial Robots: A Review. Journal of Physics: Conference Series 1267(1), 012036 (2019) https://doi.org/10.1088/1742-6596/ 1267/1/012036 [74] Dang,T.,Nguyen,K.,Huber,M.: ExtPerFC: An Efficient 2D and 3D PerceptionHardware-SoftwareFrame- work for Mobile Cobot. arXiv (2023). https://doi.org/10.48550/arXiv.2306.04853 [75] Scholz, C., Cao, H.-L., Imrith, E., Roshan- del, N., Firouzipouyaei, H., Burkiewicz, A., et al.: Sensor-Enabled Safety Sys- tems for Human–Robot Collaboration: A Review. IEEE Sensors Journal 25(1), 65–88 (2025) https://doi.org/10.1109/JSEN.2024. 3496905 43 [76] BuytheLatestJetsonProd- ucts |NVIDIADeveloper. https://developer.nvidia.com/buy-jetson (2025) [77] JetsonT5000ModulesDatasheet DS11945001 v1.1. NVIDIA (2025) [78] NVIDIA RTX Blackwell GPU Architecture v1.1. NVIDIA (2025) [79] Arrow Electronics: Connect with Elec- tronicComponents |Arrow.Com. https://w.arrow.com/en (2025) [80] Samanta, A., Hatai, I., Mal, A.K.: A Survey on Hardware Accelerator Design of Deep Learning for Edge Devices. Wireless Personal Communications 137(3), 1715–1760 (2024) https://doi.org/10.1007/ s11277-024-11443-2 [81] Woisetschl ̈ager, H., Erben, A., Wang, S., Mayer, R., Jacobsen, H.-A.: Federated Fine- Tuning of LLMs on the Very Edge: The Good, the Bad, the Ugly. In: Proceedings of the Eighth Workshop on Data Management for End-to-End Machine Learning, p. 39– 50. ACM, Santiago A Chile (2024). https: //doi.org/10.1145/3650203.3663331 [82] Yi, S., Hao, Z., Qin, Z., Li, Q.: Fog Computing: Platform and Applications. In: 2015 Third IEEE Workshop on Hot Topics in Web Systems and Technologies (HotWeb), p. 73–78. IEEE, Washington DC, DC, USA (2015). https://doi.org/10. 1109/HotWeb.2015.22 [83] Atlam, H., Walters, R., Wills, G.: Fog Computing and the Internet of Things: A Review. Big Data and Cognitive Comput- ing 2(2), 10 (2018) https://doi.org/10.3390/ bdcc2020010 [84] Boss, G., Malladi, P., Quan, D., Legregni, L.:CLOUD COMPUTING.Technical report, IBM (October 2007) [85] Ismail, B.I., Khalid, M.F., Kandan, R., Ahmad, H., Mohd Mydin, M.N., Hong Hoe, O.: Cobot Fleet Management System Using Cloud and Edge Computing. In: 2020 IEEE 7th International Conference on Engi- neering Technologies and Applied Sciences (ICETAS), p. 1–5. IEEE, Kuala Lumpur, Malaysia (2020). https://doi.org/10.1109/ ICETAS51660.2020.9484266 [86] Mystkowska, G., Monopoli, M., Nan- nipieri, P., Zulberti, L., Merodio Codi- nachs, D., Fanucci, L.: Hardware Plat- forms Enabling Edge AI for Space Appli- cations: A Critical Review. IEEE Access 13, 143939–143956 (2025) https://doi.org/ 10.1109/ACCESS.2025.3596326 [87] Huang, Z., Shen, Y., Li, J., Fey, M., Brecher, C.: A Survey on AI-Driven Digital Twins in Industry 4.0: Smart Manufacturing and Advanced Robotics. Sensors 21(19), 6340 (2021) https://doi.org/10.3390/s21196340 [88] Herrero, H., Out ́on, J., Puerto, M., Sall ́e, D., L ́opez De Ipi ̃na, K.: Enhanced Flexibil- ity and Reusability through State Machine- Based Architectures for Multisensor Intelli- gent Robotics. Sensors 17(6), 1249 (2017) https://doi.org/10.3390/s17061249 [89] Waseem,M.,Chang,Q.:Enhancing FlexibilityinSmartManufacturing: SystemPropertyEnabledMultiagent Approach for Mobile Robot Scheduling inMultiproductFlexibleManufac- turingSystems.arXiv(2025).https: //doi.org/10.48550/ARXIV.2504.20053 [90] Tellaeche, A., Maurtua, I., Ibarguren, A.: Human robot interaction in industrial robotics. Examples from research centers to industry. In: 2015 IEEE 20th Confer- ence on Emerging Technologies & Factory Automation (ETFA), p. 1–6. IEEE, Lux- embourg, Luxembourg (2015). https://doi. org/10.1109/etfa.2015.7301650 [91] Fan, H., Liu, X., Fuh, J.Y.H., Lu, W.F., Li, B.: Embodied intelligence in manufac- turing: Leveraging large language models for autonomous industrial robotics. Jour- nal of Intelligent Manufacturing 36(2), 1141–1157 (2025) https://doi.org/10.1007/ s10845-023-02294-y 44 [92] Mohammadi Amin, F., Rezayati, M., Van De Venn, H.W., Karimpour, H.: A Mixed- Perception Approach for Safe Human– Robot Collaboration in Industrial Automa- tion. Sensors 20(21), 6347 (2020) https:// doi.org/10.3390/s20216347 [93] Bonci, A., Cen Cheng, P.D., Indri, M., Nabissi, G., Sibona, F.: Human-Robot Per- ception in Industrial Environments: A Sur- vey. Sensors 21(5), 1571 (2021) https://doi. org/10.3390/s21051571 [94] Liu, Z., Liu, Q., Xu, W., Wang, L., Zhou, Z.: Robot learning towards smart robotic manufacturing: A review. Robotics and Computer-Integrated Manufacturing 77, 102360 (2022) https://doi.org/10.1016/ j.rcim.2022.102360 [95] Liu, C., Tomizuka, M.: Robot Safe Inter- action System for Intelligent Industrial Co- Robots. arXiv (2018). https://doi.org/10. 48550/arXiv.1808.03983 [96] Abuzar Yaqoob Mulla, Gauri Ansurkar: AI and Robotics: Designing Intelligent and Adaptive Robots for Industrial Automation. International Journal of Scientific Research in Science and Technology, 817–824 (2023) https://doi.org/10.32628/ijsrst523102126 [97] Villani, V., Pini, F., Leali, F., Sec- chi, C., Fantuzzi, C.: Survey on Human- Robot Interaction for Robot Program- ming in Industrial Applications. IFAC- PapersOnLine 51(11), 66–71 (2018) https: //doi.org/10.1016/j.ifacol.2018.08.236 [98] Li, K., Wang, J., Wu, X., Peng, X., Chang, R., Deng, X., et al.: Optimizing Automated Picking Systems in Warehouse Robots Using Machine Learning. arXiv (2024). https:// doi.org/10.48550/arXiv.2408.16633 [99] Wunderle, Y., Lyczkowski, E.: Sensor fusion for functional safety of autonomous mobile robots in urban and industrial environ- ments. In: 2022 IEEE 27th International Conference on Emerging Technologies and Factory Automation (ETFA), p. 1–4. IEEE, Stuttgart, Germany (2022). https:// doi.org/10.1109/etfa52439.2022.9921688 [100] Calder ́on-Cordova, C., Sarango, R., Castillo, D., Lakshminarayanan, V.: A Deep Rein- forcement Learning Framework for Control of Robotic Manipulators in Simulated Envi- ronments. IEEE Access 12, 103133–103161 (2024) https://doi.org/10.1109/access.2024. 3432741 [101] Zhang, Z., Qian, K., Schuller, B.W., Woll- herr, D.: An Online Robot Collision Detec- tion and Identification Scheme by Super- vised Learning and Bayesian Decision Theory. IEEE Transactions on Automa- tion Science and Engineering 18(3), 1144– 1156 (2021) https://doi.org/10.1109/tase. 2020.2997094 [102] Andres, A., Martinez-Seras, A., La ̃na, I., Del Ser, J.: On the black-box explainabil- ity of object detection models for safe and trustworthy industrial applications. Results in Engineering 24, 103498 (2024) https:// doi.org/10.1016/j.rineng.2024.103498 [103] Rodriguez-Guerra,D.,Sorrosal,G., Cabanes, I., Calleja, C.: Human-Robot Interaction Review: Challenges and Solu- tions for Modern Industrial Environments. IEEE Access 9, 108557–108578 (2021) https: //doi.org/10.1109/access.2021.3099287 [104] Othman, U., Yang, E.: Human–Robot Col- laborations in Smart Manufacturing Envi- ronments: Review and Outlook. Sensors 23(12), 5663 (2023) https://doi.org/10. 3390/s23125663 [105] Halme, R.-J., Lanz, M., K ̈am ̈ar ̈ainen, J., Pieters, R., Latokartano, J., Hietanen, A.: Review of vision-based safety systems for human-robot collaboration. Procedia CIRP 72, 111–116 (2018) https://doi.org/ 10.1016/j.procir.2018.03.043 [106] Robla-Gomez, S., Becerra, V.M., Llata, J.R., Gonzalez-Sarabia, E., Torre-Ferrero, C., Perez-Oria, J.: Working Together: A Review on Safe Human-Robot Collabo- ration in Industrial Environments. IEEE Access 5, 26754–26773 (2017) https://doi. 45 org/10.1109/access.2017.2773127 [107] Nahavandi, S., Alizadehsani, R., Nahavandi, D., Lim, C.P., Kelly, K., Bello, F.: Machine Learning Meets Advanced Robotic Manip- ulation. arXiv (2023). https://doi.org/10. 48550/arXiv.2309.12560 [108] Duijkeren, N., Palmieri, L., Lange, R., Kleiner, A.: An Industrial Perspective on Multi-Agent Decision Making for Inter- operable Robot Navigation Following the VDA5050 Standard. arXiv (2023). https:// doi.org/10.48550/arXiv.2311.14615 [109] Landi, C.T., Ferraguti, F., Costi, S., Bonfe, M., Secchi, C.: Safety Barrier Functions for Human-Robot Interaction with Indus- trial Manipulators. In: 2019 18th European Control Conference (ECC), p. 2565–2570. IEEE, Naples, Italy (2019). https://doi.org/ 10.23919/ecc.2019.8796235 [110] Lins, R.G., Givigi, S.N.: Cooperative RoboticsandMachineLearningfor Smart Manufacturing: Platform Design andTrendsWithintheContextof IndustrialInternetofThings.IEEE Access 9,95444–95455(2021)https: //doi.org/10.1109/access.2021.3094374 [111] Liu, S., Wang, L., Wang, X.V.: Sensor- less force estimation for industrial robots using disturbance observer and neural learn- ing of friction approximation. Robotics and Computer-Integrated Manufacturing 71, 102168 (2021) https://doi.org/10.1016/ j.rcim.2021.102168 [112] Prati, E., Villani, V., Grandi, F., Peruzzini, M., Sabattini, L.: Use of Interaction Design Methodologies for Human–Robot Collabo- ration in Industrial Scenarios. IEEE Trans- actions on Automation Science and Engi- neering 19(4), 3126–3138 (2022) https:// doi.org/10.1109/tase.2021.3107583 [113] Nakhaeinia, D., Laferriere, P., Payeur, P., Laganiere, R.: Safe Close-Proximity and Physical Human-Robot Interaction Using Industrial Robots. In: 2015 12th Confer- ence on Computer and Robot Vision, p. 237–244. IEEE, Halifax, NS, Canada (2015). https://doi.org/10.1109/crv.2015.38 [114] Hjorth, S., Chrysostomou, D.: Human– robotcollaborationinindustrial environments: A literature review on non-destructivedisassembly.Robotics andComputer-IntegratedManu- facturing73,102208(2022)https: //doi.org/10.1016/j.rcim.2021.102208 [115] De Simone, G., Greco, A., Rosa, F., Saggese, A., Vento, M.: Context-aware data aug- mentation for enhanced speech command recognition in industrial environments. Sci- entific Reports 15(1) (2025) https://doi. org/10.1038/s41598-025-01886-3 [116] Weiss, A., Huber, A., Minichberger, J., Ikeda, M.: First Application of Robot Teach- ing in an Existing Industry 4.0 Envi- ronment: Does It Really Work? Societies 6(3), 20 (2016) https://doi.org/10.3390/ soc6030020 [117] Patr ́ıcio, L., Varela, L., Silveira, Z., Felgueiras, C., Pereira, F.: A Frame- work for Integrating Robotic Process Automation with Artificial Intelligence Applied to Industry 5.0. Applied Sci- ences 15(13), 7402 (2025) https://doi.org/ 10.3390/app15137402 [118] Schreiter, T., Rudenko, A., R ̈uppel, J.V., Magnusson, M., Lilienthal, A.J.: Multi- modal Interaction and Intention Communi- cation for Industrial Robots. arXiv (2025). https://doi.org/10.48550/arXiv.2502.17971 [119] Manonmani, A., Akash, S., Aswinraj, A., Manikandan, A.: Literature Review on Advanced Cloud Robotics for Enhanced Industrial Automation. In: 2025 Inter- national Conference on Electronics and Renewable Systems (ICEARS), p. 46–51. IEEE, Tuticorin, India (2025). https://doi. org/10.1109/icears64219.2025.10940892 [120] Triantafyllou, P., Afonso Rodrigues, R., Chaikunsaeng, S., Almeida, D., Deacon, G., Konstantinova, J., et al.: A Methodology for Approaching the Integration of Complex 46 Robotics Systems: Illustration Through a Bimanual Manipulation Case Study. IEEE Robotics & Automation Magazine 28(2), 88–100 (2021) https://doi.org/10.1109/mra. 2021.3064759 [121] Quarta, D., Pogliani, M., Polino, M., Maggi, F., Zanchettin, A.M., Zanero, S.: An Exper- imental Security Analysis of an Industrial Robot Controller. In: 2017 IEEE Sympo- sium on Security and Privacy (SP), p. 268–286. IEEE, San Jose, CA, USA (2017). https://doi.org/10.1109/sp.2017.20 [122] Lo Bello, L., Mubeen, S., Saponara, S., Mariani, R., Dutta Bordoloi, U.: Guest Editorial Embedded and Networked Sys- tems for Intelligent Vehicles and Robots. IEEE Transactions on Industrial Informat- ics 15(2), 1035–1037 (2019) https://doi.org/ 10.1109/tii.2018.2886529 [123] Abdullayev,V.,Faizal,A.,Seyi- dova, I., Mikayilov, S., Mammadova, R.,Pirverdiyeva,L.,etal.:Inte- grationofArtificialIntelligenceand Roboticsintotheindustrialsector. DataandMetadata 4,209(2025) https://doi.org/10.56294/dm2025209 [124] Prajapati, V.K., Shashank, S.A., Nithya, M.: Computer Vision Enabled Pick and Place Robot for Warehouse Automa- tion. In: 2024 5th International Confer- ence on Smart Electronics and Commu- nication (ICOSEC), p. 641–646. IEEE, Trichy, India (2024). https://doi.org/10. 1109/icosec61587.2024.10722718 [125] Mezg ́ar, I., V ́ancza, J., Paniti, I., T ́oth, J.: Transforming Experimental Cobot Cell to Industrial Realization – an Ethical AI Approach. IFAC-PapersOnLine 56(2), 7335–7341 (2023) https://doi.org/10.1016/ j.ifacol.2023.10.347 [126] Shen, J.: Progress And Comprehensive Analysis of Intelligent Grasping Technol- ogy for Robotic Arms Based on Computer Vision. Highlights in Science, Engineering and Technology 103, 205–211 (2024) https: //doi.org/10.54097/3rj4ke82 [127] Yang, C., Kang, J., Eom, D.-S.: Enhanc- ing ToF Sensor Precision Using 3D Mod- els and Simulation for Vision Inspection in Industrial Mobile Robots. Applied Sci- ences 14(11), 4595 (2024) https://doi.org/ 10.3390/app14114595 [128] Chen, S.-C., Pamungkas, R.S., Schmidt, D.: The Role of Machine Learning in Improv- ing Robotic Perception and Decision Mak- ing. International Transactions on Artificial Intelligence (ITALIC) 3(1), 32–43 (2024) https://doi.org/10.33050/italic.v3i1.661 [129] D, S., R, Anil.A., Sivadasan, J., Jayaku- mar, T., Manivel, M., C, A.: Intelligent control algorithms for industrial automation systems. In: 2025 IEEE International Stu- dents’ Conference on Electrical, Electronics and Computer Science (SCEECS), p. 1– 6. IEEE, Bhopal, India (2025). https://doi. org/10.1109/sceecs64059.2025.10940929 [130] You, J., Wu, Z., Wei, W., Li, N., Yang, Y.: Evolution of Industrial Robots from the Perspective of the Metaverse: Integra- tion of Virtual and Physical Realities and Human–Robot Collaboration. Applied Sci- ences 14(14), 6369 (2024) https://doi.org/ 10.3390/app14146369 [131] Oyekanlu, E.A., Smith, A.C., Thomas, W.P., Mulroy, G., Hitesh, D., Ramsey, M., et al.: A Review of Recent Advances in Automated Guided Vehicle Technologies: Integration Challenges and Research Areas for 5G-Based Smart Manufacturing Appli- cations. IEEE Access 8, 202312–202353 (2020) https://doi.org/10.1109/access.2020. 3035729 [132] Pelleti, S., Kumar, J.R.R., Krushnasamy, V.S., Raj, G.B.M.: Intelligent Control Systems for Industrial Automation and Robotics. In: 2024 Ninth International Con- ference on Science Technology Engineering and Mathematics (ICONSTEM), p. 1–5. IEEE, Chennai, India (2024). https://doi. org/10.1109/iconstem60960.2024.10568781 [133] Linder, T., Vaskevicius, N., Schirmer, R., Arras, K.O.: Cross-Modal Analysis of 47 Human Detection for Robotics: An Indus- trial Case Study. In: 2021 IEEE/RSJ Inter- national Conference on Intelligent Robots and Systems (IROS). IEEE, Prague, Czech Republic (2021). https://doi.org/10.1109/ iros51168.2021.9636158 [134] Alt, B., Zahn, J., Kienle, C., Dvorak, J., May, M., Katic, D., et al.: Human-AI Inter- action in Industrial Robotics: Design and Empirical Evaluation of a User Interface for Explainable AI-Based Robot Program Optimization. Procedia CIRP 130, 591– 596 (2024) https://doi.org/10.1016/j.procir. 2024.10.134 [135] Menolotto, M., Komaris, D.-S., Tedesco, S., O’Flynn, B., Walsh, M.: Motion Capture Technology in Industrial Applications: A Systematic Review. Sensors 20(19), 5687 (2020) https://doi.org/10.3390/s20195687 [136] Makulaviˇcius, M., Petkeviˇcius, S., Roˇz ̇en ̇e, J., Dzedzickis, A., Buˇcinskas, V.: Indus- trial Robots in Mechanical Machining: Perspectives and Limitations. Robotics 12(6), 160 (2023) https://doi.org/10.3390/ robotics12060160 [137] Kube, D., Hadwiger, S., Meisen, T.: Beyond performance: Explaining generalisation fail- ures of Robotic Foundation Models in indus- trial simulation. Biomimetic Intelligence and Robotics 5(4), 100249 (2025) https:// doi.org/10.1016/j.birob.2025.100249 [138] Tsoumplekas, G., Li, V., Siniosoglou, I., Argyriou, V., Goudos, S.K., Moscholios, I.D., et al.: Evaluating the Energy Efficiency of Few-Shot Learning for Object Detection in Industrial Settings. In: 2024 IEEE 3rd Real-Time and Intelligent Edge Computing Workshop (RAGE), p. 1–6. IEEE, Hong Kong, Hong Kong (2024). https://doi.org/ 10.1109/RAGE62451.2024.00016 [139] Zhang, Y., Wang, C., Lu, O., Zhao, Y., Ge, Y., Sun, Z., et al.: Align- Then-stEer: Adapting the Vision-Language Action Models through Unified Latent Guidance. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.02055 [140] Bohlinger, N., Czechmanowski, G., Krupka, M., Kicki, P., Walas, K., Peters, J., et al.: One Policy to Run Them All: An End-to-end Learning Approach to Multi- Embodiment Locomotion. arXiv (2025). https://doi.org/10.48550/arXiv.2409.06366 [141] Jones, E.K., Robey, A., Zou, A., Ravichan- dran, Z., Pappas, G.J., Hassani, H., et al.: Adversarial Attacks on Robotic Vision Lan- guage Action Models. arXiv (2025). https: //doi.org/10.48550/arXiv.2506.03350 [142] Xu, H., Koh, Y.S., Huang, S., Zhou, Z.,Wang,D.,Sakuma,J.,etal.: Model-AgnosticAdversarialAttack andDefenseforVision-Language- ActionModels.arXiv(2025).https: //doi.org/10.48550/arXiv.2510.13237 [143] Liu, H., Zhang, Y., Betala, V., Zhang, E., Liu, J., Ding, C., et al.: Multi-Task Inter- active Robot Fleet Learning with Visual World Models. arXiv (2024). https://doi. org/10.48550/arXiv.2410.22689 [144] Zollo, T.P., Zemel, R.: Confidence Cal- ibration in Vision-Language-Action Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2507.17383 [145] Ren, A.Z., Dixit, A., Bodrova, A., Singh, S., Tu, S., Brown, N., et al.: Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners. arXiv (2023). https://doi.org/10.48550/arXiv.2307.01928 [146] Chisari, E., Hartz, J.O., Despinoy, F., Valada,A.:RoboticTaskAmbigu- ity Resolution via Natural Language Interaction.arXiv(2025).https: //doi.org/10.48550/arXiv.2504.17748 [147] Bucker, A., Figueredo, L., Haddadin, S., Kapoor, A., Ma, S., Vemprala, S., et al.: LATTE: LAnguage Trajectory Trans- formEr. In: 2023 IEEE International Confer- ence on Robotics and Automation (ICRA), p. 7287–7294. IEEE, London, United Kingdom (2023). https://doi.org/10.1109/ ICRA48891.2023.10161068 48 [148] Shi, L.X., Hu, Z., Zhao, T.Z., Sharma, A., Pertsch, K., Luo, J., et al.: Yell At Your Robot: Improving On-the-Fly from Lan- guage Corrections. arXiv (2024). https:// doi.org/10.48550/arXiv.2403.12910 [149] Xiang, T.-Y., Jin, A.-Q., Zhou, X.- H.,Gui,M.-J.,Xie,X.-L.,Liu, S.-Q.,etal.:VLAModel-Expert Collaboration for Bi-directional Manip- ulationLearning.arXiv(2025). https://doi.org/10.48550/arXiv.2503.04163 [150] Zhou, E., Su, Q., Chi, C., Zhang, Z., Wang, Z., Huang, T., et al.: Code-as-Monitor: Constraint-aware Visual Programming for Reactive and Proactive Robotic Failure Detection. arXiv (2025). https://doi.org/10. 48550/arXiv.2412.04455 [151] Gu, Q., Ju, Y., Sun, S., Gilitschenski, I., Nishimura, H., Itkina, M., et al.: SAFE: Multitask Failure Detection for Vision- Language-Action Models. arXiv (2025). https://doi.org/10.48550/arXiv.2506.09937 [152] Xia, W., Feng, R., Wang, D., Hu, D.: Phoenix: A Motion-based Self-Reflection Framework for Fine-grained Robotic Action Correction. arXiv (2025). https://doi.org/ 10.48550/arXiv.2504.14588 [153] Koo, M., Choi, D., Kim, T., Lee, K., Kim, C., Seo, Y., et al.: HAMLET: Switch Your Vision-Language-Action Model into a History-Aware Policy. arXiv (2025). https: //doi.org/10.48550/arXiv.2510.00695 [154] Liu, Z., Bahety, A., Song, S.: REFLECT: Summarizing Robot Experiences for Failure Explanation and Correction. arXiv (2023). https://doi.org/10.48550/arXiv.2306.15724 [155] Miao, C., Chang, T., Wu, M., Xu, H., Li, C., Li, M., et al.: FedVLA: Federated Vision- Language-Action Learning with Dual Gat- ing Mixture-of-Experts for Robotic Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2508.02190 [156] Lin, Z., Duan, J., Fang, H., Fox, D., Krishna, R., Tan, C., et al.: FailSafe: Reasoning and Recovery from Failures in Vision-Language- Action Models. arXiv (2025). https://doi. org/10.48550/arXiv.2510.01642 [157] Team, G.R., Abdolmaleki, A., Abeyruwan, S., Ainslie, J., Alayrac, J.-B., Arenas, M.G., et al.: Gemini Robotics 1.5: Push- ing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer. arXiv (2025). https: //doi.org/10.48550/arXiv.2510.03342 [158] Sun, N., Li, Y., Wang, C., Li, H., Liu, H.: CollabVLA: Self-Reflective Vision- Language-Action Model Dreaming Together with Human. arXiv (2025). https://doi.org/ 10.48550/arXiv.2509.14889 [159] Li, S., Gao, L., Wang, J., Che, C., Xiao, X., Cao, J., et al.: Information-Theoretic Graph Fusion with Vision-Language-Action Model for Policy Reasoning and Dual Robotic Control. arXiv (2025). https://doi.org/10. 48550/arXiv.2508.05342 [160] Li, C., Liu, J., Wang, G., Li, X., Chen, S., Heng, L., et al.: A Self-Correcting Vision- Language-Action Model for Fast and Slow System Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2405.17418 [161] Jones, J., Mees, O., Sferrazza, C., Sta- chowicz, K., Abbeel, P., Levine, S.: Beyond Sight: Finetuning Generalist Robot Policies with Heterogeneous Sensors via Language Grounding. arXiv (2025). https://doi.org/ 10.48550/arXiv.2501.04693 [162] Huang, H., Cen, M., Tan, K., Quan, X., Huang, G., Zhang, H.: GraphCoT- VLA: A 3D Spatial-Aware Reasoning Vision-Language-Action Model for Robotic Manipulation with Ambiguous Instruc- tions. arXiv (2025). https://doi.org/10. 48550/arXiv.2508.07650 [163] Li, Y., Deng, Y., Zhang, J., Jang, J., Mem- mel, M., Yu, R., et al.: HAMSTER: Hier- archical Action Models For Open-World Robot Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2502.05485 49 [164] Wu, H., Zhang, H., Zhang, P., Wang, J., Wang, C.: HiBerNAC: Hierarchical Brain- emulated Robotic Neural Agent Collec- tive for Disentangling Complex Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.08296 [165] Lynch, C., Sermanet, P.: Language Con- ditioned Imitation Learning over Unstruc- tured Data. arXiv (2021). https://doi.org/ 10.48550/arXiv.2005.07648 [166] Chen, J., Wang, J., Chen, L., Cai, C., Lu, J.: NanoVLA: Routing Decoupled Vision- Language Understanding for Nano-sized Generalist Robotic Policies. arXiv (2025). https://doi.org/10.48550/arXiv.2510.25122 [167] Driess, D., Xia, F., Sajjadi, M.S.M., Lynch, C., Chowdhery, A., Ichter, B., et al.: PaLM- E: An Embodied Multimodal Language Model. arXiv (2023). https://doi.org/10. 48550/arXiv.2303.03378 [168] Shukor, M., Aubakirova, D., Capuano, F., Kooijmans, P., Palma, S., Zouitine, A., et al.: SmolVLA: A Vision-Language- Action Model for Affordable and Efficient Robotics. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.01844 [169] Bu, Q., Yang, Y., Cai, J., Gao, S., Ren, G., Yao, M., et al.: UniVLA: Learning to Act Anywhere with Task-centric Latent Actions. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.06111 [170] Sun, J., Mao, P., Kong, L., Wang, J.: A Review of Embodied Grasping. Sensors 25(3), 852 (2025) https://doi.org/10.3390/ s25030852 [171] Yuan, M., Yu, T., Ge, W., Yao, X., Wang, H., Chen, J., et al.: A Survey of Behav- ior Foundation Model: Next-Generation Whole-Body Control System of Humanoid Robots. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.20487 [172] Gao, J., Belkhale, S., Dasari, S., Balakr- ishna, A., Shah, D., Sadigh, D.: A Tax- onomy for Evaluating Generalist Robot Policies. arXiv (2025). https://doi.org/10. 48550/arXiv.2503.01238 [173] Liu, Y., Chen, W., Bai, Y., Liang, X., Li, G., Gao, W., et al.: Aligning Cyber Space with Physical World: A Comprehensive Sur- vey on Embodied AI. arXiv (2025). https: //doi.org/10.48550/arXiv.2407.06886 [174] Li, J., Li, J., Yang, G., Yang, L., Chi, H., Yang, L.: Applications of Large Lan- guage Models and Multimodal Large Models in Autonomous Driving: A Comprehensive Review. Drones 9(4), 238 (2025) https:// doi.org/10.3390/drones9040238 [175] Lisondra, M., Benhabib, B., Nejat, G.: Embodied AI with Foundation Models for Mobile Service Robots: A Systematic Review. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.20503 [176] Xu, P.: Embodied AI: Bridging Simula- tion and Reality in Robotics. In: 2025 4th International Symposium on Robotics, Arti- ficial Intelligence and Information Engineer- ing (RAIIE), p. 244–248. IEEE, Zhen- jiang, China (2025). https://doi.org/10. 1109/RAIIE65740.2025.11140070 [177] Feng, T., Wang, X., Jiang, Y.-G., Zhu, W.: Embodied AI: From LLMs to World Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2509.20021 [178] Chen, S.: Exploring Embodied Multimodal Large Models: Development, Datasets, and Future Directions. Elsevier BV (2025). https://doi.org/10.48550/arXiv.2502.15336 [179] Khan, M.T., Waheed, A.: Foundation Model Driven Robotics: A Comprehensive Review. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.10087 [180] Sui, X., Tian, D., Sun, Q., Chen, R., Choi, D., Kwok, K., et al.: From Grounding to Manipulation: Case Studies of Foundation Model Integration in Embodied Robotic Systems. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.15685 50 [181] Zhang, K., Yun, P., Cen, J., Cai, J., Zhu, D., Yuan, H., et al.: Generative Artifi- cial Intelligence in Robotic Manipulation: A Survey. arXiv (2025). https://doi.org/10. 48550/arXiv.2503.03464 [182] Liu, Y., Sun, Q., Kapadia, D.R.: Integrat- ing Large Language Models into Robotic Autonomy: A Review of Motion, Voice, and Training Pipelines. AI 6(7), 158 (2025) https://doi.org/10.3390/ai6070158 [183] Mehta, V., Sharma, C., Thiyagarajan, K.: Large Language Models and 3D Vision for Intelligent Robotic Perception and Auton- omy. Sensors 25(20), 6394 (2025) https:// doi.org/10.3390/s25206394 [184] Liang, W., Zhou, R., Ma, Y., Zhang, B., Li, S., Liao, Y., et al.: Large Model Empowered Embodied AI: A Survey on Decision-Making and Embodied Learning. arXiv (2025). https://doi.org/10.48550/arXiv.2508.10399 [185] Borazjani, K., Abdisarabshali, P., Nadimi, F., Khosravan, N., Liwang, M., Wang, X., et al.: Multi-Modal Multi-Task (M3T) Feder- ated Foundation Models for Embodied AI: Potentials and Challenges for Edge Inte- gration. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.11191 [186] Han, X., Chen, S., Fu, Z., Feng, Z., Fan, L., An, D., et al.: Multimodal Fusion and Vision-Language Models: A Survey for Robot Vision. arXiv (2025). https://doi. org/10.48550/arXiv.2504.02477 [187] Liu, J., Shi, X., Nguyen, T.D., Zhang, H., Zhang, T., Sun, W., et al.: Neural Brain: A Neuroscience-inspired Framework for Embodied Agents. arXiv (2025). https: //doi.org/10.48550/arXiv.2505.07634 [188] Sartor, S., Thompson, N.: Neural Scaling Laws in Robotics. arXiv (2025). https://doi. org/10.48550/arXiv.2405.14005 [189] Wang, Y., Sun, A.: Toward Embodied AGI: A Review of Embodied AI and the Road Ahead. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.14235 [190] Salimpour, S., Fu, L., Rachwa l, K., Bertrand,P.,O’Sullivan,K.,Jakob, R., et al.: Towards Embodied Agen- tic AI: Review and Classification of LLM- and VLM-Driven Robot Auton- omyandInteraction.arXiv(2025). https://doi.org/10.48550/arXiv.2508.05294 [191] Li, D., Jin, Y., Sun, Y., A, Y., Yu, H., Shi, J., et al.: What Foundation Models Can Bring for Robot Learning in Manipulation : A Survey. arXiv (2025). https://doi.org/10. 48550/arXiv.2404.18201 [192] Gao, H., Wang, Z., Li, Y., Long, K., Yang, M., Shen, Y.: A Survey for Foundation Mod- els in Autonomous Driving. arXiv (2024). https://doi.org/10.48550/arXiv.2402.01105 [193] Jeong, H., Lee, H., Kim, C., Shin, S.: A Survey of Robot Intelligence with Large Language Models. Applied Sciences 14(19), 8868 (2024) https://doi.org/10. 3390/app14198868 [194] Cohen, V., Liu, J.X., Mooney, R., Tellex, S., Watkins, D.: A Survey of Robotic Language Grounding: Tradeoffs between Symbols and Embeddings. arXiv (2024). https://doi.org/ 10.48550/arXiv.2405.13245 [195] Kim, Y., Kim, D., Choi, J., Park, J., Oh, N., Park, D.: A survey on integration of large language models with intelligent robots. Intelligent Service Robotics 17(5), 1091–1107 (2024) https://doi.org/10.1007/ s11370-024-00550-5 [196] Xu, Z., Wu, K., Wen, J., Li, J., Liu, N., Che, Z., et al.: A Survey on Robotics with Foundation Models: Toward Embodied AI. arXiv (2024). https://doi.org/10.48550/ arXiv.2402.02385 [197] Banjanovi ́c-Mehmedovi ́c, L., Husakovi ́c, A., Gurdi ́c Ribi ́c, A., Prljaˇca, N., Karabegovi ́c, I.: Advancements in Robotic Intelligence: The Role of Computer Vision, DRL, Trans- formers and LLMs. In: Artificial Intelligence in Industry 4.0: The Future That Comes True, p. 94–127. Academy of Sciences and Arts of Bosnia and Herzegovina, ??? (2024). 51 https://doi.org/10.5644/pi2024.215.05 [198] Sanghai, N., Brown, N.B.: Advances in Transformers for Robotic Applications: A Review. arXiv (2024). https://doi.org/10. 48550/arXiv.2412.10599 [199] Lin, M.-Y., Lee, O.-W., Lu, C.-Y.: Embod- ied AI with Large Language Models: A Survey and New HRI Framework. In: 2024 International Conference on Advanced Robotics and Mechatronics (ICARM), p. 978–983. IEEE, Tokyo, Japan (2024). https://doi.org/10.1109/ICARM62033. 2024.10715872 [200] Naderi, H., Shojaei, A., Huang, L.: Foun- dation Models for Autonomous Robots in Unstructured Environments. arXiv (2024). https://doi.org/10.48550/ARXIV.2407. 14296 [201] Moroncelli,A.,Soni,V.,Shahid, A.A.,Maccarini,M.,Forgione,M., Piga,D.,etal.:IntegratingRein- forcementLearningwithFoundation Models for Autonomous Robotics: Meth- odsandPerspectives.arXiv(2024). https://doi.org/10.48550/arXiv.2410.16411 [202] Wang, J., Wu, Z., Li, Y., Jiang, H., Shu, P., Shi, E., et al.: Large Language Mod- els for Robotics: Opportunities, Challenges, and Perspectives. arXiv (2024). https://doi. org/10.48550/arXiv.2401.04334 [203] Adebayo, A.S., Chukwurah, N., Ajayi, O.O.: Leveraging Foundation Models in Robotics: Transforming Task Planning and Contextual Execution. International Journal of Multidisciplinary Research and Growth Evaluation 5(1), 1388–1394 (2024) https://doi.org/10.54660/.IJMRGE.2024.5. 1.1388-1394 [204] Atuhurra,J.:LeveragingLargeLan- guageModelsinHuman-Robot Interaction:ACriticalAnalysisof PotentialandPitfalls.arXiv(2024). https://doi.org/10.48550/arXiv.2405.00693 [205] Kawaharazuka, K., Matsushima, T., Gam- bardella, A., Guo, J., Paxton, C., Zeng, A.: Real-world robot applications of founda- tion models: A review. Advanced Robotics 38(18), 1232–1254 (2024) https://doi.org/ 10.1080/01691864.2024.2408593 [206] Pashangpour, S., Nejat, G.: The Future of Intelligent Healthcare: A Systematic Analysis and Discussion on the Integra- tion and Impact of Robots Using Large Language Models for Healthcare. Robotics 13(8), 112 (2024) https://doi.org/10.3390/ robotics13080112 [207] McCarthy, R., Tan, D.C.H., Schmidt, D., Acero, F., Herr, N., Du, Y., et al.: Towards Generalist Robot Learning from Internet Video: A Survey. arXiv (2024). https://doi. org/10.48550/arXiv.2404.19664 [208] Google DeepMind: Demonstrating Large Language Models on Robots. In: Robotics: Science and Systems XIX. Robotics: Science and Systems Foundation, ??? (2023). https: //doi.org/10.15607/RSS.2023.XIX.024 [209] Zeng, F., Gan, W., Wang, Y., Liu, N., Yu, P.S.: Large Language Models for Robotics: A Survey. arXiv (2023). https://doi.org/10. 48550/arXiv.2311.07226 [210] Xiao, X., Liu, J., Wang, Z., Zhou, Y., Qi, Y., Cheng, Q., et al.: Robot Learn- ing in the Era of Foundation Models: A Survey. arXiv (2023). https://doi.org/10. 48550/arXiv.2311.14379 [211] Huang, J., Wang, S., Lin, F., Hu, Y., Wen, C., Gao, Y.: Tactile-VLA: UnlockingVision-Language-Action Model’s Physical Knowledge for Tac- tileGeneralization.arXiv(2025). https://doi.org/10.48550/arXiv.2507.09160 [212] Zhong, Y., Huang, X., Li, R., Zhang, C.,Chen,Z.,Guan,T.,etal.: DexGraspVLA:AVision-Language- ActionFrameworkTowardsGeneral DexterousGrasping.arXiv(2025). https://doi.org/10.48550/arXiv.2502.20900 52 [213] Team, G.R., Abeyruwan, S., Ainslie, J., Alayrac, J.-B., Arenas, M.G., Armstrong, T., et al.: Gemini Robotics: Bringing AI into the Physical World. arXiv (2025). https: //doi.org/10.48550/arXiv.2503.20020 [214] Fang, H., Zhang, M., Dong, H., Li, W., Wang, Z., Zhang, Q., et al.: Robix: A Unified Model for Robot Interaction, Reasoning and Planning. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.01106 [215] Liu, X., Fu, C., Yan, C., Wu, C., Gao, H., Zhang, Y.-F., et al.: VITA-E: Natural Embodied Interaction with Con- current Seeing, Hearing, Speaking, and Acting. arXiv (2025). https://doi.org/10. 48550/arXiv.2510.21817 [216] Dai, Y., Lee, J., Fazeli, N., Chai, J.: RACER: Rich Language-Guided Failure Recovery Policies for Imitation Learning. In: 2025 IEEE International Conference on Robotics and Automation (ICRA), p. 15657–15664. IEEE, Atlanta, GA, USA(2025).https://doi.org/10.1109/ ICRA55743.2025.11127799 [217] Yuan, Y., Cui, H., Huang, Y., Chen, Y., Ni, F., Dong, Z., et al.: Embodied-R1: Reinforced Embodied Reasoning for General Robotic Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2508.13998 [218] Chen, G., Wang, M., Cui, T., Mu, Y., Lu, H., Zhou, T., et al.: VLMimic: Vision Lan- guage Models Are Visual Imitation Learner for Fine-grained Actions. arXiv (2024). https://doi.org/10.48550/arXiv.2410.20927 [219] Chen, X., Chen, Y., Fu, Y., Gao, N., Jia, J., Jin, W., et al.: InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy. arXiv (2025). https://doi.org/10.48550/arXiv.2510.13778 [220] Yang, Y., Duan, Z., Xie, T., Cao, F., Shen, P., Song, P., et al.: FPC-VLA: A Vision-Language-Action Framework with a Supervisor for Failure Prediction and Cor- rection. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.04018 [221] Li, J., Zhu, Y., Tang, Z., Wen, J., Zhu, M., Liu, X., et al.: CoA-VLA: Improving Vision- Language-Action Models via Visual-Textual Chain-of-Affordance. arXiv (2025). https:// doi.org/10.48550/arXiv.2412.20451 [222] Huang, C.-P., Wu, Y.-H., Chen, M.-H., Wang, Y.-C.F., Yang, F.-E.: ThinkAct: Vision-Language-ActionReasoningvia Reinforced Visual Latent Planning. arXiv (2025).https://doi.org/10.48550/arXiv. 2507.16815 [223] Lee, J., Duan, J., Fang, H., Deng, Y., Liu, S., Li, B., et al.: MolmoAct: Action Reasoning Models That Can Reason in Space. arXiv (2025). https://doi.org/10. 48550/arXiv.2508.07917 [224] Belkhale, S., Ding, T., Xiao, T., Ser- manet, P., Vuong, Q., Tompson, J., et al.: RT-H: Action Hierarchies Using Lan- guage. arXiv (2024). https://doi.org/10. 48550/arXiv.2403.01823 [225] Huang, W., Wang, C., Zhang, R., Li, Y., Wu, J., Fei-Fei, L.: VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models. arXiv (2023). https: //doi.org/10.48550/arXiv.2307.05973 [226] Bu, Q., Zeng, J., Chen, L., Yang, Y., Zhou, G., Yan, J., et al.: Closed-Loop Visuomo- tor Control with Generative Expectation for Robotic Manipulation. arXiv (2024). https: //doi.org/10.48550/arXiv.2409.09016 [227] Kang,G.-C.,Kim,J.,Shim,K., Lee,J.K.,Zhang,B.-T.:CLIP- RT:LearningLanguage-Conditioned RoboticPoliciesfromNaturalLan- guageSupervision.arXiv(2025). https://doi.org/10.48550/arXiv.2411.00508 [228] Li, M., Zhao, Z., Che, Z., Liao, F., Wu, K., Xu, Z., et al.: SwitchVLA: Execution-Aware Task Switching for Vision-Language-Action Models. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.03574 [229] Li, S., Gao, Y., Sadigh, D., Song, S.: Unified Video Action Model. arXiv (2025). https:// 53 doi.org/10.48550/arXiv.2503.00200 [230] Liu, H., Tong, Y., Liu, G., Ju, Z., Zhang, Z.: IDAGC: Adaptive Generalized Human-Robot Collaboration via Human Intent Estimation and Multimodal Policy Learning. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.04620 [231] Xu, K., Xia, X., Wang, K., Yang, Y., Mao, Y., Deng, B., et al.: Efficient Align- ment of Unconditioned Action Prior for Language-conditioned Pick and Place in Clutter. arXiv (2025). https://doi.org/10. 48550/arXiv.2503.09423 [232] Black, K., Nakamoto, M., Atreya, P., Walke, H., Finn, C., Kumar, A., et al.: Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Mod- els. arXiv (2023). https://doi.org/10.48550/ arXiv.2310.10639 [233] Pan, M., Zhang, J., Wu, T., Zhao, Y., Gao, W., Dong, H.: OmniManip: Towards General Robotic Manipulation via Object- Centric Interaction Primitives as Spatial Constraints. In: 2025 IEEE/CVF Con- ference on Computer Vision and Pat- tern Recognition (CVPR), p. 17359–17369. IEEE, Nashville, TN, USA (2025). https: //doi.org/10.1109/CVPR52734.2025.01618 [234] Liu, L., Wang, W., Han, Y., Xie, Z., Yi, P., Li, J., et al.: FoAM: Foresight-Augmented Multi-Task Imitation Policy for Robotic Manipulation. arXiv (2025). https://doi. org/10.48550/arXiv.2409.19528 [235] Zhang, C., Hao, P., Cao, X., Hao, X., Cui, S., Wang, S.: VTLA: Vision-Tactile-Language- Action Model with Preference Learning for Insertion Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2505.09577 [236] Ze, Y., Yan, G., Wu, Y.-H., Macaluso, A., Ge, Y., Ye, J., et al.: GNFactor: Multi-Task Real Robot Learning with Generalizable Neural Feature Fields. arXiv (2024). https: //doi.org/10.48550/arXiv.2308.16891 [237] Wen, J., Zhu, M., Zhu, Y., Tang, Z., Li, J., Zhou, Z., et al.: Diffusion-VLA: Gen- eralizable and Interpretable Robot Foun- dation Model via Self-Generated Rea- soning. arXiv (2025). https://doi.org/10. 48550/arXiv.2412.03293 [238] Li, P., Wu, H., Huang, Y., Cheang, C., Wang, L., Kong, T.: GR-MG: Leveraging Partially Annotated Data via Multi-Modal Goal-Conditioned Policy. arXiv (2024). https://doi.org/10.48550/arXiv.2408.14368 [239] Nasiriany, S., Kirmani, S., Ding, T., Smith, L., Zhu, Y., Driess, D., et al.: RT- Affordance: Affordances Are Versatile Inter- mediate Representations for Robot Manip- ulation. arXiv (2024). https://doi.org/10. 48550/arXiv.2411.02704 [240] Yang, S., Li, H., Chen, Y., Wang, B.,Tian,Y.,Wang,T.,etal.: InstructVLA:Vision-Language-Action InstructionTuningfromUnderstand- ingtoManipulation.arXiv(2025). https://doi.org/10.48550/arXiv.2507.17520 [241] Bharadhwaj, H., Vakil, J., Sharma, M., Gupta, A., Tulsiani, S., Kumar, V.: RoboA- gent: Generalization and Efficiency in Robot Manipulation via Semantic Augmentations and Action Chunking. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 4788–4795. IEEE, Yokohama, Japan (2024). https://doi.org/ 10.1109/ICRA57147.2024.10611293 [242] Li, C., Wen, J., Peng, Y., Peng, Y., Feng, F., Zhu, Y.: PointVLA: Injecting the 3D World into Vision-Language-Action Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2503.07511 [243] Qu, D., Song, H., Chen, Q., Chen, Z., Gao, X., Ye, X., et al.: EO-1: Interleaved Vision- Text-Action Pretraining for General Robot Control. arXiv (2025). https://doi.org/10. 48550/arXiv.2508.21112 [244] Tang, W., Jing, D., Pan, J.-H., Lu, Z., Liu, Y.-H., Li, L.E., et al.: Incentivizing Multi- modal Reasoning in Large Models for Direct Robot Manipulation. arXiv (2025). https: 54 //doi.org/10.48550/arXiv.2505.12744 [245] Zhai, A., Liu, B., Fang, B., Cai, C., Ma, E., Yin, E., et al.: Igniting VLMs toward the Embodied Space. arXiv (2025). https://doi. org/10.48550/arXiv.2509.11766 [246] Feng, Y., Tan, H., Mao, X., Xiang, C., Liu, G., Huang, S., et al.: Vidar: Embodied Video Diffusion Model for Generalist Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.12898 [247] Li, H., Yang, S., Chen, Y., Chen, X., Yang, X., Tian, Y., et al.: CronusVLA: Towards Efficient and Robust Manipu- lation via Multi-Frame Vision-Language- Action Modeling. arXiv (2025). https://doi. org/10.48550/arXiv.2506.19816 [248] Singh, I., Goyal, A., Birchfield, S., Fox, D., Garg, A., Blukis, V.: OG-VLA: Ortho- graphic Image Generation for 3D-Aware Vision-LanguageActionModel.arXiv (2025).https://doi.org/10.48550/arXiv. 2506.01196 [249] Huang,H.,Liu,F.,Fu,L.,Wu, T.,Mukadam,M.,Malik,J.,et al.:OTTER:AVision-Language- Action Model with Text-Aware Visual FeatureExtraction.arXiv(2025). https://doi.org/10.48550/arXiv.2503.03734 [250] Team, G., Ye, A., Wang, B., Ni, C., Huang, G., Zhao, G., et al.: GigaBrain-0: A World Model-Powered Vision-Language- Action Model. arXiv (2025). https://doi. org/10.48550/arXiv.2510.19430 [251] Sun, Q., Hong, P., Pala, T.D., Toh, V., Tan, U.-X., Ghosal, D., et al.: Emma-X: An Embodied Multimodal Action Model with Grounded Chain of Thought and Look- ahead Spatial Reasoning. arXiv (2024). https://doi.org/10.48550/arXiv.2412.11974 [252] Xue, H., Huang, X., Niu, D., Liao, Q., Kragerud, T., Gravdahl, J.T., et al.:LeVERB:HumanoidWhole-Body ControlwithLatentVision-Language Instruction.arXiv(2025).https: //doi.org/10.48550/arXiv.2506.13751 [253] Yan, G., Zhu, J., Deng, Y., Yang, S., Qiu, R.-Z., Cheng, X., et al.: ManiFlow: A Gen- eral Robot Manipulation Policy via Consis- tency Flow Training. arXiv (2025). https: //doi.org/10.48550/arXiv.2509.01819 [254] Stone, A., Xiao, T., Lu, Y., Gopalakr- ishnan, K., Lee, K.-H., Vuong, Q., et al.: Open-World Object Manipulation Using Pre-trained Vision-Language Models. arXiv (2023).https://doi.org/10.48550/arXiv. 2303.00905 [255] Liu, Z., Liu, J., Xu, J., Han, N., Gu, C., Chen, H., et al.: MLA: A Multisen- sory Language-Action Model for Multi- modal Understanding and Forecasting in Robotic Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2509.26642 [256] Garcia,R.,Chen,S.,Schmid,C.: TowardsGeneralizableVision-Language Robotic Manipulation: A Benchmark and LLM-Guided 3D Policy. In: 2025 IEEE International Conference on Robotics and Automation (ICRA), p. 8996–9002. IEEE, Atlanta, GA, USA (2025). https://doi.org/ 10.1109/ICRA55743.2025.11127315 [257] Guhur, P.-L., Chen, S., Garcia, R., Tapaswi, M., Laptev, I., Schmid, C.: Instruction- Driven History-Aware Policies for Robotic Manipulations. arXiv (2022). https://doi. org/10.48550/arXiv.2209.04899 [258] Li, P., Chen, Y., Wu, H., Ma, X., Wu, X., Huang, Y., et al.: BridgeVLA: Input- Output Alignment for Efficient 3D Manipu- lation Learning with Vision-Language Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2506.07961 [259] Wang, R., Mao, J., Hsu, J., Zhao, H., Wu, J., Gao, Y.: Programmatically Grounded, Compositionally Generalizable Robotic Manipulation. arXiv (2023). https: //doi.org/10.48550/arXiv.2304.13826 [260] Kuang, Y., Ye, J., Geng, H., Mao, J., Deng, C., Guibas, L., et al.: RAM: Retrieval-Based 55 Affordance Transfer for Generalizable Zero- Shot Robotic Manipulation. arXiv (2024). https://doi.org/10.48550/arXiv.2407.04689 [261] Luo, G., Yang, G., Gong, Z., Chen, G., Duan, H., Cui, E., et al.: Visual Embod- ied Brain: Let Multimodal Large Lan- guage Models See, Think, and Control in Spaces. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.00123 [262] Zheng, J., Li, J., Liu, D., Zheng, Y., Wang, Z., Ou, Z., et al.: Universal Actions for Enhanced Embodied Foundation Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2501.10105 [263] Ma, Y., Chi, D., Wu, S., Liu, Y., Zhuang, Y., Hao, J., et al.: Actra: Optimized Transformer Architecture for Vision-Language-Action Models in Robot Learning. arXiv (2024). https://doi.org/10. 48550/arXiv.2408.01147 [264] Goyal, A., Blukis, V., Xu, J., Guo, Y., Chao, Y.-W., Fox, D.: RVT-2: Learning Precise Manipulation from Few Demon- strations. arXiv (2024). https://doi.org/10. 48550/arXiv.2406.08545 [265] Niu, Y., Zhou, S., Li, Y., Den, Y., Wang, L.: Time-Unified Diffusion Policy with Action Discrimination for Robotic Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.09422 [266] Zhang, J., Hu, Y., Guo, Y., Chen, X., Liu, Y., Chen, W., et al.: Uni- CoD: Enhancing Robot Policy via Uni- fied Continuous and Discrete Representa- tion Learning. arXiv (2025). https://doi. org/10.48550/arXiv.2510.10642 [267] Gu,J.,Kirmani,S.,Wohlhart,P., Lu, Y., Arenas, M.G., Rao, K., et al.:RT-Trajectory:RoboticTask GeneralizationviaHindsightTrajec- torySketches.arXiv(2023).https: //doi.org/10.48550/arXiv.2311.01977 [268] Zhang, J., Wu, S., Luo, X., Wu, H., Gao, L., Shen, H.T., et al.: InSpire: Vision-Language- Action Models with Intrinsic Spatial Rea- soning. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.13888 [269] Bai, Y., Wang, Z., Liu, Y., Luo, K., Wen, Y., Dai, M., et al.: Learning to See and Act: Task-Aware Virtual View Exploration for Robotic Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2508.05186 [270] Wang, L., Chen, X., Zhao, J., He, K.: Scaling Proprioceptive-Visual Learning with Heterogeneous Pre-trained Transform- ers. arXiv (2024). https://doi.org/10.48550/ arXiv.2409.20537 [271] Song, H., Qu, D., Yao, Y., Chen, Q., Lv, Q., Tang, Y., et al.: Hume: Introduc- ing System-2 Thinking in Visual-Language- Action Model. arXiv (2025). https://doi. org/10.48550/arXiv.2505.21432 [272] Stepputtis, S., Campbell, J., Phielipp, M., Lee, S., Baral, C., Amor, H.B.: Language- Conditioned Imitation Learning for Robot Manipulation Tasks. arXiv (2020). https:// doi.org/10.48550/arXiv.2010.12083 [273] Ye, S., Jang, J., Jeon, B., Joo, S., Yang, J., Peng, B., et al.: Latent Action Pretraining from Videos. arXiv (2025). https://doi.org/ 10.48550/arXiv.2410.11758 [274] Fang, H., Grotz, M., Pumacay, W., Wang, Y.R., Fox, D., Krishna, R., et al.: SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2501.18564 [275] Hu, Y., Guo, Y., Wang, P., Chen, X., Wang, Y.-J., Zhang, J., et al.: Video Prediction Pol- icy: A Generalist Robot Policy with Predic- tive Visual Representations. arXiv (2025). https://doi.org/10.48550/arXiv.2412.14803 [276] Li, H., Feng, Q., Zheng, Z., Feng, J., Chen, Z., Knoll, A.: Language-Guided Object- Centric Diffusion Policy for Generalizable and Collision-Aware Manipulation. In: 2025 IEEE International Conference on Robotics 56 and Automation (ICRA), p. 12834– 12841. IEEE, Atlanta, GA, USA (2025). https://doi.org/10.1109/ICRA55743.2025. 11127231 [277] Fan, C., Jia, X., Sun, Y., Wang, Y., Wei, J., Gong, Z., et al.: Interleave-VLA: Enhanc- ing Robot Manipulation with Interleaved Image-Text Instructions. arXiv (2025). https://doi.org/10.48550/arXiv.2505.02152 [278] Team, T.L., Barreiros, J., Beaulieu, A., Bhat, A., Cory, R., Cousineau, E., et al.: A Careful Examination of Large Behav- ior Models for Multitask Dexterous Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.05331 [279] Saxena, S., Sharma, M., Kroemer, O.: MResT: Multi-Resolution Sensing for Real- Time Control with Vision-Language Mod- els. arXiv (2024). https://doi.org/10.48550/ arXiv.2401.14502 [280] Niu, D., Sharma, Y., Xue, H., Biamby, G., Zhang, J., Ji, Z., et al.: Pre-Training Auto- regressive Robotic Models with 4D Rep- resentations. arXiv (2025). https://doi.org/ 10.48550/arXiv.2502.13142 [281] Xu, J., Nie, X.: SPECI: Skill Prompts Based Hierarchical Continual Imitation Learn- ing for Robot Manipulation. arXiv (2025). https://doi.org/10.48550/arXiv.2504.15561 [282] Reed, S., Zolna, K., Parisotto, E., Col- menarejo, S.G., Novikov, A., Barth-Maron, G., et al.: A Generalist Agent. arXiv (2022). https://doi.org/10.48550/arXiv.2205.06175 [283] Lin, T., Li, G., Zhong, Y., Zou, Y., Du, Y., Liu, J., et al.: Evo-0: Vision-Language- Action Model with Implicit Spatial Under- standing. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.00416 [284] Huang, W., Chen, C., Qi, H., Lv, C., Du, Y., Yang, H.: MoTVLA: A Vision-Language- Action Model with Unified Fast-Slow Rea- soning. arXiv (2025). https://doi.org/10. 48550/arXiv.2510.18337 [285] Zhang, H., Ding, P., Lyu, S., Peng, Y., Wang, D.: GEVRM: Goal-Expressive Video Generation Model For Robust Visual Manipulation. arXiv (2025). https://doi. org/10.48550/arXiv.2502.09268 [286] Wen, J., Zhu, Y., Li, J., Zhu, M., Tang, Z., Wu, K., et al.: TinyVLA: Toward Fast,Data-EfficientVision-Language- Action Models for Robotic Manipulation. IEEE Robotics and Automation Let- ters 10(4),3988–3995(2025)https: //doi.org/10.1109/LRA.2025.3544909 [287] Yang, Y., Sun, J., Kou, S., Wang, Y., Deng, Z.: LoHoVLA: A Unified Vision-Language- Action Model for Long-Horizon Embod- ied Tasks. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.00411 [288] Shi, H., Xie, B., Liu, Y., Sun, L., Liu,F.,Wang,T.,etal.:Memo- ryVLA:Perceptual-CognitiveMemory inVision-Language-ActionModelsfor RoboticManipulation.arXiv(2025). https://doi.org/10.48550/arXiv.2508.19236 [289] Liu, S., Wu, L., Li, B., Tan, H., Chen, H., Wang, Z., et al.: RDT-1B: A Diffu- sion Foundation Model for Bimanual Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2410.07864 [290] NVIDIA, Bjorck, J., Casta ̃neda, F., Cherni- adev, N., Da, X., Ding, R., et al.: GR00T N1: An Open Foundation Model for Gener- alist Humanoid Robots. arXiv (2025). https: //doi.org/10.48550/arXiv.2503.14734 [291] Wu, H., Jing, Y., Cheang, C., Chen, G., Xu, J., Li, X., et al.: Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation. arXiv (2023). https: //doi.org/10.48550/arXiv.2312.13139 [292] Jiang, Y., Gupta, A., Zhang, Z., Wang, G., Dou, Y., Chen, Y., et al.: VIMA: Gen- eral Robot Manipulation with Multimodal Prompts. arXiv (2023). https://doi.org/10. 48550/arXiv.2210.03094 57 [293] Jang, E., Irpan, A., Khansari, M., Kap- pler, D., Ebert, F., Lynch, C., et al.: BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning. arXiv (2022). https:// doi.org/10.48550/arXiv.2202.02005 [294] Chen, Y., Gu, K., Wen, Y., Zhao, Y., Wang, T., Nie, L.: IntentionVLA: Generalizable and Efficient Embodied Intention Reasoning for Human-Robot Interaction. arXiv (2025). https://doi.org/10.48550/arXiv.2510.07778 [295] Shridhar, M., Manuelli, L., Fox, D.: CLI- Port: What and Where Pathways for Robotic Manipulation. arXiv (2021). https: //doi.org/10.48550/arXiv.2109.12098 [296] Xia, S., Fang, H., Lu, C., Fang, H.-S.: CAGE: Causal Attention Enables Data- Efficient Generalizable Robotic Manipu- lation. arXiv (2024). https://doi.org/10. 48550/arXiv.2410.14974 [297] Wang, S., You, J., Hu, Y., Li, J., Gao, Y.: SKIL: Semantic Keypoint Imitation Learn- ing for Generalizable Data-efficient Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2501.14400 [298] Yang,D.,Zhao,Z.,Liu,Y.:AR- VRM:ImitatingHumanMotions forVisualRobotManipulationwith AnalogicalReasoning.arXiv(2025). https://doi.org/10.48550/arXiv.2508.07626 [299] Sun, L., Xie, B., Liu, Y., Shi, H., Wang, T., Cao, J.: GeoVLA: Empowering 3D Repre- sentations in Vision-Language-Action Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2508.09071 [300] Liang, Z., Mu, Y., Ma, H., Tomizuka, M., Ding, M., Luo, P.: SkillDiffuser: Inter- pretable Hierarchical Planning via Skill Abstractions in Diffusion-Based Task Exe- cution. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 16467–16476. IEEE, Seattle, WA, USA (2024). https://doi.org/10.1109/ CVPR52733.2024.01558 [301] Chen, H., Guo, J., Wang, B., Zhang, T., Huang, X., Zheng, B., et al.: Goal-VLA: Image-Generative VLMs as Object-Centric World Models Empowering Zero-shot Robot Manipulation. arXiv (2025). https://doi. org/10.48550/arXiv.2506.23919 [302] Lin, J., Taherin, A., Akbari, A., Akbari, A., Lu, L., Chen, G., et al.: VOTE: Vision- Language-Action Optimization with Trajec- tory Ensemble Voting. arXiv (2025). https: //doi.org/10.48550/arXiv.2507.05116 [303] Wu, S., Zhu, Y., Huang, Y., Zhu, K., Gu, J., Yu, J., et al.: AffordDP: Gener- alizable Diffusion Policy with Transferable Affordance. arXiv (2025). https://doi.org/ 10.48550/arXiv.2412.03142 [304] Ranasinghe, K., Li, X., Nguyen, E.-R., Mata, C., Park, J., Ryoo, M.S.: Pixel Motion as Universal Representation for Robot Control. arXiv (2025). https://doi. org/10.48550/arXiv.2505.07817 [305] Cheang, C.-L., Chen, G., Jing, Y., Kong, T., Li, H., Li, Y., et al.: GR-2: A Gen- erative Video-Language-Action Model with Web-Scale Knowledge for Robot Manip- ulation. arXiv (2024). https://doi.org/10. 48550/arXiv.2410.06158 [306] Reuss, M., Zhou, H., R ̈uhle, M., Ya ̆gmurlu, ̈ O.E., Otto, F., Lioutikov, R.: FLOWER: Democratizing Generalist Robot Policies with Efficient Vision-Language-Action Flow Policies. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.04996 [307] Chi, X., Ge, K., Liu, J., Zhou, S., Jia, P., He, Z., et al.: MinD: Learning A Dual-System World Model for Real-Time Planning and Implicit Risk Analysis. arXiv (2025). https: //doi.org/10.48550/arXiv.2506.18897 [308] Chen, Y., Ge, Y., Tang, W., Li, Y., Ge, Y., Ding, M., et al.: Moto: Latent Motion Token as the Bridging Language for Learning Robot Manipulation from Videos. arXiv (2025). https://doi.org/10. 48550/arXiv.2412.04445 [309] Mandi,Z.,Bharadhwaj,H.,Moens, 58 V., Song, S., Rajeswaran, A., Kumar, V.: CACTI: A Framework for Scal- ableMulti-TaskMulti-SceneVisual ImitationLearning.arXiv(2023). https://doi.org/10.48550/arXiv.2212.05711 [310] Kang, J.H., Joshi, S., Dhanaraj, N., Gupta, S.K.: Task-Context-Aware Diffusion Policy with Language Guidance for Multi-task Dis- assembly. In: 2025 IEEE 21st International Conference on Automation Science and Engineering (CASE), p. 609–616. IEEE, Los Angeles, CA, USA (2025). https://doi. org/10.1109/CASE58245.2025.11163846 [311] Shentu, Y., Wu, P., Rajeswaran, A., Abbeel, P.: From LLMs to Actions: Latent Codes as Bridges in Hierarchical Robot Control. arXiv (2025). https://doi.org/10. 48550/arXiv.2405.04798 [312] Argus, M., Bratulic, J., Masnavi, H., Velikanov, M., Heppert, N., Valada, A., et al.: cVLA: Towards Efficient Camera- Space VLAs. arXiv (2025). https://doi.org/ 10.48550/arXiv.2507.02190 [313] Cheang, C., Chen, S., Cui, Z., Hu, Y., Huang, L., Kong, T., et al.: GR-3 Techni- cal Report. arXiv (2025). https://doi.org/ 10.48550/arXiv.2507.15493 [314] Xiong, Z., Li, K., Wang, Z., Jackson, M., Foerster, J., Whiteson, S.: HyperVLA: Effi- cient Inference in Vision-Language-Action Models via Hypernetworks. arXiv (2025). https://doi.org/10.48550/arXiv.2510.04898 [315] Chen, A., Yang, Y., Zhu, Z., Xu, K., Zhou, Z., Xiong, R., et al.: Toward Embod- iment Equivariant Vision-Language-Action Policy. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.14630 [316] Ji, Y., Tan, H., Shi, J., Hao, X., Zhang, Y., Zhang, H., et al.: RoboBrain: A Uni- fied Brain Model for Robotic Manipulation from Abstract to Concrete. arXiv (2025). https://doi.org/10.48550/arXiv.2502.21257 [317] Mao, J., Guan, J., Tang, Y., Hu, Q., Li, Z., Yu, J., et al.: OmniD: Generalizable Robot Manipulation Policy via Image-Based BEV Representation. arXiv (2025). https://doi. org/10.48550/arXiv.2508.11898 [318] Yu, J., Liu, H., Yu, Q., Ren, J., Hao, C., Ding, H., et al.: ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation. arXiv (2025). https://doi.org/10.48550/arXiv.2505.22159 [319] Chen, J., Zhao, W., Meng, Z., Mao, D., Song, R., Pan, W., et al.: Vision- Language Model Predictive Control for Manipulation Planning and Trajectory Gen- eration. arXiv (2025). https://doi.org/10. 48550/arXiv.2504.05225 [320] Collins, J.A., Cheng, L., Aneja, K., Wilcox, A., Joffe, B., Garg, A.: AMPLIFY: Action- less Motion Priors for Robot Learning from Videos. arXiv (2025). https://doi.org/10. 48550/arXiv.2506.14198 [321] Vo, T.V., Nguyen, T.Q., Nguyen, K.M., Nguyen, D.H.M., Vu, M.N.: ReFineVLA: Reasoning-Aware Teacher-Guided Transfer Fine-Tuning. arXiv (2025). https://doi.org/ 10.48550/arXiv.2505.19080 [322] Li, Q., Liang, Y., Wang, Z., Luo, L., Chen, X., Liao, M., et al.: CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation. arXiv (2024). https://doi.org/10.48550/arXiv.2411.19650 [323] Zhang, Z., Zheng, K., Chen, Z., Jang, J., Li, Y., Han, S., et al.: GRAPE: Generalizing Robot Policy via Preference Alignment. arXiv (2025). https://doi.org/ 10.48550/arXiv.2411.19309 [324] Yan, G., Wu, Y.-H., Wang, X.: DNAct: Diffusion Guided Multi-Task 3D Policy Learning. arXiv (2024). https://doi.org/10. 48550/arXiv.2403.04115 [325] Yang, J., Tan, R., Wu, Q., Zheng, R., Peng, B., Liang, Y., et al.: Magma: A Foundation Model for Multimodal AI Agents. arXiv (2025). https://doi.org/10. 48550/arXiv.2502.13130 59 [326] Cai, S., Zhang, B., Wang, Z., Lin, H., Ma, X., Liu, A., et al.: GROOT-2: Weakly Supervised Multi-Modal Instruction Follow- ing Agents. arXiv (2024). https://doi.org/ 10.48550/arXiv.2412.10410 [327] Liu, Z., Gu, Y., Zheng, S., Fu, Y., Xue, X., Jiang, Y.-G.: TriVLA: A Triple-System- BasedUnifiedVision-Language-Action Model with Episodic World Modeling for General Robot Control. arXiv (2025). https://doi.org/10.48550/arXiv.2507.01424 [328] Wang,B.,Sridhar,N.,Feng,C., Merwe,M.V.,Fishman,A.,Fazeli, N.,etal.:This&That:Language- GestureControlledVideoGeneration forRobotPlanning.arXiv(2025). https://doi.org/10.48550/arXiv.2407.05530 [329] Hu, S., Nagai, T., Horii, T.: TARAD: Task-AwareRobotAffordance-Centric Diffusion Policy Learned From LLM- GeneratedDemonstrations.IEEE RoboticsandAutomationLetters 10(10),10122–10129(2025)https: //doi.org/10.1109/LRA.2025.3598998 [330] Yu, A., Mooney, R.J.: Using Both Demon- strations and Language Instructions to Effi- ciently Learn Robotic Tasks. arXiv (2023). https://doi.org/10.48550/arXiv.2210.04476 [331] Shao, L., Migimatsu, T., Zhang, Q., Yang, K., Bohg, J.: Concept2Robot: Learning Manipulation Concepts from Instructions and Human Demonstrations (2020) [332] Jia, X., Wang, Q., Wang, A., Wang, H.A.,Gyenes,B.,Gospodinov,E., etal.:PointMapPolicy:Structured PointCloudProcessingforMulti- Modal Imitation Learning. arXiv (2025). https://doi.org/10.48550/arXiv.2510.20406 [333] Dalal, M., Liu, M., Talbott, W., Chen, C., Pathak, D., Zhang, J., et al.: Local Policies Enable Zero-Shot Long-Horizon Manipulation. In: 2025 IEEE International Conference on Robotics and Automation (ICRA), p. 13875–13882. IEEE, Atlanta, GA, USA (2025). https://doi.org/10.1109/ ICRA55743.2025.11128407 [334] Chen, H., Sun, B., Zhang, A., Pollefeys, M., Leutenegger, S.: VidBot: Learning Gen- eralizable 3D Actions from In-the-Wild 2D Human Videos for Zero-Shot Robotic Manipulation. arXiv (2025). https://doi. org/10.48550/arXiv.2503.07135 [335] Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., et al.: Do As I Can, Not As I Say: Grounding Language in Robotic Affordances. arXiv (2022). https:// doi.org/10.48550/arXiv.2204.01691 [336] Chen, S., Liu, J., Qian, S., Jiang, H., Li, L., Zhang, R., et al.: AC-DiT: Adap- tive Coordination Diffusion Transformer for Mobile Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2507.01961 [337] Zhao, W., Ding, P., Zhang, M., Gong, Z., Bai, S., Zhao, H., et al.: VLAS: Vision-Language-ActionModelWith SpeechInstructionsForCustomized RobotManipulation.arXiv(2025). https://doi.org/10.48550/arXiv.2502.13508 [338] Myers, V., Zheng, B.C., Mees, O., Levine, S., Fang, K.: Policy Adaptation via Lan- guage Optimization: Decomposing Tasks for Few-Shot Imitation. arXiv (2024). https:// doi.org/10.48550/arXiv.2408.16228 [339] Zheng, R., Liang, Y., Huang, S., Gao, J., Daum ́e, H., Kolobov, A., et al.: TraceVLA: Visual Trace Prompting Enhances Spatial- Temporal Awareness for Generalist Robotic Policies. arXiv (2025). https://doi.org/10. 48550/arXiv.2412.10345 [340] Chen, X., Wei, H., Zhang, P., Zhang, C., Wang, K., Guo, Y., et al.: Villa-X: Enhanc- ing Latent Action Modeling in Vision- Language-Action Models. arXiv (2025). https://doi.org/10.48550/arXiv.2507.23682 [341] Wang, Y., Zhang, Y., Huo, M., Tian, R., Zhang, X., Xie, Y., et al.: Sparse Diffusion Policy: A Sparse, Reusable, and Flexible Policy for Robot Learning. arXiv (2024). https://doi.org/10.48550/arXiv.2407.01531 60 [342] Spiridonov, A., Zaech, J.-N., Nikolov, N., Gool, L.V., Paudel, D.P.: General- ist Robot Manipulation beyond Action Labeled Data. arXiv (2025). https://doi. org/10.48550/arXiv.2509.19958 [343] Bharadhwaj, H., Dwibedi, D., Gupta, A., Tulsiani, S., Doersch, C., Xiao, T., et al.: Gen2Act: Human Video Generation in Novel Scenarios Enables Generalizable Robot Manipulation. arXiv (2024). https: //doi.org/10.48550/arXiv.2409.16283 [344] Yan, F., Liu, F., Zheng, L., Zhong, Y., Huang, Y., Guan, Z., et al.: RoboTron- Mani: All-in-One Multimodal Large Model for Robotic Manipulation. arXiv (2025). https://doi.org/10.48550/arXiv.2412.07215 [345] Zhang, J., Bai, C., He, H., Xia, W., Wang, Z., Zhao, B., et al.: SAM-E: Leveraging Visual Foundation Model with Sequence Imitation for Embodied Manip- ulation. arXiv (2024). https://doi.org/10. 48550/arXiv.2405.19586 [346] Liu, J., Chen, H., An, P., Liu, Z., Zhang, R., Gu, C., et al.: HybridVLA: Collaborative Diffusion and Autoregres- sion in a Unified Vision-Language-Action Model. arXiv (2025). https://doi.org/10. 48550/arXiv.2503.10631 [347] Yang, J., Shi, Y., Zhu, H., Liu, M., Ma, K., Wang, Y., et al.: CoMo: Learning Contin- uous Latent Motion from Internet Videos for Scalable Robot Learning. arXiv (2025). https://doi.org/10.48550/arXiv.2505.17006 [348] Zhang, J., Guo, Y., Chen, X., Wang, Y.- J., Hu, Y., Shi, C., et al.: HiRT: Enhancing Robotic Control with Hierarchical Robot Transformers. arXiv (2025). https://doi. org/10.48550/arXiv.2410.05273 [349] Zheng, R., Wang, J., Reed, S., Bjorck, J., Fang, Y., Hu, F., et al.: FLARE: Robot Learning with Implicit World Mod- eling. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.15659 [350] Tian, Y., Yang, S., Zeng, J., Wang, P., Lin, D., Dong, H., et al.: Predictive Inverse Dynamics Models Are Scalable Learners for Robotic Manipulation. arXiv (2024). https: //doi.org/10.48550/arXiv.2412.15109 [351] Wen, J., Zhu, M., Liu, J., Liu, Z., Yang, Y., Zhang, L., et al.: dVLA: Diffusion Vision- Language-Action Model with Multimodal Chain-of-Thought. arXiv (2025). https:// doi.org/10.48550/arXiv.2509.25681 [352] Yang, J., Tan, W., Jin, C., Yao, K., Liu, B., Fu, J., et al.: Transferring Foundation Models for Generalizable Robotic Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2306.05716 [353] Hou, Z., Zhang, T., Xiong, Y., Duan, H., Pu, H., Tong, R., et al.: Dita: Scaling Diffusion Transformer for Generalist Vision- Language-Action Policy. arXiv (2025). https://doi.org/10.48550/arXiv.2503.19757 [354] Yao, Y., Liu, S., Song, H., Qu, D., Chen, Q., Ding, Y., et al.: Think Small, Act Big: Primitive Prompt Learning for Lifelong Robot Manipulation (2025) https://doi.org/ 10.48550/arXiv.2504.00420 [355] Karamcheti, S., Nair, S., Chen, A., Kol- lar, T., Finn, C., Sadigh, D., et al.: Language-Driven Representation Learning for Robotics. In: Robotics: Science and Sys- tems XIX. Robotics: Science and Systems Foundation, ??? (2023). https://doi.org/10. 15607/RSS.2023.XIX.032 [356] Jiang, Y., Huang, S., Xue, S., Zhao, Y., Cen, J., Leng, S., et al.: RynnVLA-001: Using Human Demonstrations to Improve Robot Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2509.15212 [357] Lan, Z., Mao, W., Li, H., Wang, L., Wang, T., Fan, H., et al.: BFA: Best-Feature-Aware Fusion for Multi-View Fine-grained Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2502.11161 [358] Reuss, M., Ya ̆gmurlu, ̈ O.E., Wenzel, F., Lioutikov, R.: Multimodal Diffusion Trans- former: Learning Versatile Behavior from 61 Multimodal Goals. arXiv (2024). https:// doi.org/10.48550/arXiv.2407.05996 [359] Meng, Y., Bing, Z., Yao, X., Chen, K., Huang, K., Gao, Y., et al.: Pre- serving and Combining Knowledge in Robotic Lifelong Reinforcement Learn- ing. In Review (2024). https://doi.org/10. 21203/rs.3.rs-4353532/v1 [360] Ko, P.-C., Mao, J., Du, Y.: LEARN- ING TO ACT FROM ACTIONLESS VIDEOS THROUGH DENSE CORRE- SPONDENCES (2024) [361] Shafiullah, N.M.M., Rai, A., Etukuru, H., Liu, Y., Misra, I., Chintala, S., et al.: On Bringing Robots Home. arXiv (2023). https: //doi.org/10.48550/arXiv.2311.16098 [362] Qu, D., Song, H., Chen, Q., Yao, Y., Ye, X., Ding, Y., et al.: SpatialVLA: Exploring Spa- tial Representations for Visual-Language- Action Model. arXiv (2025). https://doi. org/10.48550/arXiv.2501.15830 [363] Deng, S., Yan, M., Wei, S., Ma, H., Yang, Y., Chen, J., et al.: GraspVLA: A Grasping Foundation Model Pre-trained on Billion- scale Synthetic Action Data. arXiv (2025). https://doi.org/10.48550/arXiv.2505.03233 [364] Yang, R., Chen, G., Wen, C., Gao, Y.: FP3: A 3D Foundation Policy for Robotic Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2503.08950 [365] Li, Q.: Task Reconstruction and Extrapo- lation for $π 0$ Using Text Latent. arXiv (2025).https://doi.org/10.48550/arXiv. 2505.03500 [366] Fu, M., Huang, H., Datta, G., Chen, L.Y., Panitch, W., Liu, F., et al.: ICRT: In- Context Imitation Learning via Next-Token Prediction. In: 2025 IEEE International Conference on Robotics and Automation (ICRA), p. 5937–5944. IEEE, Atlanta, GA, USA (2025). https://doi.org/10.1109/ ICRA55743.2025.11128272 [367] Bhat, V., Lan, Y.-H., Krishnamurthy, P., Karri, R., Khorrami, F.: 3D CAVLA: Lever- aging Depth and 3D Context to Gener- alize Vision Language Action Models for Unseen Tasks. arXiv (2025). https://doi. org/10.48550/arXiv.2505.05800 [368] Zhao, W., Li, G., Gong, Z., Ding, P., Zhao,H.,Wang,D.:Unveilingthe PotentialofVision-Language-Action ModelswithOpen-EndedMulti- modalInstructions.arXiv(2025). https://doi.org/10.48550/arXiv.2505.11214 [369] Huang, J., Yong, S., Ma, X., Linghu, X., Li, P., Wang, Y., et al.: An Embodied Gen- eralist Agent in 3D World. arXiv (2024). https://doi.org/10.48550/arXiv.2311.12871 [370] Li,W.,Zhang,R.,Shao,R.,He, J.,Nie,L.:CogVLA:Cognition- AlignedVision-Language-Action ModelviaInstruction-DrivenRout- ing&Sparsification.arXiv(2025). https://doi.org/10.48550/arXiv.2508.21046 [371] Song, W., Zhou, Z., Zhao, H., Chen, J., Ding, P., Yan, H., et al.: ReconVLA: Recon- structive Vision-Language-Action Model as Effective Robot Perceiver. arXiv (2025). https://doi.org/10.48550/arXiv.2508.10333 [372] Radosavovic, I., Xiao, T., James, S., Abbeel, P., Malik, J., Darrell, T.: Real-World Robot Learning with Masked Visual Pre- training. arXiv (2022). https://doi.org/10. 48550/arXiv.2210.03109 [373] Wen, J., Zhu, Y., Li, J., Tang, Z., Shen, C., Feng, F.: DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control. arXiv (2025). https://doi. org/10.48550/arXiv.2502.05855 [374] Mees, O., Borja-Diaz, J., Burgard, W.: Grounding Language with Visual Affor- dances over Unstructured Data. In: 2023 IEEE International Conference on Robotics and Automation (ICRA), p. 11576–11582. IEEE, London, United Kingdom (2023). https://doi.org/10.1109/ICRA48891.2023. 10160396 62 [375] Ze, Y., Zhang, G., Zhang, K., Hu, C., Wang, M., Xu, H.: 3D Diffusion Policy: Generaliz- able Visuomotor Policy Learning via Simple 3D Representations. arXiv (2024). https:// doi.org/10.48550/arXiv.2403.03954 [376] Luo, H., Feng, Y., Zhang, W., Zheng, S., Wang, Y., Yuan, H., et al.: Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos. arXiv (2025). https://doi.org/10.48550/arXiv.2507.15597 [377] Fang, H., Wang, C., Wang, Y., Chen, J., Xia, S., Lv, J., et al.: AirExo-2: Scaling up Gen- eralizable Robotic Imitation Learning with Low-Cost Exoskeletons. arXiv (2025). https: //doi.org/10.48550/arXiv.2503.03081 [378] Shridhar, M., Manuelli, L., Fox, D.: Perceiver-Actor: A Multi-Task Transformer for Robotic Manipulation. arXiv (2022). https://doi.org/10.48550/arXiv.2209.05451 [379] Bi, H., Wu, L., Lin, T., Tan, H., Su, Z., Su, H., et al.: H-RDT: Human Manipula- tion Enhanced Bimanual Robotic Manip- ulation. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.23523 [380] Yang, J., Zhu, H., Wang, Y., Wu, G., He, T., Wang, L.: Tra-MoE: Learning Trajectory Prediction Model from Multi- ple Domains for Adaptive Policy Condi- tioning. arXiv (2025). https://doi.org/10. 48550/arXiv.2411.14519 [381] Zhen, H., Qiu, X., Chen, P., Yang, J., Yan, X., Du, Y., et al.: 3D-VLA: A 3D Vision-Language-Action Generative World Model. arXiv (2024). https://doi.org/10. 48550/arXiv.2403.09631 [382] Ito, H., Ichiwara, H., Yamamoto, K., Mori, H., Ogata, T.: Integrated Learning of Robot Motion and Sentences: Real-Time Predic- tion of Grasping Motion and Attention based on Language Instructions. In: 2022 International Conference on Robotics and Automation (ICRA), p. 5404–5410. IEEE, Philadelphia, PA, USA (2022). https://doi. org/10.1109/ICRA46639.2022.9811815 [383] Wang, Y., Ding, P., Li, L., Cui, C., Ge, Z., Tong, X., et al.: VLA-Adapter: An Effective Paradigm for Tiny-Scale Vision-Language- Action Model. arXiv (2025). https://doi. org/10.48550/arXiv.2509.09372 [384] Mees, O.: What Matters in Language Con- ditioned Robotic Imitation Learning Over Unstructured Data (2022) [385] Kobayashi, M., Buamanee, T.: Bi-VLA: Bilateral Control-Based Imitation Learning via Vision-Language Fusion for Action Gen- eration. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.18865 [386] Lyu, J., Li, Z., Shi, X., Xu, C., Wang, Y., Wang, H.: DyWA: Dynamics-adaptive World Action Model for Generalizable Non-prehensile Manipulation. arXiv (2025). https://doi.org/10.48550/arXiv.2503.16806 [387] Hou, Z., Zhang, T., Xiong, Y., Pu, H., Zhao, C., Tong, R., et al.: Diffusion Trans- former Policy. arXiv (2025). https://doi. org/10.48550/arXiv.2410.15959 [388] Staroverov,A.,Gorodetsky,A.S., Krishtopik,A.S.,Izmesteva,U.A., Yudin,D.A.,Kovalev,A.K.,etal.: Fine-TuningMultimodalTransformer Models for Generating Actions in Vir- tualandRealEnvironments.IEEE Access 11, 130548–130559 (2023) https: //doi.org/10.1109/ACCESS.2023.3334791 [389] Bousmalis, K., Vezzani, G., Rao, D., Devin, C., Lee, A.X., Bauza, M., et al.: Robo- Cat: A Self-Improving Generalist Agent for Robotic Manipulation. arXiv (2023). https: //doi.org/10.48550/arXiv.2306.11706 [390] Chen, X., Guo, J., He, T., Zhang, C., Zhang, P., Yang, D.C., et al.: IGOR: Image-GOal Representations Are the Atomic Control Units for Foundation Models in Embodied AI. arXiv (2024). https://doi.org/10.48550/ arXiv.2411.00785 [391] Zhao, T.Z., Kumar, V., Levine, S., Finn, C.: Learning Fine-Grained Bimanual Manipula- tion with Low-Cost Hardware. arXiv (2023). 63 https://doi.org/10.48550/arXiv.2304.13705 [392] Dong, Y., Ge, H., Zeng, Y., Zhang, J., Tian, B., Zhu, H., et al.: ImitDiff: Transferring Foundation-Model Priors for Distraction- Robust Visuomotor Policy. IEEE Robotics and Automation Letters 10(11), 12125– 12132 (2025) https://doi.org/10.1109/LRA. 2025.3619835 [393] Ren, P., Zhang, K., Zheng, H., Li, Z., Wen, Y., Zhu, F., et al.: Surfer: Progressive Reasoning with World Models for Robotic Manipulation. arXiv (2024). https://doi. org/10.48550/arXiv.2306.11335 [394] Mazzaglia, P., Sancaktar, C., Peschl, M., Dijkman, D.: Hybrid Training for Vision- Language-Action Models. arXiv (2025). https://doi.org/10.48550/arXiv.2510.00600 [395] Kim, M.J., Finn, C., Liang, P.: Fine-Tuning Vision-Language-Action Models: Optimiz- ing Speed and Success. arXiv (2025). https: //doi.org/10.48550/arXiv.2502.19645 [396] Huang, K., Li, T., Liu, Y., Zhang, Z.,Wang,J.,Han,L.:VLM-TDP: VLM-guidedTrajectory-conditioned DiffusionPolicyforRobustLong- HorizonManipulation.arXiv(2025). https://doi.org/10.48550/arXiv.2507.04524 [397] Myers, V., He, A., Fang, K., Walke, H., Hansen-Estruch, P., Cheng, C.-A., et al.: Goal Representations for Instruction Follow- ing: A Semi-Supervised Language Interface to Control. arXiv (2023). https://doi.org/ 10.48550/arXiv.2307.00117 [398] Goyal, A., Xu, J., Guo, Y., Blukis, V., Chao, Y.-W., Fox, D.: RVT: Robotic View Transformer for 3D Object Manip- ulation. arXiv (2023). https://doi.org/10. 48550/arXiv.2306.14896 [399] Liu, H., Lee, L., Lee, K., Abbeel, P.: Instruction-Following Agents with Multi- modal Transformer. arXiv (2023). https:// doi.org/10.48550/arXiv.2210.13431 [400] Ma, J., Qin, Y., Li, Y., Liao, X., Guo, Y., Zhang, R.: CDP: Towards Robust Autore- gressive Visuomotor Policy Learning via Causal Diffusion. arXiv (2025). https://doi. org/10.48550/arXiv.2506.14769 [401] Fan, S., Yang, Q., Liu, Y., Wu, K., Che, Z., Liu, Q., et al.: Diffusion Trajectory-Guided Policy for Long-Horizon Robot Manipula- tion (2025) https://doi.org/10.1109/LRA. 2025.3619794 [402] Zhong, Z., Yan, H., Li, J., Liu, X., Gong, X., Zhang, T., et al.: FlowVLA: Visual Chain of Thought-based Motion Rea- soning for Vision-Language-Action Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2508.18269 [403] Tan, X., Yang, Y., Ye, P., Zheng, J., Bai, B., Wang, X., et al.: Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision- Language-Action Models. arXiv (2025). https://doi.org/10.48550/arXiv.2505.21200 [404] Li, J., Wang, Z., Zheng, J., Zhou, X., Wang, G., Song, G., et al.: Robo-MUTUAL: Robotic Multimodal Task Specification via Unimodal Learning. arXiv (2024). https:// doi.org/10.48550/arXiv.2410.01529 [405] Gao, C., Zhang, H., Xu, Z., Cai, Z., Shao, L.: FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model. arXiv (2025). https://doi.org/10. 48550/arXiv.2412.08261 [406] Deng, Y., Mo, K., Xia, C., Wang, X.: Learn- ingLanguage-ConditionedDeformable Object Manipulation with Graph Dynam- ics. In: 2024 IEEE International Conference on Robotics and Automation (ICRA), p. 7508–7514. IEEE, Yokohama, Japan (2024). https://doi.org/10.1109/ICRA57147.2024. 10610890 [407] Song, W., Chen, J., Ding, P., Zhao, H., Zhao, W., Zhong, Z., et al.: Accelerat- ing Vision-Language-Action Model Inte- grated with Action Chunking via Parallel Decoding. arXiv (2025). https://doi.org/10. 48550/arXiv.2503.02310 64 [408] Yang, J., Glossop, C., Bhorkar, A., Shah, D., Vuong, Q., Finn, C., et al.: Pushing the Limits of Cross-Embodiment Learning for Manipulation and Navigation. arXiv (2024). https://doi.org/10.48550/arXiv.2402.19432 [409] Doshi,R.,Walke,H.,Mees,O., Dasari, S., Levine, S.: Scaling Cross- EmbodiedLearning:OnePolicyfor Manipulation,Navigation,Locomo- tionandAviation.arXiv(2024). https://doi.org/10.48550/arXiv.2408.11812 [410] Zhang, H., Zhuang, Z., Zhao, H., Ding, P., Lu, H., Wang, D.: ReinboT: Amplify- ing Robot Visual-Language Manipulation with Reinforcement Learning. arXiv (2025). https://doi.org/10.48550/arXiv.2505.07395 [411] Shen, W., Liu, Y., Wu, Y., Liang, Z., Gu, S., Wang, D., et al.: Expertise Need Not Monopolize: Action-Specialized Mix- ture of Experts for Vision-Language-Action Learning. arXiv (2025). https://doi.org/10. 48550/arXiv.2510.14300 [412] Griebenow, S.: Open-Vocabulary Robotic Object Manipulation using Foundation Models (2025) [413] Ma, T., Zhou, J., Wang, Z., Qiu, R., Liang, J.: Contrastive Imitation Learning for Language-guided Multi-Task Robotic Manipulation. arXiv (2024). https://doi. org/10.48550/arXiv.2406.09738 [414] Ren, P., Zhang, K., Zheng, H., Li, Z., Wen, Y., Zhu, F., et al.: Surfer: A World Model-Based Framework for Vision- Language Robot Manipulation (2025) [415] He, H., Bai, C., Pan, L., Zhang, W., Zhao, B., Li, X.: Learning an Actionable Discrete Diffusion Policy via Large-Scale Actionless Video Pre-Training. arXiv (2024). https:// doi.org/10.48550/arXiv.2402.14407 [416] Wen, J., Zhu, Y., Zhu, M., Li, J., Xu, Z., Che, Z., et al.: Object-Centric Instruc- tion Augmentation for Robotic Manipu- lation. arXiv (2024). https://doi.org/10. 48550/arXiv.2401.02814 [417] Thomas, G., Cheng, C.-A., Loynd, R., Fru- jeri, F.V., Vineet, V., Jalobeanu, M., et al.: PLEX: Making the Most of the Avail- able Data for Robotic Manipulation Pre- training. arXiv (2023). https://doi.org/10. 48550/arXiv.2303.08789 [418] Huang, S., Chen, L., Zhou, P., Chen, S., Jiang, Z., Hu, Y., et al.: EnerVerse: Envi- sioning Embodied Future Space for Robotics Manipulation. arXiv (2025). https://doi. org/10.48550/arXiv.2501.01895 [419] Zhang, X., Liu, Y., Chang, H., Schramm, L., Boularias, A.: Autoregressive Action Sequence Learning for Robotic Manipu- lation. arXiv (2025). https://doi.org/10. 48550/arXiv.2410.03132 [420] Huang, H., Chen, X., Chen, Y., Li, H., Han, X., Wang, Z., et al.: RoboGround: RoboticManipulationwithGrounded Vision-LanguagePriors.arXiv(2025). https://doi.org/10.48550/arXiv.2504.21530 [421] Lei, Y., Mao, S., Zhou, S., Zhang, H., Li, X., Luo, P.: Dynamic Mixture of Progressive Parameter-Efficient Expert Library for Life- long Robot Learning. arXiv (2025). https: //doi.org/10.48550/arXiv.2506.05985 [422] Zhu, M., Zhu, Y., Li, J., Zhou, Z., Wen, J., Liu, X., et al.: ObjectVLA: End-to-End Open-World Object Manipulation Without Demonstration. arXiv (2025). https://doi. org/10.48550/arXiv.2502.19250 [423] Liu, F., Yan, F., Zheng, L., Feng, C., Huang, Y., Ma, L.: RoboUniView: Visual-Language Model with Unified View Representation for Robotic Manipulation. arXiv (2024). https: //doi.org/10.48550/arXiv.2406.18977 [424] Wen, Y., Gu, K., Liu, H., Zhao, Y., Wang, T., Fan, H., et al.: ROSA: Harness- ing Robot States for Vision-Language and Action Alignment. arXiv (2025). https:// doi.org/10.48550/arXiv.2506.13679 [425] Zhao, Q., Lu, Y., Kim, M.J., Fu, Z., Zhang, 65 Z., Wu, Y., et al.: CoT-VLA: Visual Chain- of-Thought Reasoning for Vision-Language- Action Models. arXiv (2025). https://doi. org/10.48550/arXiv.2503.22020 [426] Tian, J., Wang, L., Zhou, S., Wang, S., Li, J., Sun, H., et al.: PDFactor: Learn- ing Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic Manipula- tion. In: 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 15757–15767. IEEE, Nashville, TN, USA (2025). https://doi.org/10.1109/ CVPR52734.2025.01469 [427] Liu, I.-C.A., He, S., Seita, D., Sukhatme, G.: VoxAct-B: Voxel-Based Acting and Stabilizing Policy for Bimanual Manip- ulation. arXiv (2024). https://doi.org/10. 48550/arXiv.2407.04152 [428] Fan, Y., Ding, P., Bai, S., Tong, X., Zhu, Y., Lu, H., et al.: Long-VLA: Unleash- ing Long-Horizon Capability of Vision Lan- guage Action Model for Robot Manipula- tion (2025) https://doi.org/10.48550/arXiv. 2508.19958 [429] Ghosh, D., Walke, H., Pertsch, K., Black, K., Mees, O., Dasari, S., et al.: Octo: An Open-Source Generalist Robot Policy. In: Robotics: Science and Systems X. Robotics: Science and Systems Foundation, ??? (2024). https://doi.org/10.15607/RSS. 2024.X.090 [430] Caesar, A., ̈ Ozdemir, O., Weber, C., Wermter, S.: Enabling action crossmodality for a pretrained large language model. Nat- ural Language Processing Journal 7, 100072 (2024) https://doi.org/10.1016/j.nlp.2024. 100072 [431] Ge, Y., Macaluso, A., Li, L.E., Luo, P.,Wang,X.:PolicyAdaptation from Foundation Model Feedback. In: 2023 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), p. 19059–19069. IEEE, Van- couver,BC,Canada(2023).https: //doi.org/10.1109/CVPR52729.2023.01827 [432] Ding, P., Ma, J., Tong, X., Zou, B., Luo, X., Fan, Y., et al.: Humanoid-VLA: Towards Universal Humanoid Control with Visual Integration. arXiv (2025). https://doi.org/ 10.48550/arXiv.2502.14795 [433] Wang, S., Liu, S., Wang, W., Shan, J., Fang, B.: RoboBERT: An End-to-end Multimodal Robotic Manipulation Model. arXiv (2025). https://doi.org/10.48550/arXiv.2502.07837 [434] Zhu, R., Sun, E., Huang, G., Celiktutan, O.: Efficient Continual Adaptation of Pre- trained Robotic Policy with Online Meta- Learned Adapters. arXiv (2025). https:// doi.org/10.48550/arXiv.2503.18684 [435] Yang, R., Yu, Q., Wu, Y., Yan, R., Li, B., Cheng, A.-C., et al.: EgoVLA: Learning Vision-Language-Action Models from Ego- centric Human Videos. arXiv (2025). https: //doi.org/10.48550/arXiv.2507.12440 [436] Silva, A., Moorman, N., Silva, W., Zaidi, Z., Gopalan, N., Gombolay, M.: LanCon- Learn: Learning With Language to Enable Generalization in Multi-Task Manipulation. IEEE Robotics and Automation Letters 7(2), 1635–1642 (2022) https://doi.org/10. 1109/LRA.2021.3139667 [437] James, S., Wada, K., Laidlow, T., Davison, A.J.: Coarse-to-Fine Q-attention: Efficient Learning for Visual Robotic Manipulation via Discretisation. arXiv (2022). https:// doi.org/10.48550/arXiv.2106.12534 [438] Nguyen, E.-R., Zhang, Y., Ranasinghe, K., Li, X., Ryoo, M.S.: Pixel Motion Diffusion Is What We Need for Robot Control. arXiv (2025). https://doi.org/10. 48550/arXiv.2509.22652 [439] Liu, M., Shu, J., Chen, H., Li, Z., Zhao, C., Yang, J., et al.: StaMo: Unsuper- vised Learning of Generalizable Robot Motion from Compact State Represen- tation. arXiv (2025). https://doi.org/10. 48550/arXiv.2510.05057 [440] Davies, T., Huang, Y., Liu, Y., Chen, X., 66 Liu, H., Hu, L.: Tenma: Robust Cross- Embodiment Robot Manipulation with Dif- fusion Transformer. arXiv (2025). https:// doi.org/10.48550/arXiv.2509.11865 [441] Wang, Y., Zhu, H., Liu, M., Yang, J., Fang, H.-S., He, T.: VQ-VLA: Improving Vision- Language-Action Models via Scaling Vector- Quantized Action Tokenizers. arXiv (2025). https://doi.org/10.48550/arXiv.2507.01016 [442] Lu, W., Wu, X., Gao, S., He, W., Zhao, Q., Zhang, L., et al.: Research on Task Decom- position and Motion Trajectory Optimiza- tion of Robotic Arm Based on VLA Large Model. In: 2024 5th International Confer- ence on Machine Learning and Computer Application (ICMLCA), p. 80–85. IEEE, Hangzhou, China (2024). https://doi.org/ 10.1109/ICMLCA63499.2024.10754333 [443] Yang, G., Zhang, T., Hao, H., Wang, W., Liu, Y., Wang, D., et al.: Vlaser: Vision- Language-Action Model with Synergistic Embodied Reasoning. arXiv (2025). https: //doi.org/10.48550/arXiv.2510.11027 [444] Wei, Y., Sun, Y., Zheng, R., Vemprala, S., Bonatti, R., Chen, S., et al.: Is Imita- tion All You Need? Generalized Decision- Making with Dual-Phase Training. In: 2023 IEEE/CVF International Conference on Computer Vision (ICCV), p. 16175–16185. IEEE, Paris, France (2023). https://doi.org/ 10.1109/ICCV51070.2023.01486 [445] Roy, K., Dissanayake, A., Tidd, B., Moghadam,P.:M2Distill:Multi- ModalDistillationforLifelong ImitationLearning.arXiv(2025). https://doi.org/10.48550/arXiv.2410.00064 [446] Wen, Y., Lin, J., Zhu, Y., Han, J., Xu, H., Zhao, S., et al.: VidMan: Exploiting Implicit Dynamics from Video Diffusion Model for Effective Robot Manipulation. arXiv (2024). https://doi.org/10.48550/arXiv.2411.09153 [447] Han, B., Kim, J., Jang, J.: A Dual Process VLA: Efficient Robotic Manipulation Lever- aging VLM. arXiv (2024). https://doi.org/ 10.48550/arXiv.2410.15549 [448] Lv, Q., Kong, W., Li, H., Zeng, J., Qiu, Z., Qu, D., et al.: F1: A Vision-Language- Action Model Bridging Understanding and Generation to Actions. arXiv (2025). https: //doi.org/10.48550/arXiv.2509.06951 [449] Zhang, J., Chen, Y., Xu, Y., Huang, Z.,Zhou,Y.,Yuan,Y.-J.,etal.: 4D-VLA:SpatiotemporalVision- Language-ActionPretrainingwith Cross-Scene Calibration. arXiv (2025). https://doi.org/10.48550/arXiv.2506.22242 [450] Gervet, T., Xian, Z., Gkanatsios, N., Fragki- adaki, K.: Act3D: 3D Feature Field Trans- formers for Multi-Task Robotic Manip- ulation. arXiv (2023). https://doi.org/10. 48550/arXiv.2306.17817 [451] Zhang, J., Guo, Y., Hu, Y., Chen, X., Zhu, X., Chen, J.: UP-VLA: A Unified Under- standing and Prediction Model for Embod- ied Agent. arXiv (2025). https://doi.org/10. 48550/arXiv.2501.18867 [452] Zhuo, F., He, Y., Yu, F., Li, P., Zhao, Z., Sun, X.: ABM: Attention before Manipulation. In: Proceedings of the Thirty-ThirdInternationalJointCon- ferenceonArtificialIntelligence,p. 1816–1824.InternationalJointCon- ferencesonArtificialIntelligence Organization, Jeju, South Korea (2024). https://doi.org/10.24963/ijcai.2024/201 [453] Cen,J.,Yu,C.,Yuan,H.,Jiang, Y.,Huang,S.,Guo,J.,etal.: WorldVLA:TowardsAutoregressive ActionWorldModel.arXiv(2025). https://doi.org/10.48550/arXiv.2506.21539 [454] Wang, Y., Li, X., Wang, W., Zhang, J., Li, Y., Chen, Y., et al.: Unified Vision- Language-Action Model. arXiv (2025). https://doi.org/10.48550/arXiv.2506.19850 [455] Wang, H., Xiong, C., Wang, R., Chen, X.: BitVLA: 1-Bit Vision-Language-Action Models for Robotics Manipulation. arXiv (2025).https://doi.org/10.48550/arXiv. 2506.07530 67 [456] Zhou, Z., Zhu, Y., Wen, J., Shen, C.,Xu,Y.:ChatVLA-2:Vision- Language-Action Model with Open-World EmbodiedReasoningfromPre- trainedKnowledge.arXiv(2025). https://doi.org/10.48550/arXiv.2505.21906 [457] Yue, Y., Wang, Y., Kang, B., Han, Y., Wang, S., Song, S., et al.: DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Exe- cution. arXiv (2024). https://doi.org/10. 48550/arXiv.2411.02359 [458] Patratskiy,M.A.,Kovalev,A.K., Panov, A.I.: Spatial Traces: Enhanc- ing VLA Models with Spatial-Temporal Understanding.arXiv(2025).https: //doi.org/10.48550/arXiv.2508.09032 [459] Dey, S., Zaech, J.-N., Nikolov, N., Gool, L.V., Paudel, D.P.: ReVLA: Reverting Visual Domain Limitation of Robotic Foun- dation Models. arXiv (2025). https://doi. org/10.48550/arXiv.2409.15250 [460] Jin, C., Tan, W., Yang, J., Liu, B., Song, R., Wang, L., et al.: AlphaBlock: Embodied Finetuning for Vision-Language Reasoning in Robot Manipulation. arXiv (2023). https: //doi.org/10.48550/arXiv.2305.18898 [461] Bonatti, R., Vemprala, S., Ma, S., Fru- jeri, F., Chen, S., Kapoor, A.: PACT: Perception-ActionCausalTransformer for Autoregressive Robotics Pre-Training. In: 2023 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS), p. 3621–3627. IEEE, Detroit, MI, USA (2023). https://doi.org/10.1109/ IROS55552.2023.10342381 [462] Xu, R., Shen, Y., Li, X., Wu, R., Dong, H.: NaturalVLM: Leveraging Fine-grained Natural Language for Affordance-Guided Visual Manipulation. arXiv (2024). https: //doi.org/10.48550/arXiv.2403.08355 [463] Cheng,Z.,Zhang,Y.,Zhang,W., Li, H., Wang, K., Song, L., et al.: OmniVTLA:Vision-Tactile-Language- ActionModelwithSemantic-Aligned TactileSensing.arXiv(2025).https: //doi.org/10.48550/arXiv.2508.08706 [464] Driess, D., Springenberg, J.T., Ichter, B., Yu, L., Li-Bell, A., Pertsch, K., et al.: Knowledge Insulating Vision-Language- Action Models: Train Fast, Run Fast, Gen- eralize Better. arXiv (2025). https://doi. org/10.48550/arXiv.2505.23705 [465] Fang,X.,Huang,B.-R.,Mao,J., Shone, J., Tenenbaum, J.B., Lozano- P ́erez, T., et al.: Keypoint Abstraction Using Large Models for Object-Relative ImitationLearning.arXiv(2024). https://doi.org/10.48550/arXiv.2410.23254 [466] Jiang, T., Jiang, X., Ma, Y., Wen, X., Li, B., Zhan, K., et al.: The Better You Learn, The Smarter You Prune: Towards Efficient Vision-language-action Models via Differen- tiable Token Pruning. arXiv (2025). https: //doi.org/10.48550/arXiv.2509.12594 [467] Su, Y., Zhan, X., Fang, H., Li, Y.-L., Lu, C., Yang, L.: Motion Before Action: Diffusing Object Motion as Manipulation Condition. arXiv (2025). https://doi.org/ 10.48550/arXiv.2411.09658 [468] Song, H., Le, L.: Avi: Action from Volumet- ric Inference. arXiv (2025). https://doi.org/ 10.48550/arXiv.2510.21746 [469] Zhou, Z., Zhu, Y., Zhu, M., Wen, J., Liu, N., Xu, Z., et al.: ChatVLA: UnifiedMultimodalUnderstanding andRobotControlwithVision- Language-Action Model. arXiv (2025). https://doi.org/10.48550/arXiv.2502.14420 [470] Li, J., Gao, Q., Johnston, M., Gao, X., He, X., Shakiah, S., et al.: Mastering Robot Manipulation with Multimodal Prompts through Pretraining and Multi-task Fine- tuning. arXiv (2024). https://doi.org/10. 48550/arXiv.2310.09676 [471] Bing, Z., Koch, A., Yao, X., Huang, K., Knoll, A.: Meta-Reinforcement Learning via Language Instructions. arXiv (2022). https: //doi.org/10.48550/arXiv.2209.04924 68 [472] Tan, W., Wang, B., Zhi, H., Liu, C., Li, Z., Liu, J., et al.: BLM$1$: A Boundless Large Model for Cross-Space, Cross-Task, and Cross-Embodiment Learning. arXiv (2025). https://doi.org/10.48550/arXiv.2510.24161 [473] Hu, Y., Song, P., Wen, K., Detry, R.: Mini Diffuser: Fast Multi-task Diffusion Policy Training Using Two-level Mini- batches. arXiv (2025). https://doi.org/10. 48550/arXiv.2505.09430 [474] Wu, K., Zhu, Y., Li, J., Wen, J., Liu, N., Xu, Z., et al.: Discrete Policy: Learn- ing Disentangled Action Space for Multi- Task Robotic Manipulation. In: 2025 IEEE International Conference on Robotics and Automation (ICRA), p. 8811–8818. IEEE, Atlanta, GA, USA (2025). https://doi.org/ 10.1109/ICRA55743.2025.11127630 [475] Physical, I.: PI06 [476] Hung, C.-Y., Sun, Q., Hong, P., Zadeh, A., Li, C., Tan, U.-X., et al.: NORA: A Small Open-Sourced Generalist Vision Language Action Model for Embodied Tasks. arXiv (2025). https://doi.org/10. 48550/arXiv.2504.19854 [477] Li, Z., Ren, L., Yang, J., Zhao, Y., Wu, X., Xu, Z., et al.: VIRT: Vision Instructed Transformer for Robotic Manip- ulation. arXiv (2024). https://doi.org/10. 48550/ARXIV.2410.07169 [478] Chen, S., Garcia, R., Schmid, C., Laptev, I.: PolarNet: 3D Point Clouds for Language- GuidedRoboticManipulation.arXiv (2023).https://doi.org/10.48550/arXiv. 2309.15596 [479] Zhang, R., Dong, M., Zhang, Y., Heng, L., Chi, X., Dai, G., et al.: MoLe-VLA: Dynamic Layer-skipping Vision Language Action Model via Mixture-of-Layers for Effi- cient Robot Manipulation. arXiv (2025). https://doi.org/10.48550/arXiv.2503.20384 [480] AgiBot-World-Contributors, Bu, Q., Cai, J., Chen, L., Cui, X., Ding, Y., et al.: AgiBot World Colosseo: A Large-scale Manipula- tion Platform for Scalable and Intelligent Embodied Systems. arXiv (2025). https:// doi.org/10.48550/arXiv.2503.06669 [481] Chebotar, Y., Vuong, Q., Irpan, A., Haus- man, K., Xia, F., Lu, Y., et al.: Q- Transformer: Scalable Offline Reinforce- ment Learning via Autoregressive Q- Functions. arXiv (2023). https://doi.org/10. 48550/arXiv.2309.10150 [482] Wen, Y., Li, H., Gu, K., Zhao, Y., Wang, T., Sun, X.: LLaDA-VLA: Vision Language Dif- fusion Action Models. arXiv (2025). https: //doi.org/10.48550/arXiv.2509.06932 [483] Liang,W.,Sun,G.,He,Y.,Ren, Y.,Dong,J.,Cong,Y.:Never- EndingBehavior-CloningAgentfor RoboticManipulation.arXiv(2025). https://doi.org/10.48550/arXiv.2403.00336 [484] Li, H., Yan, S., Li, Y., Wang, X.: Towards Fast, Memory-based and Data-Efficient Vision-LanguagePolicy.arXiv(2025). https://doi.org/10.48550/arXiv.2503.10322 [485] Zheng, B., Zhou, J., Chen, F.: Inter- pretable Robotic Manipulation from Lan- guage. arXiv (2024). https://doi.org/10. 48550/arXiv.2405.17047 [486] Liu, J., Liu, M., Wang, Z., An, P., Li, X., Zhou, K., et al.: RoboMamba: Efficient Vision-Language-Action Model for Robotic Reasoning and Manipulation. arXiv (2024). https://doi.org/10.48550/arXiv.2406.04339 [487] Jang, H., Yu, S., Kwon, H., Jeon, H., Seo, Y., Shin, J.: ContextVLA: Vision-Language- Action Model with Amortized Multi-Frame Context. arXiv (2025). https://doi.org/10. 48550/arXiv.2510.04246 [488] Koo, J., Cho, T., Kang, H., Pyo, E., Oh, T.G., Kim, T., et al.: RetoVLA: Reusing Register Tokens for Spatial Rea- soning in Vision-Language-Action Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2509.21243 69 [489] Zhang, T., Duan, H., Hao, H., Qiao, Y., Dai, J., Hou, Z.: Grounding Actions in Camera Space: Observation-Centric Vision- Language-Action Policy. arXiv (2025). https://doi.org/10.48550/arXiv.2508.13103 [490] Li,P.,Wu,Y.,Xi,Z.,Li,W., Huang, Y., Zhang, Z., et al.: Con- trolVLA:Few-shotObject-centric AdaptationforPre-trainedVision- Language-Action Models. arXiv (2025). https://doi.org/10.48550/arXiv.2506.16211 [491] Dasari, S., Mees, O., Zhao, S., Sri- rama, M.K., Levine, S.: The Ingredi- ents for Robotic Diffusion Transform- ers. arXiv (2024). https://doi.org/10.48550/ arXiv.2410.10088 [492] Chen, Y., Li, P., Huang, Y., Yang, J., Chen, K., Wang, L.: EC-Flow: Enabling Ver- satile Robotic Manipulation from Action- Unlabeled Videos via Embodiment-Centric Flow. arXiv (2025). https://doi.org/10. 48550/arXiv.2507.06224 [493] Xian, Z., Gkanatsios, N., Gervet, T., Ke, T.-W., Fragkiadaki, K.: ChainedDiffuser: Unifying Trajectory Diffusion and Keypose Prediction for Robotic Manipulation (2023) [494] Li, Y., Zhang, T., Wei, X., Li, J., Zhao, L., Huang, D., et al.: Object-Focus Actor for Data-efficient Robot Generalization Dex- terous Manipulation. arXiv (2025). https: //doi.org/10.48550/arXiv.2505.15098 [495] Sejnova, G., Vavrecka, M., Stepanova, K.: Bridging Language, Vision and Action: Multimodal VAEs in Robotic Manipulation Tasks. In: 2024 IEEE/RSJ International Conference on Intelligent Robots and Sys- tems (IROS), p. 12522–12528 (2024). https://doi.org/10.1109/IROS58592.2024. 10802160 [496] Abouzeid, A., Mansour, M., Sun, Z., Song,D.:GeoAware-VLA:Implicit GeometryAwareVision-Language- ActionModel.arXiv(2025).https: //doi.org/10.48550/arXiv.2509.14117 [497] Song, W., Chen, J., Ding, P., Huang, Y., Zhao, H., Wang, D., et al.: CEED-VLA: Consistency Vision-Language-Action Model with Early-Exit Decoding. arXiv (2025). https://doi.org/10.48550/arXiv.2506.13725 [498] Wang, S., Yu, R., Yuan, Z., Yu, C., Gao, F., Wang, Y., et al.: Spec-VLA: Specula- tive Decoding for Vision-Language-Action Models with Relaxed Acceptance. arXiv (2025). https://doi.org/10.48550/ARXIV. 2507.22424 [499] Li, X., Wang, S., Chen, C., Wei, H., Shi, Y., Mo, H.: RoboFlamingo-Plus: Fusion of Depth and RGB Perception with Vision- Language Models for Enhanced Robotic Manipulation. In: 2025 IEEE International Conference on Real-time Computing and Robotics (RCAR), p. 311–316. IEEE, Toyama, Japan (2025). https://doi.org/10. 1109/RCAR65431.2025.11139480 [500] Nguyen, N., Vu, M.N., Ta, T.D., Huang, B., Vo, T., Le, N., et al.: Robotic-CLIP: Fine- tuning CLIP on Action Data for Robotic Applications. arXiv (2024). https://doi.org/ 10.48550/arXiv.2409.17727 [501] Budzianowski,P.,Maa,W.,Freed, M.,Mo,J.,Hsiao,W.,Xie,A., etal.:EdgeVLA:EfficientVision- Language-Action Models. arXiv (2025). https://doi.org/10.48550/arXiv.2507.14049 [502] Niu, D., Sharma, Y., Biamby, G., Quenum, J., Bai, Y., Shi, B., et al.: LLARVA: Vision- Action Instruction Tuning Enhances Robot Learning. arXiv (2024). https://doi.org/10. 48550/arXiv.2406.11815 [503] Ge, S., Zhang, Y., Xie, S., Zhang, W., Zhou, M., Wang, Z.: VGGT-DP: Generalizable Robot Control via Vision Foundation Mod- els. arXiv (2025). https://doi.org/10.48550/ arXiv.2509.18778 [504] Zhu, M., Zhu, Y., Li, J., Wen, J., Xu, Z., Liu, N., et al.: Scaling Diffusion Policy in Trans- former to 1 Billion Parameters for Robotic Manipulation. In: 2025 IEEE International Conference on Robotics and Automation 70 (ICRA), p. 10838–10845. IEEE, Atlanta, GA, USA (2025). https://doi.org/10.1109/ ICRA55743.2025.11128074 Appendix A Related Works 71 Table A1: Full list of identified related works/ surveys. Total 55 publications from survey corpus as depicted in Figure 9. RefTitleYear↓ [170]A Review of Embodied Grasping2025 [171] A Survey of Behavior Foundation Model: Next-Generation Whole-Body Control System of Humanoid Robots 2025 [13]A Survey on Vision-Language-Action Models for Embodied AI2025 [17]A Survey on Vision-Language-Action Models: An Action Tokenization Perspective2025 [172]A Taxonomy for Evaluating Generalist Robot Policies2025 [173]Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI2025 [16] An Anatomy of Vision-Language-Action Models: From Modules to Milestones and Challenges 2025 [174] Applications of Large Language Models and Multimodal Large Models in Autonomous Driving: A Comprehensive Review 2025 [15]Bridging Language and Action: A Survey of Language-Conditioned Robot Manipulation2025 [175]Embodied AI with Foundation Models for Mobile Service Robots: A Systematic Review2025 [176]Embodied AI: Bridging Simulation and Reality in Robotics2025 [177]Embodied AI: From LLMs to World Models2025 [178]Exploring Embodied Multimodal Large Models: Development2025 [179]Foundation Model Driven Robotics: A Comprehensive Review2025 [180] From Grounding to Manipulation: Case Studies of Foundation Model Integration in Embodied Robotic Systems 2025 [181]Generative Artificial Intelligence in Robotic Manipulation: A Survey2025 [182]Integrating Large Language Models into Robotic Autonomy: A Review of Motion2025 [183]Large Language Models and 3D Vision for Intelligent Robotic Perception and Autonomy2025 [184] Large Model Empowered Embodied AI: A Survey on Decision-Making and Embodied Learning 2025 [30]Large VLM-based Vision-Language-Action Models for Robotic Manipulation: A Survey2025 [185] Multi-Modal Multi-Task (M3T) Federated Foundation Models for Embodied AI: Potentials and Challenges for Edge Integration 2025 [186]Multimodal Fusion and Vision-Language Models: A Survey for Robot Vision2025 [187]Neural Brain: A Neuroscience-inspired Framework for Embodied Agents2025 [188]Neural Scaling Laws in Robotics2025 [31]OpenHelix: A Short Survey2025 [14]Parallels Between VLA Model Post-Training and Human Motor Learning: Progress2025 [32]Pure Vision Language Action (VLA) Models: A Comprehensive Survey2025 [189]Toward Embodied AGI: A Review of Embodied AI and the Road Ahead2025 [190] Towards Embodied Agentic AI: Review and Classification of LLM- and VLM-Driven Robot Autonomy and Interaction 2025 [33] Towards Generalist Robot Policies: What Matters in Building Vision-Language-Action Models 2025 continued on next page 72 RefTitleYear [35] Vision-Language Model-Based Human-Robot Collaboration for Smart Manufacturing: A State-of-the-Art Survey 2025 [34]Vision-Language-Action Models: Concepts, Progress, Applications and Challenges2025 [191]What Foundation Models Can Bring for Robot Learning in Manipulation : A Survey2025 [192]A Survey for Foundation Models in Autonomous Driving2024 [193]A Survey of Robot Intelligence with Large Language Models2024 [194]A Survey of Robotic Language Grounding: Tradeoffs between Symbols and Embeddings2024 [195]A Survey on Integration of Large Language Models with Intelligent Robots2024 [196]A Survey on Robotics with Foundation Models: Toward Embodied AI2024 [197]Advancements in Robotic Intelligence: The Role of Computer Vision2024 [198]Advances in Transformers for Robotic Applications: A Review2024 [199]Embodied AI with Large Language Models: A Survey and New HRI Framework2024 [36] Embodied Intelligence Toward Future Smart Manufacturing in the Era of AI Foundation Model 2024 [200]Foundation Models for Autonomous Robots in Unstructured Environments2024 [201] Integrating Reinforcement Learning with Foundation Models for Autonomous Robotics: Methods and Perspectives 2024 [202]Large Language Models for Robotics: Opportunities2024 [203] Leveraging Foundation Models in Robotics: Transforming Task Planning and Contextual Execution 2024 [204] Leveraging Large Language Models in Human-Robot Interaction: A Critical Analysis of Potential and Pitfalls 2024 [205]Real-World Robot Applications of Foundation Models: A Review2024 [206] The Future of Intelligent Healthcare: A Systematic Analysis and Discussion on the Integration and Impact of Robots Using Large Language Models for Healthcare 2024 [12]Toward General-Purpose Robots via Foundation Models: A Survey and Meta-Analysis2024 [207]Towards Generalist Robot Learning from Internet Video: A Survey2024 [208]Demonstrating Large Language Models on Robots2023 [9]Foundation Models in Robotics: Applications2023 [209]Large Language Models for Robotics: A Survey2023 [210]Robot Learning in the Era of Foundation Models: A Survey2023 73 Appendix B Literature Research 74 Table B2: All eleven ASQS Queries used for the RFM-related (main) corpus retrieval process as described in Section 3.2. Characters are escaped as used within the Python Script, each query is shown before being translated for the respective database. ["(robot* OR robotic) AND (general* OR generalist OR generalisation OR generalization OR foundation OR \"multi-task\" OR multitask OR \"multi task\" OR \"pre-trained\" OR pretrained OR \"pre trained\") AND (control OR manipulation OR action OR learning) AND (transformer OR model* OR policy OR agent* OR system) AND (\"vision-based\" OR \"vision based\" OR \"vision-language\" OR \"vision language\" OR \"language-guided\" OR \"language guided\" OR \"natural language\" OR \"natural-language\" OR \"multi-modal\" OR \"multi modal\" OR multimodal)", "(robot* OR robotic) AND \"foundation model*\"~3", "(robot* OR robotic) AND (\"vision-language-action\" OR \"vision language action\"~3) AND model*", "(robotic OR robot*) AND (\"foundation model\" OR \"general-purpose model\"~3 OR \"large-scale AI\"~3 OR \"pretrained AI\" OR \"generalist AI\") AND (\"multi-modal\" OR multimodal OR \"diverse datasets\" OR \"robot demonstration\" OR \"shared representation\" OR \"transfer learning\") AND (generalization OR versatility OR transferability OR adaptation OR \"broad capabilities\"~3) AND (perception OR \"decision-making\" OR actuation) NOT (\"specialized task\" OR \"specific kinematics\" OR \"highly-specialized\")", "(robot* OR robotic OR robotics) AND (\"foundation model\" OR \"general-purpose model\" OR \"large-scale model\" OR \"generalist AI\") AND (multimodal OR \"multi-modal\" OR \"diverse datasets\" OR \"robot demonstrations\" OR \"sensor data\" OR \"visual data\") AND (control OR perception OR actuation OR \"decision-making\" OR intelligence) AND (generalization OR transfer OR \"adaptable intelligence\" OR \"transferable capabilities\") NOT (specialized OR \"specific kinematics\" OR \"specific tasks\" OR \"highly-specialized\")", "(robot* OR robotic) AND (\"foundation model*\" OR \"pretrained model*\" OR versatile OR transferable OR generalist OR \"shared representation*\") AND (multimodal OR \"robot sensors\" OR \"sensor data\" OR (image OR text OR dataset*)) AND (perception OR intelligence OR actuation OR \"robot behavior\") AND (generalization OR adaptation OR versatility) NOT (specific OR specialized OR \"highly-specialized task\" OR kinematic*)", "(\"robot* foundation model\"~3 OR \"robot* generalist model\"~3 OR \"general-purpose robot* model\"~3) AND ((multimodal OR \"multi-modal\" OR \"multi modal\") AND (pretrained OR \"pre-trained\" OR \"pre trained\")) AND (generalization OR generalisation OR transfer OR adaptability OR versatility OR \"broad capabilities\") AND (perception OR \"decision-making\" OR actuation OR control) NOT (\"task-specialized\" OR \"specific robot kinematics\" OR \"highly-specialized model*\")", "(\"vision-language-action\" OR \"vision language action\"~3 OR VLA OR (vision AND language AND action)) AND ((model OR architecture OR system) AND (multimodal OR \"multi-modal\" OR \"multi modal\")) AND (robot* OR robotics) AND (manipulat* OR navigat* OR control OR action OR behavior) AND (versatile OR \"general-purpose\" OR general*) NOT (\"single-modality\" OR specialized OR narrow)", "(\"vision-language-action\"~3 OR \"vision NEAR/3 language NEAR/3 action\" OR VLA OR (multimodal AND (vision OR perception) AND language AND action)) AND (robot* OR robotics OR robotic) AND (model OR system OR architecture) AND (control OR behavior OR movement OR manipulation OR navigation) AND (\"general-purpose\" OR versatile OR integration) NOT (\"single-modality\" OR \"perception-only\" OR \"language-only\" OR narrow OR specialized)", "(multimodal AND (vision AND language AND action)) AND (robot* OR robotics) AND (\"control signals\" OR behaviors OR actions OR navigation OR manipulation) AND (versatile OR \"general-purpose\" OR integrated) NOT (narrow OR \"single-modality\" OR specialized)", "(\"vision-language-action\" OR \"vision language action\"~3 OR VLA) AND (robot* OR robotics OR \"robotic system\") AND (AI OR \"artificial intelligence\" OR \"machine learning\") AND (control OR behavior OR action OR manipulation OR navigation) AND (multimodal OR \"multi-modal\" OR integrated OR \"multi modal\") AND (model OR framework) NOT (\"perception only\" OR \"single modality\" OR \"language modeling alone\" OR \"narrow task\")"] 75 Table B3: All seven ASQS Queries used for the industrial implication corpus retrieval process as described in Section 3.3. Characters are escaped as used within the Python Script, each query is shown before being translated for the respective database. ["(robot OR robotics) AND (industrial OR \"industrial application\"~3 OR \"industrial use-case\"~3 OR \"applied robotics\"~3 OR \"industrial environment\"~2) AND (review OR analysis OR viewpoint OR guideline OR perspective) AND (hardware OR sensor OR \"compute resource\"~2 OR integration OR compliance OR standards OR workflow)", "(robotics system* OR \"industrial robotics\" OR \"applied robot*\"~2) AND (design OR performance OR deployment OR scalability OR robustness OR reliability OR safety OR operator interaction) AND (industrial needs OR \"industrial requirement*\"~2 OR \"industrial-focused\" OR \"real-world application\"~3)", "(robotics OR robot* OR \"robotic systems\") AND (industrial OR \"industry-specific\" OR \"industrial automation\") AND (design OR deployment OR \"performance requirement*\"~2) AND (\"machine learning\" OR ML OR \"artificial intelligence\" OR AI) AND (analyze OR review OR implication* OR considerations)", "(\"industrial robots\" OR robotics OR \"robot hardware\" OR \"robotic systems\") AND (sensor* OR \"compute resources\" OR integration OR standards OR compliance OR regulatory) AND (manufacturing OR logistics OR assembly OR inspection OR \"quality control\") AND (property OR characteristic OR \"industrial requirements\")", "(\"industrial deployment\" OR \"industrial automation\" OR \"real-world deployment\") AND (\"robotics\" OR \"robot\") AND (\"challenges\" OR scalability OR robustness OR reliability OR safety OR \"operator interaction\") AND (\"machine learning\" OR AI)", "(robotics OR AI) AND (review OR \"case study\" OR analysis) AND (industrial OR \"manufacturing environments\" OR logistics OR \"assembly line\" OR warehouse OR inspection OR \"quality assurance\") AND (constraints OR \"sector-specific needs\" OR adaptability)", "(industrial robotics OR \"robotic system\") AND (hardware OR \"compute resources\" OR sensor OR \"sensor types\" OR integration OR \"regulatory compliance\" OR standards OR \"performance requirements\") AND (manufacturing OR logistics OR factory OR automotive OR electronics OR warehousing) NOT (academic OR prototype)"] 76 Appendix C Industrial Implications 77 Table C4: Full criteria catalogue for industrial applicability rating as discussed in Section 7.3. Derived from the general industrial implications (see Section 7.1) ImplicationCategory/ Criterion I1:A. Multi-Hardware and Configuration Adaptability Adaptability & Flexibility 1. Multi-Hardware Support: Operates with different kinematics, sensors, and end-effectors without retraining. 2. Hot-Swappable & Dimensionality Adaptation: Sensors/actuators can be replaced, added, or removed during operation. Automatically detects and configures input/output channels at runtime according to connected hardware and environmental context: Flexibly adapts the dimensionality, type and interpretation without manual intervention or system restart. B. Task Adaptability and Strategy Flexibility 3. Multi-Task Generalisation: Handles multiple task types/categories, including unseen variations thereof, without retraining. Adapts strategies dynamically for task variations and can switch between tasks online. 4. Restricted Autonomous Task Switching: Only switches between distinct tasks upon explicit user instruction: autonomy remains bounded, so adherence to the operator’s goal stays primary (may autonomously adapt solution approaches/ strategies). 5. Strategy-Goal Alignment: Any adaptation of solution strategies ensures adherence to the commanded task goal: deviating too far from user instructions triggers feedback, halt, or request for approval. C. Environmental Adaptation and Robustness 6. Generalisation to Environmental Variation: Maintains consistent performance across diverse environmental conditions (lighting, weather, background, friction, etc.), object/material variations, and scene clutter, with minimal degradation. 7. Autonomous Handling of Distractors/Obstacles: Identifies and robustly manages/ handles unknown, distracting, or non-required objects in the scene or obstacles along the path/ trajectory (including incorporation into task strategy if needed). 8. Online Environment Adaptation & Response: Recognises and continuously adapts to environmental changes, such as errors, faults, sensor degradation or external influences during operation: e.g. logging, halting, or feedback-requesting as confidence drops. 9. Autonomous Fallbacks: Capable of autonomously choosing context-appropriate error/fault responses (e.g. request feedback, switch input modalities, logging, etc.) and behaviour (e.g. continue operation, transition to safe state, adapt strategy, etc.) without requiring predefined fallback behaviour. D. Instruction and Feedback Flexibility 10. Multimodal Command Interface: Accepts user-input (task instructions, correction/feedback, etc.) via diverse modalities (natural language in multiple languages, images, teaching by demonstration, video, etc.). 11. Mid-Execution Feedback Responsiveness: Remains receptive to user feedback and correction throughout execution, enabling course correction and adjustment without re-initiating trajectories. E. Self-Flexibility and Adaptive Behaviour 78 ImplicationCategory/ Criterion 12. Operational Adaptation Confidence: Continuously outputs a confidence score reflecting its present predicted ability to execute the instructed task using the currently (adapted) solution strategy, in the current environment, with the currently active hardware configuration and sensor context. The score may be used to trigger automatic logging, operator feedback requests, or safe halting if confidence drops below a set threshold, ensuring transparent awareness and operational safety during dynamic adaptation. 13. Autonomous Incorporation of Skills: Dynamically selects and integrates known robot skills (e.g. action primitives) as needed for current task/environment into the current solution strategy (like tool-calling). 14. Online/Mid-Execution Adaptation: Integrates operator/environmental feedback, (input/ task)-quality degradation and detected faults dynamically during operation, adapting strategy, input importance or output. 15. Explained Adaptation Decisions: Explains, upon request, its performed/ intended adaptation of strategies, handling of distractors, input importance/quality (for each modality), and response to uncertainties or noise. 16. Memory of Previous Faults/Distractors: Remembers prior encountered variations, distractors, faults, anomalies, environmental changes and corresponding solutions, reducing the need for repeated user intervention for recurring scenarios. 17. Energy/System-Aware Adaptation: Adapts solution strategies and predicted actions based on real-time resource constraints and consumption (e.g. battery charge, system health, energy/ compute utilisation). 18. Self-Evaluation & Command Adherence: Continuously self-evaluates its own adapted behaviour and adherence to user command(s): Flags or readjusts if deviation from expected outcome or operator intent is detected. I2:A. Verification, Standards, and Certification Safety & Compliance 1. Standards-Compliant: Behaviour and safety features are checked to align with relevant industrial standards, evaluated by reproducible tests. 2. Integrable Safety Systems: System interfaces and safety signals are compatible with industrial plant safety infrastructure, supporting seamless integration and monitoring (e.g. model-generated safety stop can be directly used to stop respective production line). 3. Safety Documentation & Traceability: All safety features, integration, and reasoning are well documented, including considered safety rules, internal processes, output explanations, and verified compliance with standards. B. Safety Reasoning, Adaptation, and Confidence 4. Safety Confidence Output: Outputs a confidence score estimating the safety margin of each predicted operation and for the complete task before starting operation (in the current environment). The confidence score quantifies the system’s estimated likelihood that all specified safety requirements are currently satisfied, using sensor/hardware/context data. 5. Persistence of Safety Capabilities in Adaptation: Safety rules/modules remain enforced, and system safety is verified after any model adaptation, retraining, or finetuning→ standalone safety logic cannot be bypassed by learning changes. 6. Reactive, Real-Time Safety: System performs continuous, real-time evaluation of safety (considering sensor noise, hardware degradation, dynamic obstacles, harmful objects, OOD, and environment changes) – e.g. leveraging confidence – with immediate adaptation (halting/ triggering stop, slowing, or replanning) if risks arise. 79 ImplicationCategory/ Criterion 7. Proactive, Plan-Level Safety: Safety reasoning is incorporated into high-level planning/solution strategies (not only reactive), evaluating alternative actions and solutions by their associated predicted safety confidence. C. Safety Rules, Enforcement, and Supervisory Control 8. Rule-Adherence & Autonomous Rejection: The system always respects predefined safety/behavioural rules and can autonomously reject operator goals/commands that pose unacceptable risk (e.g. “grab knife and throw”), e.g. supervised by independently certified controllers/modules. 9. Command Compliance with Online-Safety: System adheres strictly to operator commands as long as safety requirements are not compromised: Command execution is continuously monitored and suspended/modified if unacceptable risk is detected, providing transparent feedback (to the user/ logs/ supervisor/ etc.). 10. Redundant, Independent Safety Mechanisms: Employs multiple, independent (possibly redundant and diverse-modality) mechanisms for safety assessment and enforcement (e.g. dual model prediction, dedicated supervisor, independent safety layers). 11. Decoupled, Real-Time Safety Control: Safety checks and action execution operate independently: a ”safety semaphore” or gating mechanism ensures actions are only predicted/executed when confidence exceeds threshold. D. Operator/User and External Interfaces 12. Authorised Operator Detection & Override: Only authorised/ recognised operators may control safety overrides or cause safety-critical actions; interfaces securely verify operator identity and permissions. 13. External Safety Integration: Safety outputs/signals are shareable with other robots/systems in plant (including alarms, safe state signals, zone definitions, and status reports). E. Fault Detection and Recovery 14. Fault Detection & Sensing: Actively detects major faults or safety violations. Distinguishes between sensor fault and safety/ environmental hazard and the degree of severity. 15. Safe, Autonomous Recovery: Recovers autonomously to a safe state after detecting major faults, safety loss, low safety-related confidence in own actions or accidents. 16. Learning from Safety Events: Continuously adapts and improves safety confidence/logic or recovery behaviour based on feedback from operator, supervisor modules, comparison of actual vs. predicted safety outcomes and past incidents. F. Transparency and Logging 17. Transparent Safety Reasoning: The system can explain its safety confidence, risk assessment, and chosen/blocked actions to operators, both live and retrospectively (with logs/reports or upon querying). 18. Persistent Safety Logs: All safety-relevant events, decisions, adaptations, actions, and outcomes are logged, both user-readable and for self-improvement. I3:A. Communication Modalities & Operator Interfaces HRI & HRC 1. Multimodal Instruction & Output: Accepts and generates instructions, queries, and feedback through multiple operator interfaces (speech, text, gesture, visual). Chooses the most suitable modalities based on available hardware, environmental context (e.g. noise level, lighting), operator preferences, and task demands/ instruction. 80 ImplicationCategory/ Criterion 2. Instruction Confidence: For every operator command or query, provides a confidence score reflecting certainty in both (a) correct recognition of the input (e.g. speech recognition, gesture identification), and (b) correct interpretation of the operator’s intent. When confidence is low, proactively seeks clarification or reframing from the operator, and communicates the reason for uncertainty (ambiguity, low signal, novel command, etc.). 3. Intuitive Feedback Acceptance: Integrates operator clarifications, corrections, or input refinements (e.g. specifying objects, adjusting approach, clarifying intent) online, into the ongoing action interpretation or prediction – adapting its immediate responses and future understanding, mostly without pausing or disrupting the main task flow. B. Responsiveness & Interaction Control 4. Real-time Interruptibility: Allows operators/ users to instantly and directly pause, modify, query, or halt ongoing execution or task, providing full control and safe intervention across all interfaces and modalities. 5. Continuous Responsiveness: Maintains complete HRI/HRC accessibility and interaction throughout all operational phases (idle, executing, transitioning, recovering), ensuring operators can always intervene, instruct, or query – never leaving the system ”locked out” or unresponsive. C. Operator Management, Authorisation & Privacy 6. Active Operator Exclusivity: Guarantees that only one authorised/ known operator/ user is ”active” and in control at any time. Enforces explicit, secure handover or switching protocols, and reliably ignores unverified or distracting bystanders in both input and output. 7. Operator Identification & Authorisation: Identifies and responds exclusively to securely recognised and authorised personnel/ user. Actively prevents unauthorised commands, even under disruptive or noisy conditions, and logs all identity-related access attempts. 8. Operator Data Privacy: Encrypts and protects operator-specific data and interaction history; permits authorised deletion or review in accordance with relevant privacy standards. Ensures only authorised personnel can access, modify, or request deletion of such data. 9. Learned Operator Preferences: Adapts interface layout, communication style, behaviour, or task sequencing to individual operators/ users – learning and persisting preferences across sessions for efficiency and comfort unless instructed otherwise. D. Human-Robot Collaboration & Behaviour Adaptation 10. Human Presence Awareness in Workspace: Continuously senses human presence and proximity throughout and beyond the workspace using multiple sensor modalities (vision, audio, proximity, etc.). Dynamically adapts operation speed, safety margins, and communication in response to detected changes (especially in cases of unexpected entries, operator absence, etc.). 11. Collaborative Adaptation: Reactively modulates solution strategies, action plans, and interaction traits in direct response to real-time operator feedback, demonstrations, or correction. Behaves as an adaptive teammate, tailoring technical and social behaviours to maximise collaboration efficiency. 12. Human-Centric Adaptation: Dynamically and autonomously adjusts behaviour, speed, and language/output based on human presence, operator skill/preference, perceived operator condition, task safety factors, and perceived risk. E. Reasoning, Standards & Exception Handling 81 ImplicationCategory/ Criterion 13. Transparent Reasoning Explanation: Provides operator-friendly, on-demand explanations during operation of its own reasoning, uncertainty, behaviour, and decision-making – using appropriate modalities (natural language, images, charts) and tailoring explanations to the operator’s degree of expertise. 14. Industrial HRI/HRC Standards Compliance: Adheres to relevant industrial safety, communication, and collaboration standards and best-practices for HRI and HRC; supports integration with company-specific norms, exception rules, and documentation requirements. 15. Exception Handling & Safe State Transition: Reliably recognises faults, abnormal human behaviours, and communication breakdowns during HRI/HRC. Notifies the operator, explains the situation, and if necessary, transitions to a pre-defined safe state. I4:A. Fault Detection, Diagnosis & Monitoring Robustness & Reliability 1. Multi-Level Fault & Anomaly Detection: Continuously detects and logs faults, anomalies, and degradations not only in system hardware and sensors, but also in environment, manipulated objects, operator actions, and internal representations – utilising diverse sensor inputs, automatic evaluation, supervisor components or confidence scores to spot issues before failures occur. 2. Reproducible & Benchmarked Fault Diagnosis: Integrated fault/anomaly detection delivers consistent, reproducible evaluations under matching conditions, with detection precision benchmarked against standardised scenarios. 3. Threat & Attack Detection: Autonomously monitors all inputs and interface signals for security threats, adversarial manipulations, and unauthorised commands, triggering alerts or adapting strategies or behaviour to mitigate risk and reduce attack impact. B. Error Handling, Recovery & Adaptation 4. Automated Recovery & Safe-State Transition: Autonomously initiates recovery routines or shifts into a safe state when critical faults, severe anomalies, unrecoverable disturbances, or low confidences are detected (prioritisation of safety and system integrity over performance and task completion). 5. Graceful Degradation & Strategy Adaptation: Adjusts execution plans, sensor usage, and action strategies in real-time when encountering partial sensor failures, orientation/config changes, or slightly degraded environmental conditions, striving to maintain performance without outright task failure. 6. Autonomous Fault Categorisation & Severity Response: Recognises, categorises, and prioritises faults by severity, autonomously selecting and executing the most appropriate recovery behaviour, fallback routine, or adaptation strategy. 7. Robust Redundancy & Modality Management: Ensures reliable operation and fault tolerance by accepting, integrating, and dynamically switching between redundant input and output modalities. When any channel degrades or fails, the system autonomously maintains task progress through alternative modalities, adapting detection, execution, and communication strategies to suit the current hardware, environment, and operator constraints. C. Robustness to Disturbance, Environment & Variation 8. Environmental Robustness & Adaptation: Operates effectively across varying environmental conditions (e.g. lighting, noise, object/material variation), detecting and alerting for changes or disturbances that could risk execution reliability. 82 ImplicationCategory/ Criterion 9. Disturbance Resilience & Continuous Operation: Maintains task progress and adherence to commands even when exposed to distractors or minor disturbances, recovering or adapting autonomously when possible. (e.g. arm pushed aside, etc.) D. Learning, Feedback & Continuous Improvement 10. Adaptive Error Learning & Memorisation: Continuously learns from past faults, errors, environmental changes, and operator feedback –memorising causes and resolutions to improve future fault/ error detection, prediction, and recovery. 11. Online Confidence & Reliability Assessment: Continuously generates a clear reliability confidence score for further operation, quantifying the risk and likelihood of fault occurrence and successful execution. E. Supervision, Oversight & Trust 12. Supervisor Oversight & Fault Review: Employs dedicated supervisor components to monitor for faults in execution, hardware/sensor input, and operator commands, critically evaluating instructions and requesting operator engagement for suspected errors. 13. Logging, Explanation & Operator Guidance: Logs and communicates faults, anomalies, and noise through varied modalities, providing operators with actionable explanations (what/why/when/how to resolve) and performance insights for transparency and trust. 14. Consistent & Documented Fault Handling: Ensures both fault/anomaly-detection and response-behaviours remain reproducible and consistent across time and conditions, with all capabilities and accepted modalities documented and evaluated. I5:A. Specification Compliance Precision & Accuracy 1. v/P &A Instructability: For every task specification, the system can be instructed to meet an explicit v/P&A target. 2. Outcome Logging & Failure Reasoning: Gathers/ creates and logs all available v/P&A-relevant task execution metrics (such as velocity and, if sensor capabilities permit, achieved precision & accuracy) after each run. If direct measurement is not possible, the system leverages predicted values, confidence scores, or proxy indicators. For any deviation or failure to meet specified standards, the system documents and explains likely reasons based on the given information. 3. User-Defined Trade-offs: Supports explicit operator input for v/P&A-trade-offs (e.g. maximising speed at the cost of precision, or vice versa if either cannot be ensured), before or during execution and ensures system behaviour adapts accordingly. B. Evaluation, Reporting & Confidence 4. v/P &A-Confidence Score & Explanation: For every execution, outputs a confidence score estimating the predicted likelihood that required v/P&A will be achieved, factoring in task, object/material, environment, and other sensor data. When low confidence is predicted, system can explain likely causes (e.g. out-of-distribution scenario, sensor noise, disturbances). 5. Achievable v/P &A-Proposals: Alongside the confidence score and explanation, the system can propose and explain alternative v/P&A-combinations likely to succeed in the given context (task, object, environment), helping users adjust requirements if needed. 6. Transparent Documentation of v/P &A Capabilities: Achievable v/P&A-combinations for every tested task, object, and context is provided alongside the model (including limitations, known influence factors, etc.). C. Consistency & Disturbance Handling 83 ImplicationCategory/ Criterion 7. Consistency Over Time & Contexts: Demonstrates consistent v/P&A across known tasks, objects, and environments, and under repeated test conditions. 8. Disturbance & Uncertainty Management: Identifies and factors in disturbances, distractors, sensor anomalies, or environmental changes when predicting, achieving, and reporting v/P&A metrics. Explains in real-time what current conditions may affect v/P&A and adapts behaviour accordingly. 9. Autonomous v/P &A Adaptation: Automatically adapts its own v/P&A-settings/ targets and solution strategies in response to changing conditions or goals (e.g. sensor quality, energy saving) within the user-defined bounds (e.g. reduces speed within user-defined limits to save energy). D. Learning, Feedback & Continuous Improvement 10. v/P &A Learning from Experience and Feedback: Memorises and utilises past v/P&A-failures (e.g. slower than/ not as precise as demanded) and successful adaptations of v/P&A to improve prediction confidence, action planning, and further relevant decisions for similar future tasks. 11. HRI/HRC Integration: Accepts v/P&A-related user/operator feedback during or after execution to update confidence or task strategy and leverages help from operators or other robots to meet v/P&A requirements. E. External/ Separate Component Integration 12. Skill Integration with Known v/P &A: Incorporates predefined and learned skills/tools for improved v/P&A outcomes, selecting solutions with documented effectiveness whenever possible. 13. Supervision & Verification Module: Includes separated modules or capabilities to explicitly verify actual outcome v/P&A during and post-execution, not just prediction. Alerts/warns model/operator if outcome may not meet requirements, explaining why and offering alternatives. I6:A. Real-Time Robustness & Generality Real-time Performance 1. Task, Input & Environment Independence: Meets real-time requirements (e.g. inference, control, responsiveness) regardless of task complexity, environment, objects, input modality, number or dimensionality of in/outputs, or presence of distractors, OOD conditions, or hardware degradation – demonstrated through benchmarked worst-case test scenarios on a given hardware-setup. 2. System-wide Real-time Compliance: Real-time operation is maintained holistically across all relevant system components – perception, reasoning, actuation, sensing, communication, feedback, and logging – rather than only for a subset. B. Adaptive Resource & Real-Time Management 3. Dynamic Resource Management & Determinism: Autonomously adapts compute usage (e.g. through output/inference frequency modulation) to maintain user- or norm-specified real-time targets. When under load or during resource shortages, it disables or degrades non-critical functions to uphold core real-time properties and safety measures. Benchmarked under varying hardware constraints and external disturbances. 4. Autonomous Adaptation & Safety Response: Upon detecting real-time violations or imminent deadline risks, triggers fallback routines, prioritises safety-critical functionality, and can alert operators or other systems as needed. Adaptation and prioritisation logic is explainable and logged. 84 ImplicationCategory/ Criterion 5. Goal-Responsive Solution Strategies: Comprehends and adapts to new or updated user goals and instructions ”in real time” based on current workload – ensuring planning and execution adaptation meets pre-defined latency requirements. C. Monitoring, Reporting, & Continuous Improvement 6. Live Metrics, Logging & Explainability: Continuously monitors, reports, and logs real-time metrics (e.g. frequencies, latencies, dropped cycles, deadline adherence), makes these accessible, and provides clear explanations for performance drops, adaptation events, or system changes (e.g. what, why, when a module was throttled/disabled). 7. Real-Time Confidence/Alerting: Continuously outputs a confidence score predicting current and future compliance with real-time constraints (factoring in compute status, history, planned actions, hardware utilisation). Issues alerts or recommendations if violation risk increases. 8. Historical Data Utilisation: Integrates historical real-time performance/logs and adaptation outcomes to optimise future resource use, prediction, confidence and adaptation strategies, coupled with continued evaluation. 9. Transparent Documentation & Compliance: Maintains up-to-date, accessible documentation on real-time system capabilities, adaptation mechanisms, hardware requirements (e.g. which hardware-specifications are required for which set of enabled functionalities) with evaluated compliance. All possible real-time related adaptations, violation cases and countermeasures are documented. I7:A. Hardware & Operational Cost-Efficiency Cost- effectiveness & Integration Capabilities 1. Low-Cost Hardware Readiness: Delivers reliable robot control and core model capabilities on commercially available, affordable compute hardware (e.g. edge devices, embedded platforms), supporting lower-spec sensors and actuators. 2. Dynamic Resource Allocation: Dynamically adjusts its resource usage – such as modulating inference frequency, simplifying/ deactivating optional model operations, adapting velocity, etc. – minimising energy use or execution duration given current environmental factors (e.g. current required throughput) to reduce operating/ production-line costs. B. Universal & Modular Integration 3. Universal Integration without Retraining: Seamlessly integrates with diverse robot embodiments, kinematics, sensor and actuator types, and plant automation systems, without requiring model retraining, software changes, or adaptation for each hardware configuration change. Expected performance is preserved wherever hardware capacities allow. 4. Hot-Swapping & Auto-Detection: Supports hot-plugging and automatic detection of new, swapped, or removed input/output channels (sensors, actuators, interfaces, communication), updating its operation and configuration with no need for manual operator intervention. C. Commissioning, Flexibility & Self-Adaptation 5. Guided & Autonomous Commissioning: Features intuitive commissioning workflows, including automated calibration and adaptation procedures, providing feedback on missing/ expected/ proposed in-/outputs or adjustments, enabling non-expert users to deploy and onboard the model with minimal manual setup. 85 ImplicationCategory/ Criterion 6. Flexible Input/Output Adaptation: Automatically configures itself to accept changing sensor modalities, command interfaces, and action outputs, functioning consistently even if components are substituted, or temporarily degraded/ missing – requiring no manual software changes. 7. Autonomous Fault Handling & Feedback: Autonomously detects and diagnoses sensor, actuator or integration faults and missing interfaces – provides clear, actionable feedback or explanations to users for resolution and continued safe operation. D. Documentation & Integration Performance 8. Comprehensive Integration Documentation: Includes detailed, accessible documentation for all supported platforms, integration methods, autonomous adaptation and recognition mechanisms, commissioning and troubleshooting steps, supplemented with example configurations and practical best practices. 9. Integration Performance Monitoring & Learning: Continually logs integration performance, adaptation strategies/ speed, and incident recoveries – using historical deployment data to inform, accelerate, and improve future onboarding and commissioning processes. 10. Cost-Reduction Memory & Adaptive Optimisation: Systematically records operational experiences that relate to resource and cost efficiency – such as energy- conserving behaviours, internal module optimisations, and hardware usage patterns – and leverages this memory to autonomously propose or implement more cost-effective action plans, configurations, and adaptation strategies in future deployments. I8:A. Capability, Limitation & Reasoning Transparency Explainability & Trust 1. Task Feasibility & Capability Explanation: Supports clear, user-facing explanations of what tasks it can or cannot solve – including limitations, current capabilities, and the feasibility of the currently instructed task – both before and during execution, with accessible rationale for any constraint. 2. Reason Logging & Accessible Justification: Autonomously generates, logs, and exposes the reasoning behind its critical decisions, action choices, adaptations, and detected failures, ensuring each is transparently documented and retrievable for audit, debugging, self-improvement and user review. 3. Explains Influences & Disturbances: Provides explanations regarding environmental factors (e.g. lighting, scene setup, object/material properties, distractors, detected disturbances) and how these influence its reasoning, behaviour, or performance, along with advice or solutions for mitigation. B. Self-Assessment, Learning, & Feedback Integration 4. Continuous Self-Assessment with Explainability: Consistently self-evaluates capabilities, performance, and limitations and their impacts in the current context (e.g. hardware, sensor setup, noise conditions, distribution shifts), providing the evaluation alongside an explanation and leveraging these insights for ongoing online learning, self-improvement and user-insight. 5. User Feedback & Adaptive Improvement: Supports intuitive interfaces for user feedback and corrections during and after execution (e.g. thumbs up/down, follow-up queries, correction inputs), acknowledges understanding, adapts behaviour accordingly, and provides explanations of resulting changes; uses feedback to refine future reasoning. 6. Monitors & Logs Explainability for Learning: Keeps persistent records of all generated explainability outputs (reasoning, assessment, error reports), using these logs to continuously improve future behaviours, onboarding, and commissioning processes. 86 ImplicationCategory/ Criterion C. Explainability Methods, Interfaces & Outputs 7. Flexible, Intuitive Explainability Interfaces: Offers explanation outputs in multiple modalities (text, speech, visualisation, etc.), supporting both non-expert operators and technical users, with methods to visualise/explain impact of user commands, input/output importance, intermediate steps, and internal processes at runtime. 8. Confidence Score Explanation: All confidence scores (for actions, predictions, or decisions) are supplied with user-facing explanations – detailing what influenced high or low values, such as data loss, fallback mode, sensor unreliability, or ambiguous scene attributes. 9. Explainability Across System Components: Integrates explainability tools and interfaces across all modules, enabling explainable actions, outputs, and internal reasoning during setup, operation, debugging, and benchmarking; all methods are thoroughly documented with guidance for triggering and interpreting results. D. Security, Compliance, & Benchmarking Explainability 10. Security & Authorisation Feedback: Detects and explains potential security issues, adversarial attacks, or authorisation events – providing specifics on why the event was flagged, input used, type of threat, and how to strengthen defences or permissions for future operations. 11. Rule & Norm Compliance with Explainability: Evaluates and explains compliance with operational guidelines, safety norms, and domain-specific rules (both upon being instructed and during/after execution), visualising/documenting adherence, deviations, and model reasoning behind its decision. 12. Performance Benchmarking & OOD Feedback: Integration of explainability methods in benchmarking, evaluation, and reporting – e.g. visualising focus, highlighting difficulties in OOD situations, or explaining how improvements, data changes, or model modifications affect learned skills or generalisation, etc. E. Trust & Transparency during Runtime 13. Behaviour Consistency & Pre-Adaptation Notification: Maintains consistent, transparent explanations for identical inputs and conditions, avoids “super-hero syndrome” (thinking to be capable to solve every task), and informs the user before adapting behaviour due to error, environment change, or new context. 14. Error, Anomaly & Safeguard Explanations: Immediately triggers explanation processes for detected errors, anomalies, or failed safeguards – detailing what happened, why, and the recovery strategies attempted or proposed. Explanation is automatically provided to relevant system components/ logs/ etc. I9:A. Multimodal Acceptance & Fusion Sensor Fusion & Perception 1. Multimodal Sensor Acceptance & Fusion: Accepts diverse sensors and supports their real-time fusion and processing across modalities (visual, tactile, audio, etc.), including redundant interchangeable sensors, with the ability to interpret and combine inputs from multiple simultaneous sources for robust perception. 2. Flexible Input Combination & Goal Specification: Accepts task instructions, goals, and feedback through multiple modalities (speech/text, gesture, image, etc.), and can leverage several modalities simultaneously for improved perception, understanding, and command robustness. 3. Security in Sensor Operations: Utilises various sensor sources/ modalities to detect adversarial attacks, facilitates secure authorisation, and ensures safe detection of humans/workers. 87 ImplicationCategory/ Criterion B. Input Monitoring, Fault Tolerance & Adaptation 4. Continuous Input Monitoring & Feedback: Continuously checks sensor inputs for noise, faults, degradation, absence, or sub-optimal orientation, providing real-time feedback during integration, commissioning and operation, reporting missing or degraded inputs as well as recommendations for improvement. 5. Robust Fault Tolerance & Degradation Handling: Detects and actively compensates for sensor failures, degradation, or external disturbances by dynamically shifting input weighting, switching perception strategies, or prioritising the most reliable sensor(s) for the current scene, environment, and task context. 6. Autonomous Disturbance Recognition & Countermeasures: Recognises environmental disturbances that critically affect perception (e.g. poor lighting, moisture, obstructed view) and initiates context-sensitive countermeasures (like halting, alerting users, removing obstacles, or activating cleaning routines), and adapts perception strategies online. C. Dynamic Input Prioritisation & Adaptation 7. Dynamic Sensor Prioritisation: Automatically ranks and prioritises active sensors and modalities based on current instruction, scene conditions, quality, relevance, historical context and reliability, adapting input strategy in real-time to environmental changes and operational demands. 8. Online Learning for Perception Skills: Continuously learns and adapts perception strategy (e.g. how to interpret or counteract noisy or degraded sensor values, shifting focus to non-degraded sensors, . . . ) using historical data and reasoning explanations to improve future sensor fusion and perception capabilities. D. Scene Understanding, Interpretation & Output 9. Comprehensive Scene Understanding via Fusion: Recognises objects, materials, spatial arrangements and environmental constraints by integrating information from multiple, complementary modalities – supplies rich scene interpretation, spatial awareness, and accurate environmental representations for downstream robot control. 10. Perceptual Confidence & Explanation Output: Provides a perceptual confidence score, summarising the certainty of scene/object perception and input quality. Complemented by interpreting input importance (which sensors/modalities influenced results) and supply of user-friendly explanations visualising/ describing reasoning behind understanding, degraded input handling, and current limitations. 11. Perceptual Feature Exposure & Modularity: Exposes interpretable intermediate representations and perceptual features to system components or external interfaces, supporting modularity, human supervision, and transparent debugging of perception skills. E. Evaluation, Documentation & Knowledge Exchange 12. Modular Integration & Documentation: Enables flexible, exchangeable or adaptive observation spaces (autonomous/manual adaptation; fine-tuning optional) with support for hot-swapping sensor setups, providing clear feedback during integration and maintaining up-to-date documentation of all supported sensor types, modality usage, and perception features. 13. Perception Evaluation & Benchmarking: Benchmarks and documents perception skills (input importance ranking, noise adaptation, consistency, and reproducibility), and supplies results for model evaluation, external audit, and improvement. 14. Peer-to-Peer Perception Knowledge Exchange: Can exchange sensor/perception information and interpretations with other systems, enabling collective/global perception enhancement and collaborative scene/environmental understanding. 88 ImplicationCategory/ Criterion I10:A. Coverage of Tasks, Domains, & Environments Standardised Benchmarking & Evaluation 1. Benchmark Diversity & Scope: Evaluated using standardised benchmarks that span a wide range of tasks, domains, environments, object types, and material properties – including known, partially known, and unknown scenarios – explicitly testing performance in worst-case distribution shift or OOD situations. 2. Hardware Platform Benchmarking: Evaluates model on multiple hardware platforms/ uses a hardware-aware benchmark, documenting resource utilisation, energy consumption, real-time performance, and cost-effectiveness across devices and sensor/actuator setups. B. Robustness, Generalisation & Transparency 3. Repeatability & Reproducibility: Demonstrates repeatable, reproducible model/system behaviour in benchmark runs, validating consistency over multiple trials, hardware settings, and environmental variations. 4. Robustness & Fault Tolerance Evaluation: Systematically benchmarks performance against faults, disturbances, distractor objects, sensor degradation, and intentional adversarial attacks; measures recovery strategies, fallback/safeguard execution, and online adaptation when exposed to (un)seen disturbances. 5. Generalisation & Adaptation Assessment: Evaluates the model’s generalisation and online adaptation capabilities – including response to distribution shifts, novel commands, unknown inputs, and task or environmental changes – using both synthetic and real-world scenarios. 6. Transparent Metrics & Reporting: Publishes clear, traceable benchmark metrics – including performance, accuracy (v/P&A), safety, robustness, fault tolerance, energy usage, real-time performance and compliance – with full documentation for all test scenarios. C. Instruction, Output & Integration Assessment 7. Instruction, Command & Modality Coverage: Benchmarks instructability and goal specification for all supported input modalities (e.g. speech, text, images, gestures), focusing on interpretation accuracy, and solution strategy adherence. 8. Integration & Commissioning Evaluation: Assesses ease and effectiveness of integration into different systems and workflows, including input/output adaptation, fine-tuning procedures, and usability before, during, and after commissioning. D. Specific Benchmarking 9. Module & Skill-Specific Benchmarking: Separate evaluation of (a) independent core modules, learned skills, tool selection quality, etc. and (b) whole-system-level performance/ metrics. 10. Safety, Security & Norm Compliance Validation: Explicitly evaluates implemented authorisation, safety features, resilience against adversarial inputs, and system compliance with relevant industrial standards, rules, and domain-specific guidelines. 11. Explainability, Reasoning & Feedback Benchmarks: Evaluation of built-in explainability methods, overall reasoning transparency, feedback usage/ usability, and correction/ user guidance understanding – demonstrating black-box mitigation and feedback consistency. 12. Data Requirement & Autonomous Runtime Assessment: Verifies the model’s data requirements for pretraining, fine-tuning, and online adaptation. Benchmarks autonomous self-evaluation during runtime learning, adaptation, and self-improvement. 89 ImplicationCategory/ Criterion 13. Human Interaction & Object Recognition Evaluation: Evaluates human-robot interaction and cooperation capabilities. Benchmarks HRC/HRI-related object/human recognition for known and unknown cases, considering detection rates, explanation quality, and adaptation of resulting behaviour. I11:A. Data Efficiency & Finetuning Data Requirements & Usage 1. Low-Data Adaptation & Few-Shot Requirements: Needs only few or zero new examples for effective adaptation to novel tasks, environments, hardware, or operators, using suitable techniques (e.g. in-context learning, online self-adaptation, modular fine-tuning) with minimal resource utilisation. 2. Efficient Task, Environment & Modality Adaptation: Supports adaptation of observation/action spaces or input modalities without needing retraining or extensive new data and supports independent module adaptation to minimise both overall and domain-specific data requirements. B. Pretraining Coverage & Robustness 3. Comprehensive & Diverse Pretraining Dataset: Pretraining covers broad task types, environments, domains, object/material variations, instruction modalities, distractors, noise, faults, adversarial attacks, redundant/fallback routines, and safety/data privacy scenarios, using both synthetic and real-world data for robust generalisation and OOD handling. 4. Documented Data Composition & Simulation-Real Transfer: Maintains accessible documentation on training/finetuning data composition, variation, and procedures. Including evaluation of transfer between simulation and real-world settings and tracks performance on transfer degradation. C. Data Sufficiency, Self-Assessment & Explainability 5. Automatic Data Sufficiency Assessment & Feedback: Continuously analyses historic and incoming data for sufficiency, coverage gaps, redundancy, and quality issues during adaptation or operation. Provides alerts and actionable recommendations for missing, underrepresented, or low-quality data (e.g. based on task, scene understanding, . . . ). 6. Explainability on Data Usage & Improvement: Supplies explainable analyses of data composition/ identified gaps, adaptation impact, improvement, and focus shift during training/finetuning. Recommends targeted data enrichment or refinement to address detected blind spots or weaknesses. D. Data Flexibility, Interface & Compliance 7. Flexible Data Type Acceptance: Accepts and learns (on- and offline) from diverse data formats and modalities (e.g. trajectory, demonstration, egocentric, multi-modal sensory). 8. Compliance with Privacy, Ethics & Law: Ensures all data usage can comply with provided legal, ethical, and corporate policies – leveraging privacy safeguards (e.g. data anonymisation, operator-specific data isolation, encrypted logging, . . . ) where needed. E. Ongoing Monitoring, Confidence & Robustness 9. Ongoing Adaptation/Finetuning Confidence Scores: Outputs initial and ongoing confidence scores after (fine)tuning and during online adaptation, estimating both data sufficiency/appropriateness and the model’s expected operational robustness and readiness under changing environments, hardware, or tasks. 10. Resilience Through Redundancy & Fault Tolerance: Incorporates redundancy, noise, faults, and fallback examples in pretraining and fine-tuning to support robust adaptation, autonomous countermeasure activation, and continued reliable operation under adverse data scenarios. 90 ImplicationCategory/ Criterion 11. Self-Assessment to Reduce Input & Modalities: Evaluates and documents own data and input usage, minimising dependence on unnecessary modalities and optimising resource efficiency during all operational phases. 91 Appendix D RFM-Overview 92 Table D5: Overview of all 324 identified RFMs that include manipulation capabilities (for control & integrated; see Section 3.2 for composition details). Reporting their proportional industrial applicability ratings for each individual implication, and a total score over all implications. Per-criterion ratings are created utilising the publication’s context, by leveraging our LLM-based pipeline as described in Section 8.1. Values are rounded to two decimal places. RefModelTotal↓I1I2I3I4I5I6I7I8I9I10I11 [157] Gemini Robotics 1.5 0.120.440.060.070.070.000.000.000.070.140.230.09 [158] Collab VLA 0.110.280.000.200.070.080.110.000.070.210.080.09 [159] GF- VLA 0.110.280.000.000.140.080.000.000.070.140.230.27 [6] OneTwo VLA 0.110.330.000.200.140.000.110.100.070.140.000.09 [160] SC- VLA 0.110.330.060.000.210.230.110.000.000.000.150.09 [164] HiBer NAC 0.100.220.110.000.000.080.110.100.000.210.150.09 [143] Sirius- Fleet 0.100.390.220.000.140.000.000.000.000.000.080.09 [211] Tactile- VLA 0.100.280.000.000.070.150.000.000.140.070.150.18 [212] DexGrasp VLA 0.090.280.000.000.070.080.000.000.070.000.310.18 [213] Gemini Robotics 0.090.220.000.000.070.000.000.000.140.140.230.18 [214]Robix0.090.220.110.200.000.000.000.000.000.000.310.09 [215] VITA- E 0.090.280.110.200.000.000.110.000.070.000.080.09 [167] PaLM- E 0.090.330.000.000.290.000.000.000.000.070.000.18 [216]RACER0.090.280.000.000.070.000.000.000.140.000.150.27 [217] Embodied- R1 0.090.280.000.000.000.080.000.000.000.140.230.18 [218]VLMimic0.080.330.000.000.140.000.000.000.070.000.150.09 [219] Intern VLA- M1 0.080.280.000.070.070.000.000.000.000.070.150.18 [220] FPC- VLA 0.080.280.060.000.140.080.000.000.000.000.230.00 [163]HAMSTER 0.070.110.000.000.000.000.000.100.000.070.230.36 [221] CoA- VLA 0.070.110.000.000.000.080.000.100.070.140.230.09 [222]ThinkAct0.070.280.000.000.140.000.000.000.000.000.150.18 continued on next page 93 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [223]MolmoAct 0.070.170.000.070.000.080.000.000.070.070.150.18 [224]RT-H0.070.220.000.130.000.080.000.000.070.070.000.18 [18]OpenVLA 0.070.220.000.000.070.000.000.000.000.000.380.09 [5]RoboDual 0.070.170.000.000.000.000.110.000.000.070.230.27 [225]VoxPoser0.070.170.000.130.070.150.000.000.000.000.150.09 [169]UniVLA0.070.110.000.000.000.080.000.000.070.000.230.36 [226]CLOVER0.070.280.000.000.070.000.000.000.070.000.230.00 [166]NanoVLA 0.070.110.000.000.070.080.000.200.000.000.310.00 [227] CLIP- RT 0.070.170.000.130.070.000.000.000.070.000.080.18 [165]LCIL0.070.110.000.270.000.000.110.000.000.000.080.18 [228] Switch VLA 0.060.220.000.070.000.000.110.000.000.000.080.18 [229]UVA0.060.110.000.000.070.080.000.000.000.000.310.09 [230]IDAGC0.060.110.000.130.000.000.000.000.000.140.080.18 [231]A20.060.060.000.000.000.150.000.000.070.070.150.18 [232]SuSIE0.060.110.000.000.000.080.000.000.000.000.230.27 [233] Omni Manip 0.060.280.000.000.070.000.000.000.000.000.150.09 [234]FoAM0.060.170.000.000.070.080.000.000.000.000.080.27 [235]VTLA0.060.060.000.000.140.080.000.000.000.070.150.18 [236]GNFactor0.060.170.000.000.000.000.000.000.070.070.150.18 [237] Diffusion- VLA 0.060.220.000.000.070.000.000.000.070.070.000.18 [238] GR- MG 0.060.220.000.000.000.000.000.000.000.000.230.18 [239] RT- Affordance 0.060.170.000.000.000.000.000.000.070.070.080.27 [240] Instruct VLA 0.060.170.000.000.000.000.000.000.070.070.230.09 [241] Robo Agent 0.050.170.000.000.140.000.000.000.000.000.150.09 [242]PointVLA 0.050.060.000.000.000.080.000.000.000.000.230.27 [243]EO-10.050.110.000.000.000.000.110.000.000.070.150.18 [244] Reason Manip 0.050.220.000.000.000.000.000.000.000.140.080.09 [245] WALL- OSS 0.050.170.000.070.000.000.110.000.000.000.080.18 [246]Vidar0.050.110.000.000.000.000.000.000.070.000.150.27 [247] Cronus VLA 0.050.060.000.000.140.000.000.000.070.000.230.09 continued on next page 94 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [248] OG- VLA 0.050.110.000.000.070.000.000.000.000.000.310.09 [249]OTTER0.050.170.000.000.000.000.000.000.070.000.230.09 [250] Giga Brain-0 0.050.060.000.000.000.000.000.100.000.140.080.27 [251] Emma- X 0.050.110.000.000.000.000.000.000.070.070.230.09 [252]LeVERB0.050.170.000.000.000.000.000.000.000.000.230.18 [253]ManiFlow 0.050.110.000.000.070.000.000.000.000.000.230.18 [1] Diffusion Policy 0.050.110.060.000.140.080.110.000.000.000.080.00 [254]MOO0.050.220.000.000.000.000.000.000.000.070.000.27 [255]MLA0.050.110.000.000.000.000.000.000.070.140.150.09 [256]GemBench 0.050.170.000.000.000.000.000.000.070.000.310.00 [257]IDHAP0.050.170.000.000.070.000.000.000.070.000.080.18 [258] Bridge VLA 0.050.170.000.000.000.080.000.000.000.000.230.09 [259] PRO- GRAM PORT 0.050.110.000.000.000.000.000.000.070.070.150.18 [260]RAM0.050.220.000.000.000.080.000.000.000.000.080.18 [261]VeBrain0.050.280.000.000.000.080.000.000.000.000.080.09 [262]UniAct0.050.110.000.000.000.080.000.000.000.000.230.18 [263]Actra0.050.170.000.000.070.000.000.000.070.000.080.09 [264]RVT-20.050.170.000.000.070.000.000.000.000.000.150.09 [265]TUDP0.050.000.000.000.000.080.000.100.000.070.150.18 [266]UniCoD0.050.170.000.000.000.080.000.000.000.000.150.09 [267] RT- Trajectory 0.050.110.000.000.000.000.000.000.070.000.150.18 [268]InSpire0.050.220.000.000.000.000.000.000.070.000.150.00 [269]TVVE0.050.110.000.000.000.080.000.000.000.000.310.00 [270]HPT0.050.060.000.000.070.000.000.000.000.000.150.27 [271]Hume0.050.220.000.000.070.000.000.000.000.000.150.00 [272]LCIL20.050.110.000.000.070.000.000.000.070.070.080.09 [273] LAP- VLA 0.050.060.000.000.000.000.000.000.000.000.230.27 [28]RT-10.050.170.000.000.000.000.000.000.070.000.150.09 [161]FuSe0.050.060.000.000.000.000.000.000.000.210.000.27 [274] SAM2 Act 0.050.110.000.000.000.080.000.000.000.000.310.00 continued on next page 95 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [275] Video Predic- tion Policy 0.050.060.000.000.000.000.000.000.000.070.150.27 [276] Lan- o3dp 0.050.170.060.000.070.000.000.000.000.000.080.09 [277] Interleave VLA 0.050.170.000.000.000.000.000.000.070.000.080.18 [278]LBMfDM0.050.060.000.000.000.000.000.000.000.000.230.27 [279]MResT0.050.110.000.000.000.080.110.000.000.070.080.09 [162] GraphCoT VLA 0.050.060.000.000.000.000.000.000.210.140.000.09 [280]ARM4R0.050.170.000.000.070.080.000.000.000.000.080.09 [281]SPECI0.040.110.000.000.000.000.000.000.000.000.150.18 [282]Gato0.040.000.000.000.000.000.000.000.070.000.150.27 [283]Evo-00.040.110.000.000.070.080.000.000.000.000.150.00 [284] MoT VLA 0.040.110.000.000.070.000.000.000.000.000.150.09 [285]GEVRM0.040.110.000.000.070.000.000.000.000.000.150.09 [286] Tiny VLA 0.040.170.000.000.070.000.000.000.000.000.150.00 [287] LoHo VLA 0.040.220.000.000.070.000.000.000.000.000.080.00 [288] Memory VLA 0.040.170.000.000.000.000.000.000.000.000.230.00 [289] RDT- 1B 0.040.110.000.000.000.080.110.000.000.000.000.18 [290] GR00T N1 0.040.110.000.000.000.000.000.000.000.000.080.27 [291]GR-10.040.170.000.000.070.000.000.000.000.000.000.18 [292]VIMA0.040.170.000.000.070.000.000.000.000.000.000.18 [293]BC-Z0.040.110.000.070.000.000.000.000.000.000.080.18 [294] Intention VLA 0.040.170.000.000.000.000.110.000.000.070.080.00 [295]CLIPort0.040.110.000.000.000.000.000.000.000.000.230.09 [7] Fast-in- Slow 0.040.170.000.000.000.080.000.000.000.000.080.09 [296]CAGE0.040.110.000.000.000.000.000.000.000.000.230.09 [297]SKIL0.040.060.000.000.000.080.000.000.000.000.150.18 [298] AR- VRM 0.040.110.000.000.000.000.000.000.070.000.080.18 [299]GeoVLA0.040.060.000.000.000.000.000.000.000.000.310.09 continued on next page 96 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [300] Skill Dif- fuser 0.040.110.000.000.000.000.000.000.070.070.150.00 [301] Goal- VLA 0.040.110.000.000.000.080.000.000.000.000.150.09 [302]VOTE0.040.110.000.000.000.080.000.100.000.000.150.00 [303]AffordDP0.040.110.000.000.000.000.000.000.000.000.080.27 [304]LangToMo 0.040.060.000.000.000.000.000.000.070.070.150.09 [305]GR-20.040.170.000.000.000.000.000.000.000.000.150.09 [20] π 0∗.6 0.040.000.000.000.070.150.000.000.070.000.080.09 [306]FLOWER 0.040.220.000.000.000.000.000.000.000.000.150.00 [307]MinD0.040.000.000.000.140.000.000.000.070.070.150.00 [308]Moto0.040.110.000.000.000.000.000.000.070.000.080.18 [309]CACTI0.040.220.000.000.000.000.000.000.000.000.080.09 [310] TCADP- LG 0.040.060.000.000.000.080.000.100.000.070.080.09 [311]LCB0.040.110.000.070.000.000.000.100.070.070.000.00 [168] Smol VLA 0.040.060.000.000.070.000.000.200.000.000.150.00 [312]cVLA0.030.060.000.000.000.000.000.000.000.000.080.27 [313]GR-30.030.110.000.000.000.000.000.000.000.000.080.18 [314] Hyper VLA 0.030.060.000.000.000.080.000.000.000.000.150.09 [315]E2VLA0.030.060.000.000.070.080.000.000.070.000.000.09 [316] Robo Brain 0.030.060.000.000.000.000.000.000.140.000.080.09 [317]OmniD0.030.000.000.000.000.080.000.000.000.000.230.09 [318] Force VLA 0.030.110.000.000.140.000.000.000.000.000.000.09 [319]VLMPC0.030.110.000.000.000.080.000.000.000.000.150.00 [320]AMPLIFY 0.030.060.000.000.000.000.000.000.000.000.150.18 [321] ReFine VLA 0.030.110.000.000.000.000.000.000.140.070.000.00 [322]CogACT0.030.110.000.000.000.000.000.000.000.000.230.00 [323]GRAPE0.030.110.000.000.000.080.000.000.000.000.150.00 [324]DNAct0.030.110.000.000.000.000.000.000.000.070.150.00 [325]Magma0.030.110.000.000.000.000.000.000.000.000.000.27 [326] GROOT- 2 0.030.110.000.000.000.000.000.000.070.000.080.09 [327]TriVLA0.030.110.000.000.000.000.000.000.000.000.150.09 [328]This&That 0.030.060.000.000.000.000.000.000.000.070.150.09 [329]TARAD0.030.170.000.000.000.000.000.000.000.000.000.18 continued on next page 97 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [330] DeL- TaCo 0.030.060.000.000.000.000.000.000.000.000.150.18 [331] Concept2 Robot 0.030.060.000.000.000.080.000.000.000.000.150.09 [332] Point Map Policy 0.030.000.000.000.000.000.000.000.070.070.150.09 [333] LP-ZS- LH 0.030.170.000.000.000.000.000.000.000.000.080.09 [334]VidBot0.030.060.060.000.000.000.110.000.000.000.080.09 [335]SayCan0.030.110.000.000.000.000.000.000.070.070.080.00 [336] AC- DiT 0.030.000.000.000.000.000.000.000.070.070.150.09 [337]VLAS0.030.060.000.000.000.000.000.000.000.070.150.09 [338]PALO0.030.110.000.000.000.000.000.000.000.000.150.09 [339] Trace VLA 0.030.170.000.000.000.000.000.000.000.000.150.00 [340]Villa-X0.030.000.000.000.000.000.000.000.000.070.230.09 [341] Sparse Diffu- sion Policy 0.030.110.000.000.000.000.000.000.070.070.000.09 [342] GRM- BALD 0.030.060.000.000.000.000.000.000.000.000.150.18 [343]Gen2Act0.030.110.000.000.000.000.000.000.070.000.000.18 [344] Robo Tron- Mani 0.030.060.000.000.000.000.000.000.000.070.150.09 [345]SAM-E0.030.060.000.000.000.000.000.000.070.070.080.09 [346] Hybrid VLA 0.030.110.060.000.000.000.000.000.000.000.150.00 [147]LATTE0.030.110.000.000.000.080.000.000.000.000.080.09 [347]CoMo0.030.060.000.000.000.000.000.000.000.000.150.18 [348]HiRT0.030.170.000.000.000.000.000.000.000.000.150.00 [349]FLARE0.030.060.000.000.000.000.000.000.000.070.000.18 [350]Seer0.030.110.000.000.000.000.000.000.000.000.080.09 [351]dVLA0.030.060.000.000.000.000.000.000.070.070.000.09 [352]TPM0.030.060.000.000.070.000.000.000.000.000.080.09 [353]Dita0.030.110.000.000.000.000.000.000.000.000.080.09 [354] Think Small- PPL 0.030.060.000.000.000.000.000.000.000.070.080.09 [355]LDRL0.030.110.000.000.070.000.000.000.000.000.080.00 continued on next page 98 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [356] Rynn VLA- 001 0.030.060.000.000.000.000.000.100.000.000.080.09 [357]BFA0.030.000.000.000.000.000.000.100.070.070.080.00 [358] M Diff Tran 0.030.060.000.000.000.000.000.000.000.000.080.18 [359] Robo L-RL 0.030.060.000.000.000.080.000.000.000.000.150.00 [360] LAA VTDC 0.030.110.000.000.000.000.000.000.000.000.080.09 [361]Dobb·E0.030.000.000.000.000.000.000.100.000.000.150.09 [19] Robo Flamingo 0.030.060.000.000.000.000.000.100.000.000.080.09 [362] Spatial VLA 0.030.110.000.000.070.000.000.000.000.000.080.00 [27]π 0.5 0.030.060.000.000.000.000.000.000.000.000.150.09 [363] Grasp VLA 0.030.060.000.000.000.000.000.000.000.000.080.18 [364]FP30.030.110.000.000.070.000.000.000.000.000.000.09 [365] π 0 - Text Latent 0.030.060.000.000.000.000.000.000.070.000.150.00 [366]ICRT0.030.110.000.000.000.000.000.000.000.000.000.18 [367] 3D- CAVLA 0.030.060.000.000.000.000.000.000.000.070.080.09 [368] OE- VLA 0.030.110.000.000.000.000.000.000.000.070.000.09 [369]LEO0.030.060.000.000.000.000.000.000.000.070.080.09 [370] Cog VLA 0.030.060.000.000.000.080.000.000.070.000.080.00 [371] Recon VLA 0.030.110.000.000.000.000.000.000.070.000.080.00 [372] Real World- RL- MVP 0.030.060.000.000.000.000.000.000.000.000.080.18 [373]DexVLA0.030.060.000.000.000.000.000.000.000.000.150.09 [374]GLVA0.030.110.000.000.000.000.000.000.000.000.150.00 [375] 3D Dif- fusion Policy 0.030.110.000.000.000.000.000.000.000.000.080.09 [376] Being- H0 0.030.060.000.000.000.000.000.000.000.000.080.18 continued on next page 99 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [377] AirExo- 2 0.030.000.000.000.000.000.000.100.000.070.150.00 [29]RT-20.030.060.000.000.000.000.000.000.000.000.150.09 [378] Perceiver- Actor 0.030.060.000.000.000.000.000.000.000.000.080.18 [379]H-RDT0.030.060.000.000.000.000.000.000.000.000.080.18 [380] Tra- MoE 0.030.000.000.000.000.080.000.000.000.000.150.09 [381] 3D- VLA 0.030.060.000.000.000.000.000.000.000.070.080.09 [382]ILRMS0.030.000.000.000.070.000.000.000.070.000.000.18 [383] VLA- Adapter 0.030.060.000.000.000.000.000.000.000.000.230.00 [384] Lang Cond- RoboIL 0.030.060.000.000.000.000.000.000.000.000.150.09 [385] Bi- VLA 0.030.110.000.000.000.000.000.000.000.000.000.18 [386]DyWA0.030.060.000.000.000.080.000.000.000.000.150.00 [3] Dream VLA 0.030.060.000.000.000.000.000.000.000.070.080.09 [387] Diffusion Trans- former Policy 0.030.110.000.000.000.000.000.000.070.000.000.09 [26]π 0 0.020.060.000.000.000.000.000.000.000.000.080.09 [388] Rozum Former 0.020.000.000.000.000.000.000.000.000.000.000.27 [389]RoboCat0.020.000.000.000.000.000.000.000.000.000.150.09 [390]IGOR0.020.060.000.000.000.000.000.000.000.000.000.18 [391]ACT0.020.000.000.000.000.000.000.100.000.000.080.09 [392]ImitDiff0.020.110.000.000.000.000.000.000.000.000.080.00 [393]Surfer0.020.110.000.000.000.000.000.000.000.000.080.00 [394] HyT- VLA 0.020.000.000.000.000.000.000.000.000.070.150.00 [395] Open VLA OFT 0.020.000.000.000.000.000.000.000.000.000.150.09 [396] VLM- TDP 0.020.060.000.000.000.000.000.000.000.000.150.00 [397]GRIF0.020.000.000.000.000.000.000.000.000.000.080.18 [398]RVT0.020.000.000.000.000.080.000.000.000.000.150.00 [399] Instruct RL 0.020.060.000.000.000.000.000.000.000.000.000.18 continued on next page 100 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [400]CDP0.020.060.000.000.000.000.000.000.000.000.080.09 [401]DTP0.020.060.000.000.000.000.000.000.000.000.080.09 [402] Flow VLA 0.020.060.000.000.000.000.000.000.000.000.080.09 [403] FLASH VLA 0.020.000.000.000.000.000.000.100.000.000.080.09 [404] Robo- MUTUAL 0.020.000.000.000.000.000.000.000.000.000.150.09 [405]FLIP0.020.060.000.000.000.000.000.000.070.000.080.00 [406]LCDOM0.020.060.000.000.000.000.000.000.000.000.000.18 [407] PD- VLA 0.020.060.000.000.070.000.000.000.000.000.080.00 [408] Cross EmbL 0.020.060.000.000.000.000.000.000.000.000.080.09 [409] Cross Former 0.020.110.000.000.000.000.000.000.000.000.000.09 [410]ReinboT0.020.000.000.000.000.000.000.000.000.000.080.18 [411]AdaMoE0.020.060.000.000.000.000.000.000.000.000.150.00 [412]OWL×T5 0.020.060.000.000.000.000.000.000.000.000.000.18 [413] P - agent 0.020.060.000.000.000.000.000.000.070.000.080.00 [414]Surfer0.020.110.000.000.000.000.000.000.000.000.080.00 [415] ADDP- AVP 0.020.000.000.000.000.000.000.000.000.000.080.18 [416]OCI0.020.000.000.000.000.000.000.000.000.000.150.09 [417]PLEX0.020.060.000.000.000.000.000.000.000.000.000.18 [418]EnerVerse 0.020.000.000.000.000.000.000.000.070.000.080.09 [419]ARP0.020.060.000.000.000.000.000.000.070.000.080.00 [420] Robo Ground 0.020.060.000.000.000.000.000.000.000.070.080.00 [421]DMPEL0.020.060.000.000.000.000.000.000.000.000.080.09 [422] Object VLA 0.020.000.000.000.000.000.000.000.000.000.080.18 [423] Robo Uni View 0.020.060.000.000.000.080.000.000.000.000.080.00 [2] Discrete Diffu- sion VLA 0.020.110.000.000.000.000.000.000.000.000.080.00 [424]ROSA0.020.110.000.000.000.000.000.000.000.000.080.00 [425] CoT- VLA 0.020.000.000.000.000.000.000.000.000.070.080.09 continued on next page 101 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [426]PDFactor0.020.000.000.000.000.080.000.000.000.000.080.09 [427] VoxAct- B 0.020.060.000.000.000.000.000.100.000.000.080.00 [428] Long- VLA 0.020.110.000.000.000.000.000.000.000.000.000.09 [429]Octo0.020.060.000.000.000.000.000.000.000.000.000.18 [430]CrossT50.010.000.000.000.000.000.000.000.000.000.000.18 [431]PAFMF0.010.000.000.000.000.000.000.000.000.000.150.00 [432] Humanoid VLA 0.010.060.000.000.000.000.000.000.000.000.000.09 [433] Robo BERT 0.010.060.000.000.000.000.000.000.000.000.080.00 [434]OMLA0.010.000.000.000.000.000.000.000.000.000.080.09 [4]Dreamitate 0.010.060.000.000.000.000.000.000.000.070.000.00 [435]EgoVLA0.010.060.000.000.000.000.000.000.000.000.080.00 [436] LanCon- Learn 0.010.060.000.000.000.000.000.000.000.000.000.09 [437] CtF Q- attention 0.010.000.000.000.000.000.000.000.000.000.080.09 [438]DAWN0.010.000.000.000.000.000.000.000.000.070.080.00 [439]StaMo0.010.000.000.000.000.000.000.000.000.000.080.09 [440]Tenma0.010.060.000.000.000.000.000.000.000.000.000.09 [441] VQ- VLA 0.010.000.000.000.000.000.000.000.000.000.080.09 [442] VLA- Arm- TDMTO 0.010.000.000.000.000.000.000.000.000.000.150.00 [443]Vlaser0.010.000.000.000.000.000.000.000.000.000.080.09 [444]DualMind 0.010.000.000.000.000.000.000.000.070.000.000.09 [445]M2Distill0.010.000.000.000.000.000.000.000.000.000.080.09 [446]VidMan0.010.060.000.000.000.000.000.000.000.000.080.00 [447] DP- VLA 0.010.000.000.000.000.000.110.000.000.000.080.00 [448]F10.010.000.000.000.000.000.000.000.070.000.080.00 [153]HAMLET 0.010.000.000.000.000.000.000.000.070.000.080.00 [449] 4D- VLA 0.010.060.000.000.000.000.000.000.000.000.080.00 [450]Act3D0.010.060.000.000.000.000.000.000.000.000.000.09 [451] UP- VLA 0.010.060.000.000.000.000.000.000.000.000.080.00 [452]ABM0.010.060.000.000.000.000.000.000.000.000.000.09 [453] World VLA 0.010.000.000.000.000.000.000.000.000.000.080.09 continued on next page 102 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [454]UniVLA0.010.000.000.000.000.000.000.000.000.000.000.18 [455]BitVLA0.010.000.000.000.000.000.000.100.000.000.000.09 [456] Chat VLA-2 0.010.060.000.000.000.000.000.000.000.000.080.00 [457] DeeR- VLA 0.010.060.000.000.000.000.000.100.000.000.000.00 [458] Spatial Traces 0.010.000.000.000.000.000.000.000.000.000.080.09 [459]ReVLA0.010.060.000.000.000.000.000.000.000.000.080.00 [460] Alpha Block 0.010.060.000.070.000.000.000.000.000.000.000.00 [461]PACT0.010.000.000.000.000.000.000.000.000.000.000.18 [462] Natural VLM 0.010.060.000.000.000.000.000.000.000.000.000.09 [463] Omni VTLA 0.010.000.000.000.000.000.000.000.070.000.000.09 [464]KIVLA0.010.060.000.000.000.000.000.000.000.000.000.09 [465]KALM0.010.000.000.000.000.000.000.000.000.000.080.09 [466] Light VLA 0.010.000.000.000.000.000.000.100.070.000.000.00 [467]MBA0.010.000.000.000.000.080.000.000.000.000.080.00 [468]Avi0.010.060.000.000.000.000.000.000.000.070.000.00 [469] Chat VLA 0.010.060.000.000.000.000.000.000.000.000.000.09 [470]MIDAS0.010.060.000.000.000.000.000.000.000.000.000.09 [471] Meta RL-LI 0.010.060.000.000.000.000.000.000.000.000.080.00 [472]BLM10.010.060.000.000.000.000.000.000.000.000.000.09 [473] Mini Dif- fuser 0.010.060.000.000.000.000.000.000.000.000.000.09 [474] Discrete Policy 0.010.000.000.000.000.000.000.000.000.000.080.09 [475] π 0 .6 (model card) 0.010.060.000.000.000.000.000.000.000.000.000.09 [476]NORA0.010.060.000.000.000.000.000.000.000.000.000.00 [477]VIRT0.010.000.000.000.000.000.000.000.000.000.080.00 [478]PolarNet0.010.000.000.000.000.000.000.000.000.000.080.00 [479] MoLe- VLA 0.010.000.000.000.000.000.000.100.000.000.000.00 [480] AgiBot World Colosseo 0.010.000.000.000.000.000.000.000.000.000.000.09 continued on next page 103 RefModelTotalI1I2I3I4I5I6I7I8I9I10I11 [481] Q- Trans- former 0.010.000.000.000.000.000.000.000.000.000.080.00 [482] LLaDA- VLA 0.010.060.000.000.000.000.000.000.000.000.000.00 [483]NBAgent0.010.060.000.000.000.000.000.000.000.000.000.00 [484]LiteVLP0.010.000.000.000.000.000.000.000.000.000.000.09 [485] Ex- PERACT 0.010.000.000.000.000.000.000.000.000.000.000.09 [486] Robo Mamba 0.010.000.000.000.000.000.000.000.000.000.080.00 [487] Context VLA 0.010.000.000.000.000.000.000.000.000.000.080.00 [488]RetoVLA0.010.000.000.000.000.000.000.000.000.000.080.00 [489] OC- VLA 0.010.000.000.000.000.000.000.000.000.000.000.09 [490] Control VLA 0.010.000.000.000.000.000.000.000.000.000.000.09 [491]RoboDiT0.010.000.000.000.000.000.000.000.000.000.000.09 [492] EC- Flow 0.010.000.000.000.000.000.000.000.000.000.000.09 [493] Chained Dif- fuser 0.010.000.000.000.000.000.000.000.000.000.000.09 [494]OFA0.010.000.000.000.000.000.000.000.000.000.000.09 [495] VAE- Rob Manip 0.010.000.000.000.000.000.000.000.000.000.000.09 [496] Geo Aware- VLA 0.000.000.000.000.000.000.000.000.000.000.000.00 [497] CEED- VLA 0.000.000.000.000.000.000.000.000.000.000.000.00 [498] Spec- VLA 0.000.000.000.000.000.000.000.000.000.000.000.00 [499] Robo Flamingo- Plus 0.000.000.000.000.000.000.000.000.000.000.000.00 [500] Robotic- CLIP 0.000.000.000.000.000.000.000.000.000.000.000.00 [501]EdgeVLA0.000.000.000.000.000.000.000.000.000.000.000.00 [502]LLARVA0.000.000.000.000.000.000.000.000.000.000.000.00 [503] VGGT- DP 0.000.000.000.000.000.000.000.000.000.000.000.00 [504]ScaleDP0.000.000.000.000.000.000.000.000.000.000.000.00 104 Table D6: Criterion fulfilment for I1 – Adaptability & Flexibility. Criterion P ↓ % fulfilled Multi-Task Generalisation16049.38% Generalisation to Environmental Variation 10131.17% Autonomous Handling of Distractors/Obstacles 8726.85% Online/Mid-Execution Adaptation 4915.12% Autonomous Incorporation of Skills 226.79% Online Environment Adaptation and Response 216.48% Self-Evaluation & Command Adherence 195.86% Multi-Hardware Support164.94% Mid-Execution Feedback Responsiveness 144.32% Explained Adaptation Decisions103.09% Multimodal Command Interface72.16% Memory of Previous Faults/Distractors 41.23% Restricted Autonomous Task Switching 30.93% Autonomous Fallbacks20.62% Energy/System-Aware Adaptation 10.31% Operational Adaptation Confidence 10.31% Strategy-Goal Alignment10.31% Hot-Swappable & Dimensionality Adaptation 00.00% Table D7: Criterion fulfilment for I2 – Safety & Compliance. Criterion P ↓ % fulfilled Redundant, Independent Safety Mechanisms 51.54% Command Compliance with Online-Safety 41.23% Proactive, Plan-Level Safety30.93% Learning from Safety Events20.62% Decoupled, Real-Time Safety Control 10.31% Reactive, Real-Time Safety10.31% Rule-Adherence & Autonomous Rejection 10.31% Authorised Operator Detection & Override 00.00% External Safety Integration00.00% Fault Detection & Sensing00.00% Integrable Safety Systems00.00% Persistence of Safety Capabilities in Adaptation 00.00% Persistent Safety Logs00.00% Safe, Autonomous Recovery00.00% Safety Confidence Output00.00% Safety Documentation & Traceability 00.00% Standards-Compliant00.00% Transparent Safety Reasoning00.00% 105 Table D8: Criterion fulfilment for I3 – HRI & HRC. Criterion P ↓ % fulfilled Intuitive Feedback Acceptance134.01% Collaborative Adaptation82.47% Transparent Reasoning Explanation 51.54% Real-time Interruptibility41.23% Continuous Responsiveness20.62% Active Operator Exclusivity00.00% Exception Handling & Safe State Transition 00.00% Human Presence Awareness in Workspace 00.00% Human-Centric Adaptation00.00% Industrial HRI/HRC Standards Compliance 00.00% Instruction Confidence00.00% Learned Operator Preferences00.00% Multimodal Instruction & Output 00.00% Operator Data Privacy00.00% Operator Identification & Authorisation 00.00% Table D9: Criterion fulfilment for I4 – Robustness & Reliability. Criterion P ↓ % fulfilled Disturbance Resilience & Continuous Operation 4413.58% Environmental Robustness & Adaptation 144.32% Graceful Degradation & Strategy Adaptation 61.85% Reproducible & Benchmarked Fault Diagnosis 51.54% Adaptive Error Learning & Memorisation 30.93% Logging, Explanation & Operator Guidance 10.31% Automated Recovery & Safe-State Transition 00.00% Autonomous Fault Categorisation & Severity Response 00.00% Consistent & Documented Fault Handling 00.00% Multi-Level Fault & Anomaly Detection 00.00% Online Confidence & Reliability Assessment 00.00% Robust Redundancy & Modality Management 00.00% Supervisor Oversight & Fault Review 00.00% Threat & Attack Detection00.00% 106 Table D10: Criterion fulfilment for I5 – Precision & Accuracy. Criterion P ↓ % fulfilled Consistency Over Time & Contexts3811.73% Skill Integration with Known v/P&A 92.78% v/P&A Learning from Experience & Feedback 61.85% HRI/HRC Integration30.93% User-Defined Trade-offs20.62% Disturbance & Uncertainty Management 10.31% v/P&A Instructability10.31% Achievable v/P&A-Proposals00.00% Autonomous v/P&A Adaptation00.00% Outcome Logging & Failure Reasoning 00.00% Supervision & Verification Module00.00% Transparent Documentation of v/P&A Capabilities 00.00% v/P&A-Confidence Score & Explanation 00.00% Table D11: Criterion fulfilment for I6 – Real-time Performance. Criterion P ↓ % fulfilled Goal-Responsive Solution Strategies 123.70% Historical Data Utilisation20.62% System-wide Real-time Compliance10.31% Task, Input & Environment Independence 10.31% Autonomous Adaptation & Safety Response 00.00% Dynamic Resource Management & Determinism 00.00% Live Metrics, Logging & Explainability 00.00% Real-Time Confidence/Alerting00.00% Transparent Documentation & Compliance 00.00% Table D12: Criterion fulfilment for I7 – Cost-effectiveness & Integration Capabilities. Criterion P ↓ % fulfilled Dynamic Resource Allocation164.94% Low-Cost Hardware Readiness82.47% Comprehensive Integration Documentation 10.31% Autonomous Fault Handling & Feedback 00.00% Cost-Reduction Memory & Adaptive Optimisation 00.00% Flexible Input/Output Adaptation00.00% Guided & Autonomous Commissioning 00.00% Hot-Swapping & Auto-Detection00.00% Integration Performance Monitoring & Learning 00.00% Universal Integration without Retraining 00.00% 107 Table D13: Criterion fulfilment for I8 – Explainability & Trust. Criterion P ↓ % fulfilled Performance Benchmarking & OOD Feedback 6319.44% Explains Influences & Disturbances41.23% Reason Logging & Accessible Justification 30.93% Task Feasibility & Capability Explanation 30.93% Error, Anomaly & Safeguard Explanations 20.62% User Feedback & Adaptive Improvement 20.62% Behaviour Consistency & Pre-Adaptation Notification 00.00% Confidence Score Explanation00.00% Continuous Self-Assessment with Explainability 00.00% Explainability Across System Components 00.00% Flexible, Intuitive Explainability Interfaces 00.00% Monitors & Logs Explainability for Learning 00.00% Rule & Norm Compliance with Explainability 00.00% Security & Authorisation Feedback00.00% Table D14: Criterion fulfilment for I9 – Sensor Fusion & Perception. Criterion P ↓ % fulfilled Perceptual Feature Exposure & Modularity 329.88% Comprehensive Scene Understanding via Fusion 298.95% Multimodal Sensor Acceptance & Fusion 113.40% Flexible Input Combination & Goal Specification 103.09% Dynamic Sensor Prioritisation20.62% Peer-to-Peer Perception Knowledge Exchange 10.31% Autonomous Disturbance Recognition & Countermeasures 00.00% Continuous Input Monitoring & Feedback 00.00% Modular Integration & Documentation 00.00% Online Learning for Perception Skills 00.00% Perception Evaluation & Benchmarking 00.00% Perceptual Confidence & Explanation Output 00.00% Robust Fault Tolerance & Degradation Handling 00.00% Security in Sensor Operations00.00% 108 Table D15: Criterion fulfilment for I10 – Standardised Benchmarking & Evaluation. Criterion P ↓ % fulfilled Repeatability & Reproducibility14544.75% Module & Skill-Specific Benchmarking 12939.81% Generalisation & Adaptation Assessment 9027.78% Benchmark Diversity & Scope4714.51% Instruction, Command, & Modality Coverage 30.93% Explainability, Reasoning, & Feedback Benchmarks 10.31% Integration & Commissioning Evaluation 10.31% Data Requirement & Autonomous Runtime Assessment 00.00% Hardware Platform Benchmarking00.00% Human Interaction & Object Recognition Evaluation 00.00% Robustness & Fault Tolerance Evaluation 00.00% Safety, Security, & Norm Compliance Validation 00.00% Transparent Metrics & Reporting00.00% Table D16: Criterion fulfilment for I11 – Data Requirements & Usage. Criterion P ↓ % fulfilled Flexible Data Type Acceptance15848.77% Low-Data Adaptation & Few-Shot Requirements 13641.98% Documented Data Composition & Simulation-Real Transfer 3310.19% Efficient Task, Environment, & Modality Adaptation 144.32% Resilience Through Redundancy & Fault Tolerance 20.62% Self-Assessment to Reduce Input & Modalities 10.31% Automatic Data Sufficiency Assessment & Feedback 00.00% Compliance with Privacy, Ethics, & Law 00.00% Comprehensive, Diverse Pretraining Dataset 00.00% Explainability on Data Usage & Improvement 00.00% Ongoing Adaptation/Finetuning Confidence Scores 00.00% 109