Paper deep dive
InternW0: A Foundational Physical World Model for Efficient Real-World Interactions
Jisong Cai, Yao Mu, Ganlin Yang, Zhe Cao, Zhangzheng Tu, Xing Gao, Kailin Li, Xinyu Zhan, Lixin Yang, Yangkun Zhu, Haoxiang Ma, Ming Zhou, Qiaojun Yu, Yufei Xue, Liqun He, Yifei Yao, Yifan Zhu, Long Ling, Bingqi Jiang, Haoyu Guo, Xueyue Zhu, Bowen Zhou, Bin Zhao, Tianfan Xue, Chunhua Shen, Weinan Zhang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/26/2026, 3:14:24 AM
Summary
The paper introduces InternW0, a foundational physical world model from Shanghai AI Laboratory designed for efficient real-world robot interactions. It employs an asymmetric video-action architecture with flow matching, featuring a high-capacity video expert for long-horizon prediction and a lightweight action expert for fast control. Key innovations include asynchronous multi-frequency processing, observation-conditioned context routing to reuse cached K/V states, and contact-aware post-training using force/tactile signals. The model is trained on 7,200 hours of heterogeneous data, including the EgoLab dataset, and demonstrates success in simulation benchmarks and real-world tasks like metal-organic framework synthesis and dexterous pipetting.
Entities (9)
Relation Signals (8)
InternW0 → developedby → Shanghai AI Laboratory
confidence 100% · We introduce InternW0, the first instantiation of the InternW physical world model series from Shanghai AI Laboratory
InternW0 → trainedon → EgoLab
confidence 95% · We train InternW0 on approximately 7,200 hours of heterogeneous robot and egocentric data, including EgoLab
InternW0 → usesmethod → Flow Matching
confidence 95% · InternW0 jointly learns ... through an asymmetric video--action architecture with flow matching
InternW0 → performstask → Metal-organic framework synthesis
confidence 90% · Evaluation spans ... real-world scientific tasks, including a 15-stage metal–organic framework synthesis workflow
InternW0 → performstask → Quantitative pipetting
confidence 90% · Evaluation spans ... 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipetting
InternW0 → usescomponent → Wan VAE
confidence 90% · A frozen Wan VAE (Wan et al., 2025) ... provide visual representations
InternW0 → usescomponent → DINOv3
confidence 90% · current chunk-level visual observations are encoded by a frozen DINOv3 encoder
InternW0 → usesmethod → Context Routing
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Physical intelligence requires more than predicting how the world may evolve: predictions must remain actionable as the world continues to change. We introduce InternW0, the first instantiation of the InternW physical world model series from Shanghai AI Laboratory, built around omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observations and external influences. InternW0 jointly learns future visual dynamics and continuous robot control through an asymmetric video--action architecture with flow matching. A high-capacity video expert provides longer-horizon predictive context, while a lightweight action expert operates at a faster timescale. Instead of regenerating the future for every action update, InternW0 reuses layerwise K/V and adapts it to newly observed states through observation-conditioned context routing. Domain-specific interfaces and soft prompts support heterogeneous embodiments, while contact-aware post-training incorporates force and tactile signals for contact-rich manipulation. We train InternW0 on approximately 7,200 hours of heterogeneous robot and egocentric data, including EgoLab, a 275-hour real-laboratory egocentric dataset. Evaluation spans simulation benchmarks and real-world scientific tasks, including a 15-stage metal--organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipetting. These results advance scalable, asynchronous, and science-native physical world models for universal and efficient real-world interactions.
Tags
Links
- Source: https://arxiv.org/abs/2609.27656v1
- Canonical: https://arxiv.org/abs/2609.27656v1
Trouble viewing inline? Open PDF directly →
Full Text
88,224 characters extracted from source content.
Expand or collapse full text
— -- 上海人工智能实验室 -:=-.==.. � Shanghai Artificial Intelligence Laboratory InternW0: A Foundational Physical World Model for Efficient Real-World Interactions Physical Intelligence Team, Shanghai AI Laboratory InternW0 Toward General Physical World Models From embodied experience to general physical intelligence. Perceive Predict Act Observe Vision Language Force / Tactile Action 1 Heterogeneous Pretraining Data 2 InternW0 Asynchronous World-Action World Modeling with contact 3 Simulation Results 7.2K Hours Contact-Rich Data Real-Robot Data Simulation Data Ego-Centric Data Unified State-Action Interface VideoStateActionForce/Tactile Future Video Action [Force] ... Video Expert (Slow) Action Expert (Fast) Context Routing Video Encoder Video Vision Encoder Video (Slow) Action (Fast) ......... ... Predict slowly, act quickly. ... Real-World Results Make Sandwich Real-World Success Rate (%) π 0.5 Fast-WAM InternW0 73.3 40.0 73.3 53.1 50.6 82.7 86.7 66.7 88.9 51.6 10.7 68.4 46.7 18.7 65.3 Make Sandwich Pick Industrial Parts Sort Tubes MoF Experiments Quantitative Pipetting Inference Speed Frequency (Hz) ↑ 0.54 5.26 16.47 MotusFast-WAM InternW0 1866.10 ms 0.10× 190.00 ms 1.00× 60.73 ms 3.13× 4 Pick Industrial Parts Sort Tubes MoF Experiments Quantitative Pipetting RoboTwin 2.0-Full CleanRandomized 88.66 87.02 92.90 91.50 93.20 93.04 Motus LingBotVA InternW0 RoboTwin 2.0-C2R CleanRandomized 68.0 20.9 66.8 67.9 83.20 68.0 X-VLA GigaBrain-0.7 InternW0 Libero 94.4 Motus InternW0 π 0 Fast-WAMLingbotVA 96.9 97.7 97.6 98.5 98.6 π 0.5 Average ... Soft Prompt Action Encoder Abstract Physical intelligence requires more than predicting how the world may evolve: predictions must remain actionable as the world continues to change. We intro- duce InternW0, the first instantiation of the InternW physical world model series from Shanghai AI Laboratory, built around omnimodal interfaces, asynchronous multi-frequency processing, and local physical modeling under partial observa- tions and external influences. InternW0 jointly learns future visual dynamics and continuous robot control through an asymmetric video–action architecture with flow matching. A high-capacity video expert provides longer-horizon predictive context, while a lightweight action expert operates at a faster timescale. Instead of regenerating the future for every action update, InternW0 reuses layerwise K/V and adapts it to newly observed states through observation-conditioned context routing. Domain-specific interfaces and soft prompts support heterogeneous em- bodiments, while contact-aware post-training incorporates force and tactile signals for contact-rich manipulation. We train InternW0 on approximately 7,200 hours of heterogeneous robot and egocentric data, including EgoLab, a 275-hour real-laboratory egocentric dataset. Evaluation spans simulation benchmarks and real-world scientific tasks, includ- ing a 15-stage metal–organic framework synthesis workflow and 5-stage contact- and force-aware dexterous manipulation for general-purpose quantitative pipet- ting. These results advance scalable, asynchronous, and science-native physical world models for universal and efficient real-world interactions. Project page: https://internrobotics.github.io/internw0 arXiv:2609.27656v1 [cs.RO] 23 Sep 2026 1 Design Motivation and Overall Framework Physical intelligence concerns an AI agent’s ability to interact with the physical world, requiring it to connect its understanding of the environment and the task with the consequences of its actions. This connection demands more than recognizing objects or interpreting instructions: the agent must estimate the current physical state, anticipate how candidate actions will change it, and use feedback to revise its decisions. These capabilities must operate under partial observations (Kael- bling et al., 1998) and limited computation, even as the environment continues to evolve. A world model establishes this connection by representing how physical states evolve in response to actions, enabling the agent to translate its understanding of the environment and the task into informed de- cisions (Ha and Schmidhuber, 2018; Hafner et al., 2025). Recent advances in physical world modeling have increasingly focused on world action models (WAMs), which couple future visual prediction with robot action generation and provide a promis- ing way to ground predictive dynamics in executable interaction (Wang et al., 2026a; Yuan et al., 2026; Bi et al., 2026; Ye et al., 2026). A unified WAM is expected to support heterogeneous em- bodiments and richer physical modalities such as proprioception, force, and tactile feedback, while avoiding the latency introduced by synchronously updating expensive video prediction and fast action generation. However, these remain insufficiently addressed by existing WAMs. In this report, we develop the InternW world model series to provide this connection through a shared model of perception, physical dynamics, and action. Following the resource-constrained perspective from Shanghai AI Laboratory (Chen et al., 2026c), a physical world model is viewed as a compact approximation of physical state transitions under finite sensing and computational re- sources. This perspective motivates three requirements of world modeling: omnimodal perception and interaction, asynchronous processing at multiple frequencies, and local environment modeling under external influences. • Omnimodal interfaces: InternW aims to unify diverse modality inputs and predict actions within a shared physical representation, enabling embodied agents to perceive, reason, and interact with a wide range of physical environments. • Asynchronous and multi-frequency signal processing: InternW accommodates different sensing and control frequencies to balance prediction quality, responsiveness, and compu- tational cost. • Local environment and action modeling: InternW estimates local physical state from partial observations and predicts the effects of both agent actions and external disturbances, updating its estimates as new evidence arrives. Shared Representation and Specialized Experts. These requirements motivate the InternW ar- chitecture described here at the series level. We draw on the modality-specific parameterization principle of mixture-of-transformers (MoT) (Liang et al., 2025). Our series-level design adapts this principle to modalities (or data channels) for language semantics, visual dynamics, spatial geome- try, contact mechanics, and body actions. Modality-specific processing thus supports joint reason- ing about task intent, object configuration, contact conditions, and executable motion. The cross- modality interfaces are extensible, allowing individual InternW models to select the modalities (or data channels) and computational capacity appropriate for their deployment setting. Latent Dynamics for Prediction and Action. At the center of the framework, latent dynamics connect the estimated physical state to possible future states and interactions. Learning latent dynamics for planning from visual observations has been explored in prior model-based control work (Hafner et al., 2019). To express this role conceptually, letH t contain the timestamped obser- vations and executed actions available up to time t, and let z t denote a compact representation of local state and its uncertainty. We write z t = F θ (H t ),p θ (z t+∆ | z t , u t:t+∆ ,∆),(1) 2 where u t:t+∆ denotes a candidate action sequence and∆ is the prediction horizon. Unobserved external influences contribute to the uncertainty of the transition distribution. This formulation specifies the modeling role rather than prescribing a particular probabilistic implementation. Pre- diction interfaces translate the latent representation into future observations or task-relevant states, while action interfaces generate commands conditioned on the task and available physical informa- tion. Their shared representation connects what the agent expects to happen with what it chooses to do. Under finite computational resources, we seek latent predictions that preserve physical in- formation relevant to task outcomes and action choices, consistent with control-centric world mod- eling (Hansen et al., 2024). In the InternW framework, we additionally require predictive informa- tion to become available in time to influence ongoing execution. We therefore regard predictive accuracy, decision relevance, and timeliness as three design criteria for our framework. Duplex Interaction and Closed-Loop Feedback. InternW couples perception, prediction, and ac- tion through duplex interaction: it continues to receive new observations while generating predic- tions and taking actions. Duplex interaction describes this concurrent input–output flow, whereas asynchronous, multi-frequency processing determines when each sensor stream and computa- tional component is updated. Together, these mechanisms allow visual, contact, and body-state feedback to refine subsequent outputs while an operation is in progress. Each executed action produces new observations that update the local physical representation and inform the next prediction–action cycle. In scientific workflows, this closed loop connects experimental operations to their measured outcomes and supplies evidence for subsequent model refinement. Correct oper- ation requires timestamp alignment and causal consistency: each update must account for actions that have already been executed while remaining able to revise commands that have not yet been issued. Environment Geometry Video Contact Semantic Action InternW Series Shared Physical Representation Task Planning Place the egg in the plate. LocateGraspPlace Video Prediction 3D Prediction Contact Prediction Action Generation Future Physical States Predict the consequences of actions in the physical world Multimodal Observations Observe and understand the current physical world Task “ Make me a breakfast.” Visual 3DContact State Figure 1: Framework of InternW world model series. Multimodal Feedback for Contact-Aware Manipulation. InternW is designed to incorporate force, tactile, and proprio- ceptive feedback into action generation alongside visual predictive context. These physical modalities provide complemen- tary information about contact and the robot’s physical state, including changes that may not be directly observable in im- ages (Lee et al., 2019; Chen et al., 2023; Song et al., 2026; Yu et al., 2026).Be- cause these signals arrive at different tem- poral resolutions, the model separates long-horizon predictive context from re- cent feedback used to update each ac- tion.This enables local action correc- tion at every feedback step without re- quiring the full predictive representation to be recomputed. Specifically, InternW0 fuses the temporal history of interaction forces with visual and proprioceptive ob- servations, while jointly predicting the end-effector poses and six-dimensional in- teraction wrenches.Each wrench com- prises three force and three moment com- ponents (Lynch and Park, 2017). The his- torical force inputs encode observed con- tact, whereas the predicted wrenches represent anticipated interaction loads. Together, they pro- vide information about both past contact and anticipated loads to support contact-aware action 3 selection. Active hybrid position–force interaction is a core manipulation capability that coordi- nates motion with contact-force regulation (Raibert and Craig, 1981; Li et al., 2026c). InternW0 as an Initial Instantiation. Building on the above design principles, InternW0 con- nects physical prediction and responsive interaction through a mixture-of-transformers (MoT) ar- chitecture. Specialized video and action experts are jointly optimized through video–action flow matching, coupling future visual dynamics with executable motion while preserving distinct pro- cessing pathways. Within the action expert, force feedback is integrated with visual and propri- oceptive observations, enabling changes in contact and body state to condition subsequent action updates. This fusion provides a direct pathway from real-time sensory feedback to local motion ad- justment. To coordinate future prediction with real-time observation, an observation-conditioned context-routing interface uses current visual features to query predictive video representations and construct chunk-specific context for the action expert. The resulting duplex interaction allows longer-horizon predictive context to guide execution while incoming observations continually in- form subsequent actions, without requiring a new prediction cycle for every interaction update. Domain-specific state and action encoders, action decoders, and soft prompts adapt this shared ar- chitecture to heterogeneous datasets and embodiments. These mechanisms concretely realize the framework’s multimodal interfaces, asynchronous processing, and observation-driven local up- dates. Within the InkStone scientific discovery platform (InkStone, Shanghai AI Laboratory), our system connects the scientific reasoning and tool-use capabilities of Intern-S2-Preview (Bai et al., 2026) to physical experimentation through InternW0. Execution records and measured outcomes can support subsequent evaluation and model improvement. The following sections detail the architecture and training pipeline, data recipe, simulation benchmarks, real-robot and wet-lab ex- periments, as well as the future work of InternW world model series. 2 InternW0 Architecture and Training Pipeline 2.1 Architecture InternW0 instantiates the preceding framework with a mixture-of-transformers (MoT) backbone comprising a video expert for future visual prediction and a lightweight action expert for robot con- trol. Each joint layer contains modality-specific blocks with separate parameters and token streams, connected through an observation-conditioned video-context interface. This structure allows the two experts to use different capacities while coupling visual prediction with action generation. The central design goal is to keep future predictions useful as new sensory signals arrive during execution. This requires coordinating prediction and control at different timescales: repeatedly recomputing a complete predictive plan at every control update is expensive, whereas executing against an unchanged plan context risks conditioning the policy on an increasingly stale view of the world. InternW0 addresses this tension through asynchronous duplex inference, in which predictive video modeling and action generation operate at different temporal scales while remaining coupled through an observation-conditioned context interface. The heterogeneous pretraining stage learns this coupling from visual observations, proprioceptive states, and robot actions. For contact-rich downstream tasks, the action interface is further extended during post-training with force and tac- tile observations and joint prediction of future interaction signals. Embodiment-specific interfaces and soft prompts support transfer across heterogeneous control spaces and sensing configurations. The interaction is intentionally asymmetric. The video expert builds a shared predictive repre- sentation from visual observations and language without consuming robot-specific action tokens or domain identities. The action expert, in turn, reads the video context together with the lat- est observation, proprioception, language conditioning, and embodiment-specific interfaces. This separation keeps physical prediction broadly shared while allowing control to specialize across heterogeneous embodiments. 4 IDM-mask f 0 f 1 a 1 f 0 f 1 a 1 Video DiT Learn from vast video data Action DiT Learn action dynamics across domains Ego & Internet Videos No action labels required Heterogeneous Robot Data State–action supervision Required Optional (Posttrain) Context routing Vision Encoder Prompts Force Action Actions [Force] ... VAE Decoder Cross Attention & FFN Self Attention AdaLN & Q K V Wan VAE Encoder Action Decoder Cross Attention & FFN Joint Attention AdaLN & Q K V Soft prompt Action encoder Figure 2: Architecture of InternW0. Modality-specific video and action experts are coupled through an observation-conditioned chunk K/V editor. The editor adapts cached video context to chunk-level observations, allowing it to be reused across action updates. A frozen Wan VAE and DINOv3 encoder provide visual representations, while precomputed language embeddings con- dition both experts. Domain-specific interfaces and soft prompts support heterogeneous embodi- ments; force and tactile channels are introduced optionally during contact-aware post-training. As shown in Figure 2, video frames are encoded by a frozen Wan VAE (Wan et al., 2025), while current chunk-level visual observations are encoded by a frozen DINOv3 encoder (Siméoni et al., 2026). Language instructions are represented by precomputed text embeddings. Trainable projec- tions connect these representations to the corresponding experts, and proprioceptive states provide chunk-specific conditioning for action generation. During post-training on contact-rich tasks, force and tactile histories can be introduced as additional action-side observations without modifying the shared video-prediction backbone. Asynchronous Duplex Inference. Future visual prediction provides anticipatory context, but its computational cost makes regenerating a plan for every control update impractical. Meanwhile, action generation must incorporate observations that arrive after a plan was produced. InternW0 separates these timescales: the video expert updates its prediction on a slower schedule, while the action expert generates short chunks from the latest available plan and current sensory feedback. After an initial plan is generated, action inference can continue while the next video prediction is computed. This coordination is implemented through a cached video context and a layerwise chunk K/V editor. Following the observation-guided video-context routing design of AHA-WAM (Cai et al., 2026a), the latest chunk-level observation is used to adapt the cached predictive context before it is con- sumed by the action expert. For each action chunk n, the chunk-aligned RGB observation is first encoded by DINOv3 and projected into a visual context. A learned query encoder summarizes this context into a small set of routing queries. 5 Let (K ℓ , V ℓ ) denote the valid video context keys and values at layerℓ. The editor first lets the observation queries q retrieve task-relevant information from the predictive video representation, R ℓ,n = Attn W q ℓ q n , K ℓ , V ℓ .(2) It then routes the retrieved information back to the original video-token positions, D ℓ,n = Attn ( K ℓ , R ℓ,n , R ℓ,n ) ,(3) from which a lightweight projection predicts residual corrections (∆K ℓ,n ,∆V ℓ,n ). The chunk-specific video context is e K ℓ,n = K ℓ + σ(γ ℓ )∆K ℓ,n , e V ℓ,n = V ℓ + σ(γ ℓ )∆V ℓ,n , (4) where γ ℓ is a learned gate. The final residual projection is zero-initialized, so the editor begins as an identity mapping and learns to introduce observation-dependent corrections progressively. The underlying video context remains unchanged: each action chunk constructs its own edited view from the same predictive context using its newly observed visual state. Consequently, the model can preserve longer-horizon predictive structure while adapting its local control context as execution deviates from the previously imagined future. Soft Prompts for Cross-embodiment Pretraining. Sharing interaction knowledge across robots requires retaining the differences in their action semantics and sensing configurations. Inspired by X-VLA’s design of embodiment-specific prompts for cross-embodiment learning (Zheng et al., 2026), InternW0 associates each training domain d with learned soft prompts P (d) , prepended to every action chunk. Domains distinguish data sources and embodiment configurations, and need not correspond one-to-one to robot morphologies. The prompts condition shared expert compu- tation on these differences, while domain-specific input and output projections adapt action and sensor representations to the shared token space. Per-dimension validity masks distinguish miss- ing channels from valid zero values: unavailable inputs are zeroed before token processing and excluded from the corresponding loss. Together, these interfaces allow joint training without re- quiring identical control or sensing semantics across robots. Mixed Supervision for Prediction and Control. Action-free egocentric videos provide observa- tions of physical interactions, while action-labeled robot trajectories connect those interactions to executable control. InternW0 trains the same video expert on both sources, whereas robot trajecto- ries additionally supervise the action expert. Real-robot and simulated trajectories follow the same action-supervised pathway, while egocentric videos contribute predictive supervision without re- quiring robot action annotations. Both experts use continuous flow matching. For a clean target x 0 , representing either video latents or a continuous action chunk, Gaussian noise ε, and normalized flow time τ ∈ [0, 1], the noisy input and target velocity are x τ = (1− τ)x 0 + τε,v ∗ = ε− x 0 .(5) Flow times are sampled using the shifted continuous flow-matching schedule, with one flow time sampled for each video example and independent flow times for individual action chunks. Flow times are sampled per video example and per action chunk. Video attention uses a first-frame- causal mask: the observed first latent frame cannot attend to future video tokens, whereas future latent frames can attend to the observed frame and interact bidirectionally with one another. The first frame remains clean as a temporal anchor and is excluded from the video prediction loss. The pretraining objective combines the video and action flow-matching losses, L video and L joint , respectively, over samples ξ drawn from the mixed training distribution D mix : L pre = E ξ∼D mix λ v L video (ξ) + λ a I act (ξ)L joint (ξ) ,(6) where λ v and λ a balance their contributions, and I act indicates whether action supervision is avail- able. Both terms are scheduler-weighted, masked squared velocity errors. Temporal padding, the clean current video frame, and unavailable action dimensions are excluded from the corresponding losses. Egocentric samples therefore contribute video supervision without requiring action labels. 6 IDM Conditioning. During action-supervised training, InternW0 maintains a second video stream that provides predictive conditioning for the action expert. This independent-denoising con- dition stream shares all VideoDiT parameters with the primary video flow-matching stream but uses an independently sampled noise realization and flow time. For each training example, the condition stream is kept clean with probability 0.5; otherwise, it is perturbed using an independently sampled shifted flow time. In both cases, the first latent frame remains clean. The primary video stream is supervised by the video flow-matching ob- jective, whereas the condition stream does not receive a separate reconstruction loss. Instead, its layerwise K/V features are consumed by the observation-conditioned editor and subsequently by the action expert. Importantly, the condition K/V features are not detached from the video network. Gradients from the action flow-matching objective therefore propagate through the K/V editor into the shared video expert. Predictive representation learning is consequently shaped by both future-video mod- eling and its usefulness for action generation, rather than by visual reconstruction alone. At inference time, the ground-truth condition stream is replaced by a generated video context. Its layerwise K/V features are cached and exposed to the action expert through the same context interface. Contact-Aware Post-training with Joint Action–Contact Prediction. The heterogeneous pre- training stage does not require force or tactile annotations. For downstream tasks in which contact dynamics are critical, we extend the pretrained action interface with measured interaction signals and jointly predict their future evolution together with robot actions. Visual observations alone may not fully resolve the contact conditions that determine whether a manipulation succeeds. During contact-aware post-training, the action expert is therefore condi- tioned on available force and tactile histories, and its prediction target at physical timestep i is augmented as y i = a ⊤ i , f ⊤ i , h ⊤ i ⊤ ,(7) where a i is the action representation, f i is the force/torque signal, and h i is the tactile represen- tation. For each chunk, only contact history available before its generation is supplied as condi- tioning; future contact channels are noised and denoised jointly with the action channels. This formulation links observed interaction history to both subsequent commands and their anticipated sensory consequences. The action channels are decoded for execution, while the contact channels represent predicted feedback. This extension preserves the shared predictive video backbone while augmenting the action-side interface with task-specific physical interaction signals. Together, these mechanisms connect future visual context, ongoing sensory feedback, and embodiment-specific control: predictions guide action generation, new observations revise sub- sequent chunks, and shared training links diverse interaction experience to executable behavior. 2.2 Training Pipeline The training pipeline for InternW0 is organized as a three-layer stack that strictly separates data sampling, distributed execution, and model optimization. The data plane is built on an in-house data engine, which curates heterogeneous robot-episode corpora into unified streams of training-ready samples. The execution plane is built on the top of Ray (Moritz et al., 2018), which manages all cluster processes and co-locates data materialization with GPU-side training under explicit resource and memory budgets. The optimization plane is implemented on top of an in-house distributed train- ing framework: its gradient trainer owns the training loop proper, while parallelism is delegated to a pluggable backend protocol, into which InternW0 installs the sharding, checkpointing, and compilation strategy outlined below. Streaming Data Pipeline. To maximize data-sampling throughput of data engine, we decouple metadata from the actual sample payloads throughout the data pipeline. Instead of directly shuf- 7 fling and communicating raw data, samples are first represented as compact, contiguous descrip- tors, and only metadata is reordered during sampling, substantially reducing communication and data-movement overhead. The corresponding payloads are materialized on demand according to the requirements of downstream training tasks. When combining heterogeneous data sources, we apply domain-level mixing weights to prevent any single corpus from dominating the training stream simply because of its raw data volume. The resulting stream is then partitioned across train- ing nodes while preserving block order, such that batches processed locally on each node remain relatively homogeneous in their domain distribution. Sample-level random choices are generated deterministically from sample identities rather than execution or scheduling order, allowing a re- sumed run to reproduce exactly the same data stream. In addition, to avoid redundant I/O for video-heavy workloads, we implement a node-shared caching mechanism for video data, enabling samples already fetched and decoded by one worker to be reused across workers on the same node. This significantly reduces repeated reads and improves the overall data-loading throughput. Parallelism, Precision, and Compilation. InternW0 uses composable FSDP2 whose sharding units follow the model’s structure: each joint mixture-of-transformers layer and each expert stage is sharded independently, while the narrow per-domain interfaces are replicated across ranks. A single-node job uses a flat sharding mesh; in multi-node execution, the same configuration trans- parently promotes to a hierarchical replicate-and-shard mesh, with no changes to the model or checkpoint schema. Mixed-precision training retains FP32 master parameters and optimizer states, while model computation is performed in BF16. Distributed gradient reduction is performed in FP32 to improve numerical stability. Each joint layer is compiled as a fixed-shape, full-graph unit before activation-checkpointing and sharding wrappers are applied. Selective activation check- pointing then rematerializes only a configurable prefix of layers, trading memory for recomputa- tion without changing the objective. Training state is persisted transactionally through distributed checkpointing, enabling exact mid-epoch resumption. 3 Data Recipe The pretraining corpus of InternW0 contains approximately 7,200 hours of data from seven datasets, spanning real and simulated robot manipulation as well as human egocentric labora- tory activities. Actions and states from the robot datasets are mapped to a unified 37-dimensional representation. Numerical preprocessing, data filtering, and sampling-index construction are per- formed offline; during training, the pipeline reads only the resulting artifacts and precomputed normalization statistics. To balance contributions across datasets, we sample them in proportion to the square root of their numbers of effective anchors. This gives larger datasets more weight while limiting the extent to which any one dataset dominates training. Robot data cleaning targets defects that affect the supervision signals: static trajectories, intervals affected by head or waist motion, brief gripper pulses, and missing or out-of-range values. These defects are identified on each dataset’s native timeline and handled with validity masks or window invalidation, retaining usable data while excluding unreliable supervision. Data Composition. As shown in Table 1, the corpus comprises four real-robot datasets, includ- ing AgibotWorld (Bu et al., 2025), Galaxea (Jiang et al., 2025), RoboCOIN (Wu et al., 2025), Mol- moAct (Lee et al., 2025), also two simulated datasets, InternData-A1 (Tian et al., 2026) and Ro- boDojo (Chen et al., 2026b), and one dataset of human egocentric laboratory videos (EgoLab) col- lected by our team. The robot datasets include single-arm, dual-arm, and mobile manipulation. Dexterous-hand robot data and dataset copies whose format conversion is incomplete are excluded from this pretraining corpus. We define training domains by data provenance, robot morphology, and subset identity; thus, domains do not map one-to-one to physical robot morphologies. This round comprises 25 training domains and 24 configured statistics groups, with the two Franka grip- per variants sharing a single group. Of these, 23 groups provide robot-state/action normalization statistics; the remaining group corresponds to EgoLab, which contributes only video supervision during pretraining. For language conditioning, we use a valid sub-task instruction when available and otherwise fall back to the task-level instruction. 8 Table 1: Pretraining data composition. DatasetTypeSubsets#Domains #Hours #Episodes InternData-A1 (Tian et al., 2026) Simulated Franka (Panda / Robotiq85), Genie1, Lift2, Split ALOHA, AC-One 63,494.3568,194 AgibotWorld (Bu et al., 2025)Realgripper and tactile subsets22,620.0158,380 RoboCOIN (Wu et al., 2025)RealAgiBot G1 (2 subsets), Cobot Magic (5 subsets), Galbot G1, R1 Lite, RMC-AIDA-L, Split ALOHA (2 subsets), Tianqin A2 13438.159,952 Galaxea (Jiang et al., 2025)RealR1 Lite1314.915,362 EgoLabEgovideo-only pathway1275.43,192 MolmoAct (Lee et al., 2025)RealYAM bimanual170.33,424 RoboDojo (Chen et al., 2026b)Simulated dual-arm ARX X5120.53,465 Total257,233.5811,969 EgoLab Dataset. We collect the EgoLab dataset to support WAMs in carrying out real-world lab- oratory experiments. It comprises 275 hours of egocentric RGB video of bimanual human ma- nipulation in real wet-laboratory settings. These demonstrations capture manipulation priors in laboratory instrument handling and encode implicit task specifications and procedural constraints through the ordering of operations and the handling of samples and vessels. The dataset aims at infusing this laboratory-specific operational knowledge into WAMs. Video Acquisition. Participants perform specified laboratory tasks with selected apparatus, follow- ing task descriptions while controlling the recording with a wristband device. A DJI Action-series camera is secured to each participant’s head with a strap or helmet. Videos are recorded at 2.5K resolution and 50 fps and cover operations such as pipetting, sample weighing, titration, filtration, grinding with a mortar and pestle, stirring, and rinsing glassware. Figure 3 shows representative laboratory activities, reconstructed hand annotations, and the distribution of activity categories. Vessel preparation Open reagent bottle Liquid handling Pour the liquid Mixing & grinding Grind in a mortar Equipment Load a centrifuge Solid handling Scrape powder Cleaning 6.1% Solids 8.0% Equipment 11.9% Mixing 20.0% Prep & records 3.3% Vessels 25.7% Liquids 25.1% Original Annotated Figure 3: EgoLab dataset. Left: example activities with original frames above and hand annotations below. Right: distribution of activity categories. 3D Motion Reconstruction. Camera motion and the trajectories of both hands are reconstructed at the native recording rate of 50 fps. We use HaWoR (Zhang et al., 2025) for general processing and a separate reconstruction method for challenging cases. The reconstructed trajectories provide auxiliary annotations for filtering; EgoLab is used only during InternW0 pretraining through the video-only pathway. Subtask Partitioning. We use a vision-language model accessed through a public API for open- vocabulary subtask partitioning. Frames sampled at 0.5 fps are used to generate a global activity caption and fine-grained descriptions of visible hand–object interactions in 10 s clips. Adjacent descriptions are merged into a subtask when they refer to the same laboratory operation, primary sample or vessel, experimental cycle, and intended outcome. The resulting partitions are mapped to training clips using the same rules as for robot datasets. 9 Unified Action and State Interface. Robot actions and states are mapped to a fixed 37- dimensional representation. Each arm occupies 17 dimensions: 7 for joint values, 3 for end-effector translation, 6 for orientation in the rotation-6D representation, and 1 for the gripper. The remain- ing 3 dimensions are reserved for the mobile base. Missing channels are zero-filled and marked by validity masks so that they do not contribute to the action loss. Head and waist degrees of freedom are outside the supervision space. left joints (7D), left E∆pos (3D), left E∆rot (6D), left gripper (1D), right joints (7D), right E∆pos (3D), right E∆rot (6D), right gripper (1D), base (∆x,∆y,∆yaw) Action semantics depend on the channel type. Joint actions follow each source’s convention of ab- solute next-waypoint targets. End-effector actions encode the translation delta and relative rotation from the observation at the current control step to the next-waypoint target. Base actions encode planar displacement and yaw change over the corresponding interval. In the state representation, end-effector positions and orientations remain absolute, with orientations represented by the first two columns of the rotation matrix. Joint, gripper, and base states and actions are normalized sep- arately. For 6D end-effectors, only the translation delta is normalized in the action representation; absolute states and action rotation-6D values remain in their raw form. Temporal Sampling and Sample Construction. We enforce 15 fps model-time convention through temporal subsampling across all datasets. Native 60/50 fps streams use stride 4, 30/25 fps streams use stride 2, and 15 fps streams retain their original stride. For this convention, 50 fps streams are treated as 60 fps and 25 fps streams as 30 fps, while video frames are still read from their native frame positions. In stride-4 sequences, starts at frames 0, 1, 2, and 3 define four phases. Effective anchors are counted on phase-0 only; at sampling time, a legal phase for the selected an- chor is chosen uniformly at random. For robot data, a legal action-window offset is then chosen within that phase. Images, states, actions, and masks remain aligned throughout each sample. Each robot training sample contains 13 multi-view frames and 64 action steps. The multi-view images are stitched into a 384× 320 canvas. Consecutive video frames are separated by 8 control steps, so the 13 frames span 96 control steps. The action sequence is divided into four 16-step chunks, each conditioned on the image and robot state at the start of that chunk. The action window is randomly offset by 0-15 steps, and no history images are used. EgoLab follows the same video-window and temporal-sampling scheme. Domain-Mixture Sampling. Prior robot foundation models use sublinear data reweighting to mitigate imbalances in training data: RDT-1B initializes dataset weights proportionally to the square root of dataset size (Liu et al., 2024), while π 0 weights each task-robot combination pro- portionally to m 0.43 , where m is its number of training samples (Black et al., 2024). Following this principle, we use fixed square-root weights at the domain level. Let N d denote the number of effective phase-0 anchors in domain d; its target sampling probability is p d = √ N d ∑ j p N j .(8) This weighting samples larger domains more frequently while reducing their dominance relative to size-proportional sampling. Each GPU batch contains samples from a single domain, although suc- cessive batches may switch domains and different GPUs may sample different domains. Domain- homogeneous batches are required by the per-domain action interfaces. Source-level target pro- portions are the sums of their constituent domain weights; actual consumption proportions are monitored throughout training. Offline Cleaning and Filtering. Numerical preprocessing, filtering, and sampling-index con- struction are completed before training. For each source, we first standardize fields, gripper en- codings, and the temporal relationship between actions and feedback to construct unnormalized 10 37-dimensional states and actions. We then generate legal intervals, phases, action windows, and effective-anchor indices. Training reads these derived artifacts directly. Cleaning is based on the signals used for supervision: • Static trajectories. We use raw motion signals to determine whether valid motion occurs and remove episodes with no valid action throughout. This test uses neither downsampled sequences nor normalized values. Trajectories containing only gripper or base motion are retained. • Head and waist motion. Intervals affected by head or waist motion are excluded, along with any action window that intersects them. Other legal portions of the same episode are retained; action sequences are never concatenated across excluded intervals. • Gripper short pulses. We detect open-close-open and close-open-close round trips on the native sequence using values normalized by the corresponding statistics group, side, and signal, before clipping. High and low signal levels are identified by normalized values of at least 0.6 and at most−0.6, respectively. We then check each sampling phase to determine whether subsampling omits the intermediate reversal. If it does, the action chunk containing the event is excluded from valid supervision, and the affected window offsets are removed. • Missing channels and boundaries. Missing action dimensions and temporal padding are marked with validity masks. Zero-filled missing channels do not contribute to the loss, whereas genuinely valid zeros remain supervised. Window boundaries and action validity are checked separately for each phase; validity determined for phase 0 is not assumed to hold for other phases. Normalization Statistics and Split Hygiene. Normalization statistics are computed separately for each statistics group from values that can be sampled from the training set. The computation covers all legal phases and excludes missing channels, padding, and validation data. Gripper- pulse detection initially uses reference statistics computed from the training set. Once filtering is complete, final statistics are recomputed from the retained valid supervision. Absolute end-effector states and rotation-6D values bypass this normalization. 4 Simulation Benchmarks We evaluate InternW0 in simulation to examine two complementary properties of the pretrained model: its ability to support strong downstream manipulation after task-specific post-training, and its ability to retain useful physical representations when the downstream evaluation distribution shifts beyond the demonstrations used for adaptation. We consider LIBERO (Liu et al., 2023) and RoboTwin 2.0 (Chen et al., 2025a). LIBERO provides a standard single-arm manipulation benchmark spanning spatial reasoning, object interaction, goal- conditioned control, and long-horizon execution. RoboTwin 2.0 provides a broader 50-task bi- manual setting and, importantly, allows us to contrast two downstream data regimes: Full, where both clean and randomized demonstrations are available during post-training, and Clean2Random, where post-training uses only clean demonstration data and randomization is introduced only at evaluation time. Together, these evaluations separate downstream task adaptation from generalization beyond the downstream training distribution. We compare InternW0 with representative vision-language- action (VLA) and world-action model (WAM) baselines. For each benchmark, we emphasize com- parisons under the same evaluation setting and distinguish in-domain performance from out-of- domain generalization where applicable. 4.1 Single-Arm Manipulation We first evaluate whether large-scale physical world-model pretraining translates into strong task- specific manipulation performance on LIBERO (Liu et al., 2023). We evaluate InternW0 on the 11 Table 2: LIBERO Full setting. MethodSpatialObjectGoalLongAverage π 0 (Black et al., 2024)98.096.894.488.494.4 InternVLA-M1 (Chen et al., 2025b)98.099.093.892.695.9 π 0.5 (Intelligence et al., 2025)98.898.298.092.496.9 GR00T-N1.7 (NVIDIA et al., 2025)97.798.597.594.497.0 OpenVLA-OFT (Kim et al., 2025)97.698.497.994.597.1 Fast-WAM (Yuan et al., 2026)98.2100.097.095.297.6 Motus (Bi et al., 2026)96.899.896.697.697.7 LingBot-VA (Li et al., 2026b)98.599.697.298.598.5 InternW099.499.498.697.098.6 standard LIBERO benchmark, covering all 40 tasks across the four official task suites: LIBERO- Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long. Post-training setup. For LIBERO, we initialize from the pretrained InternW0 checkpoint and post-train the model for 50 epochs with a learning rate of 5× 10 −5 . We retain the same unified 37- dimensional action interface used during pretraining. Since LIBERO is a single-arm manipulation benchmark, its native absolute end-effector action representation is mapped to the corresponding left-arm slots in the 37-dimensional action space, while the unused dimensions are zero-padded and masked out from training supervision. This allows the shared video–action backbone to be preserved while adapting only the valid control dimensions required by the LIBERO embodiment. Following the benchmark protocol, we conduct 50 rollouts for each task and report the average success rate for each suite, as well as the macro-average over the four suites. We compare InternW0 against representative vision-language-action (VLA) and world-action model (WAM) baselines, in- cluding π 0 (Black et al., 2024), π 0.5 (Intelligence et al., 2025), GR00T-N1.7 (NVIDIA et al., 2025), OpenVLA-OFT (Kim et al., 2025), InternVLA-M1 (Chen et al., 2025b), Xiaomi-Robotics-0 (Cai et al., 2026b), Motus (Bi et al., 2026), LingBot-VA (Li et al., 2026b), and Fast-WAM (Yuan et al., 2026). Results and Analysis. Table 2 summarizes the results on the standard LIBERO benchmark. In- ternW0 achieves an average success rate of 98.6%, outperforms LingBot-VA (Li et al., 2026b) and Fast-WAM (Yuan et al., 2026) by 0.1 and 1.0 percentage points, respectively. In particular, InternW0 achieves the best performance on LIBERO-Spatial, reaching a success rate of 99.4% and exceeding the strongest reported baseline result by 0.6 percentage points. It also obtains 99.4% on LIBERO- Object and 98.6% on LIBERO-Goal, with the latter being only 0.2 percentage points below the best reported result. On the more challenging long-horizon suite, InternW0 achieves a success rate of 97.0%. Overall, the consistently strong performance across spatial reasoning, object manipulation, goal-conditioned control, and long-horizon tasks demonstrates that the pretrained model transfers effectively to single-arm manipulation and provides a competitive foundation for general-purpose robot control. 4.2 Bimanual Manipulation We evaluate InternW0 on bimanual manipulation using RoboTwin 2.0. We consider two com- plementary settings. RoboTwin 2.0-Full measures in-domain multi-task performance when both clean and randomized demonstrations are available during training, whereas RoboTwin 2.0- Clean2Random evaluates out-of-domain generalization from clean training environments to un- seen randomized scenes. Post-training setup. Both RoboTwin settings initialize from the pretrained InternW0 checkpoint and retain the future-IDM formulation used during pretraining. We use a 64-step action horizon divided into four 16-step chunks and a random action offset in [0, 15]. Chunk observations are 12 Table 3: RoboTwin 2.0 Full setting. MethodClean (%)Randomized (%)Avg. (%) π 0 (Black et al., 2024)65.9258.4062.16 π 0.5 π 0.5 (Intelligence et al., 2025)82.7476.7679.75 ABot-M0 (Yang et al., 2026b)81.2080.4080.80 Motus (Bi et al., 2026)88.6687.0287.84 Fast-WAM (Yuan et al., 2026)91.8891.7891.83 LingBot-VA (Li et al., 2026b)92.9091.5092.20 AHA-WAM (Cai et al., 2026a)93.4092.2092.80 OpenWAM-α (Wang et al., 2026b)93.7493.4693.60 ABot-M0.5 (Chen et al., 2026a)94.0094.2094.10 InternW093.2093.0493.12 encoded with the frozen DINOv3 (Siméoni et al., 2026) encoder, and the action and proprioceptive signals are represented through the unified 37-dimensional interface. The native 14-dimensional ALOHA joint-action representation is mapped to its corresponding slots in this interface, while the remaining dimensions are zero-padded and masked out from training supervision. The condition- video stream is independently noised with probability 0.5, and the Video-DiT remains trainable so that downstream action supervision can continue to shape the predictive representation. The model architecture and overall optimization recipe are kept aligned between Full and Clean2Random, and both settings are post-trained for 5 epochs with a learning rate of 2× 10 −5 . 4.2.1 RoboTwin 2.0-Full We first evaluate InternW0 on RoboTwin 2.0-Full, covering all 50 bimanual manipulation tasks. Following the standard multi-task training protocol, the model is fine-tuned jointly on the clean and randomized training demonstrations and evaluated separately under the two test settings. We report the task-averaged success rate. Results and Analysis. Table 3 summarizes the results on RoboTwin 2.0-Full. InternW0 achieves 93.20% success in the clean setting and 93.04% under randomized evaluation, yielding an aver- age success rate of 93.12%. The difference between the two settings is only 0.16 percentage points, showing that the model maintains highly stable performance when scene randomization is intro- duced during both training and evaluation. Among the WAM baselines, InternW0 improves over Fast-WAM (Yuan et al., 2026) and AHA- WAM (Cai et al., 2026a) by 1.29 and 0.32 percentage points in average success, respectively, and exceeds LingBot-VA (Li et al., 2026b) by 0.92 points. Its average performance is also close to OpenWAM-α (Wang et al., 2026b), with a gap of only 0.48 points. ABot-M0.5 (Chen et al., 2026a) achieves the highest average success rate among the compared methods at 94.10%, while InternW0 remains within 0.98 points. Overall, these results show that InternW0 can effectively adapt to a diverse set of bimanual manipulation tasks while maintaining stable performance across clean and randomized environments. 4.2.2 RoboTwin 2.0-Clean2Random We further evaluate InternW0 under the RoboTwin 2.0-Clean2Random setting to measure out-of- domain generalization. The model is fine-tuned exclusively on clean demonstrations and is then evaluated under both clean and randomized environments. Because randomized scene configu- rations are not observed during training, this setting directly tests transfer to previously unseen visual and environmental conditions. Results and Analysis. Table 4 presents the results under Clean2Random evaluation. InternW0 achieves 83.20% success on clean scenes and 68.0% on randomized scenes, resulting in an average success rate of 75.60%. This is the highest average success rate among all compared methods, 13 Table 4: RoboTwin 2.0 Clean2Random setting. MethodClean (%)Randomized (%)Avg. (%) Fast-WAM (Yuan et al., 2026)77.81.939.9 X-VLA (Zheng et al., 2026)68.020.944.5 X-WAM (Guo et al., 2026)70.025.847.9 Spatial Forcing (Li et al., 2026a)77.226.752.0 π 0.5 (Intelligence et al., 2025)70.746.058.4 4D-WAM (Yang et al., 2026a)81.541.861.6 GigaBrain-0.7 (Team et al., 2026)66.867.967.3 InternW083.2068.075.60 exceeding GigaBrain-0.7 (Team et al., 2026), the strongest baseline in terms of average success, by 8.30 percentage points. On the clean split, InternW0 reaches 83.20%, surpassing the strongest clean- split baseline, 4D-WAM (Yang et al., 2026a), at 81.5% by 1.70 percentage points. Under randomized evaluation, InternW0 also achieves the highest success rate among the listed methods, marginally outperforming GigaBrain-0.7 (Team et al., 2026) at 67.9% by 0.1 percentage points. The resulting 15.2-point gap between clean and randomized evaluation indicates that unseen scene variations remain challenging. Nevertheless, the model preserves a higher absolute performance level than the compared baselines on both splits. Taken together, the Clean2Random results pro- vide evidence that InternW0 transfers effectively from clean training data to previously unseen randomized environments. 4.2.3 Egocentric Video Pretraining with Limited Data No pretraining Robot only Robot + EgoLab 0 20 40 60 80 Clean success (%) 55.20 72.13 76.40 +4.27 p CleanRandomized 0 10 20 30 40 Randomized success (%) 6.73 13.73 21.97 +8.24 p Figure 4: Pretraining data ablation on RoboTwin C2R. Outlines: Clean (left axis); solid fills: Randomized (right axis). Gains are relative to robot-only pretraining (p). We investigate how egocentric video contributes to downstream generalization under the RoboTwin 2.0- Clean2Random protocol. This ablation uses restricted pretraining subsets, whereas the main InternW0 result in Table 4 uses the full pretraining corpus We further compare three pretraining configurations: no pretrain- ing, robot-data-only pretraining, and joint pretraining on robot data and EgoLab videos. Pretraining uses ~50 h of robot data, plus ~50 h of EgoLab videos for the joint vari- ant. All variants are subsequently fine-tuned on clean demonstrations and evaluated on both clean and ran- domized environments. Results and Analysis. Figure 4 summarizes the results within this limited-data setting. Compared with robot- data-only pretraining, adding EgoLab videos improves clean success from 72.13% to 76.40% and randomized success from 13.73% to 21.97%, corresponding to gains of 4.27 and 8.24 percentage points. The larger gain on ran- domized scenes suggests that adding egocentric video is particularly beneficial for generalization beyond the clean fine-tuning distribution in this limited-data setting. 14 Sort PartsZip Bag Operate OvenUnscrew Cap Make Sandwich Sort Rubbish Place Markpen PlaceFlower Sweep Trash Wipe Stain Sort PartsZip Bag Operate OvenUnscrew Cap Make Sandwich Sort Rubbish Place Markpen PlaceFlower Sweep Trash Wipe Stain Make Sandwich Pick Industrial Parts Sort Tubes MoF Experiments Quantitative Pipetting Figure 5: Five real-world tasks. From left to right: Make Sandwich, Pick Industrial Parts, Sort Tubes, MoF Experiments, and Quantitative Pipetting. The first four are performed on a dual-arm platform, and the last one on a dexterous-hand platform. 5 Real-Robot and Wet-Lab Experiments 5.1 Real-Robot Experiments We evaluate on five real-robot tasks, shown in Figure 5, which fall into two categories: four dual-arm manipulation tasks performed with two parallel-gripper arms, and one dexterous-hand manipulation task performed with a multi-fingered hand mounted on a single arm. Dual-Arm Manipulation. Make Sandwich is a simple long-horizon task in which the robot picks up two slices of bread and a piece of meat and stacks them in order. Pick Industrial Parts requires the robot to identify each part and place it into its matching box. Sort Tubes requires the robot to follow a language instruction that specifies an arm and a tube color (e.g., “Using the right arm, pick up the orange-tipped tube and place it into the right box”) and place the matching tubes into the box on the corresponding side. MoF Experiments reproduces the solution preparation stage of a metal-organic framework synthesis: the robot pours liquid from a graduated cylinder into a flask through a funnel, removes the funnel, transfers the flask onto a stirrer, inserts the stopper, and turns on the stirrer, totaling 15 sequential subtasks that require precise insertion and progress tracking across visually similar stages. Dexterous-Hand Manipulation. Quantitative Pipetting evaluates dexterous manipulation and compliant tool use with a 20-degree-of-freedom (DoF) dexterous hand coupled to a 7-DoF robotic arm with hybrid force–position control. The task comprises five sequential subtasks: (i) Pickup & re- orientation, in which the robot grasps and reorients the pipette through coordinated multi-finger in- hand manipulation; (i) Tip attachment, in which the robot attaches a disposable pipette tip through hybrid force–position control; (i) Liquid aspiration, in which the robot aspirates a prescribed vol- ume of liquid; (iv) Liquid dispensing, in which the liquid is dispensed into a designated test tube; and (v) Tip ejection & return, in which the used tip is ejected and the pipette is returned to its initial position. The task requires precise hand–arm coordination, stable contact maintenance, compliant interaction control, and reliable execution of the complete manipulation sequence. To quantify performance, we use three complementary metrics that match the structure of each task. For every task, we run 15 trials with randomized initial object positions and orientations. For Make Sandwich, we report the episode-level success rate (SR), where a trial is successful only when all required actions are completed correctly. Pick Industrial Parts and Sort Tubes each present multiple target objects in a single trial, so we report the object-level success rate, i.e., the fraction of target objects that are correctly picked and placed across all trials, which measures the recall of the policy over the targets on the table. For the two multi-stage tasks, MoF Experiments and Quantitative Pipetting, all methods share the same VLM for subtask generation and we report the progress rate, computed as the proportion of subtasks completed correctly and in the prescribed order, averaged over all trials. As shown in Table 5, InternW0 matches or outperforms π 0.5 on all five tasks and outperforms Fast- WAM by a large margin. On Make Sandwich, both pretrained policies reach 73.3%, as basic pick- and-stack behaviors are already well covered by pretraining. On Pick Industrial Parts, InternW0 improves the object-level success rate from 53.1% to 82.7%: most failures of π 0.5 are recognition 15 Table 5: Real-world results on all five tasks. We report success rate (%) for Make Sandwich, Pick Industrial Parts, and Sort Tubes, and progress rate (%) for MoF Experiments and Quantitative Pipet- ting. Method Success Rate (%)↑Progress Rate (%)↑ Make Sandwich Pick Industrial Parts Sort Tubes MoF Experiments Quantitative Pipetting π 0.5 73.353.186.750.246.7 Fast-WAM40.050.666.710.718.7 InternW073.382.788.968.465.3 Table 6: Subtask success rates and overall task progress rates on MoF Experiments. #Subtaskπ 0.5 Fast-WAMInternW0 1Pick up funnel from rack100.0100.0100.0 2Insert funnel into flask100.046.7100.0 3Pick up graduated cylinder100.013.393.3 4Pour liquid into flask via funnel73.30.080.0 5Place graduated cylinder back73.30.080.0 6Hold flask steady on platform73.30.080.0 7Pick up funnel from flask73.30.080.0 8Insert funnel into rack hole46.70.066.7 9Return to initial position46.70.066.7 10Place flask onto stirrer33.30.066.7 11Pick up stopper from rack6.70.066.7 12Insert stopper into flask6.70.066.7 13Press left button of stirrer6.70.026.7 14Press right button of stirrer6.70.026.7 15Return to initial position6.70.026.7 Progress rate50.210.768.4 errors where a part is placed into the wrong box, which we attribute to the stronger part-level discrimination provided by the visual pretraining of the WAM backbone. On Sort Tubes, both pretrained models follow the instruction reliably (88.9% vs. 86.7%), while Fast-WAM (66.7%) often closes the gripper above the thin tubes, as its visual representation trained from scratch fails to localize the correct grasp point. The benefit of prediction is clearest on MoF Experiments, where InternW0 improves the progress rate from 50.2% to 68.4%. The subtask breakdown in Table 6 reveals two failure modes of π 0.5 . First, at the pouring step (subtask 4), π 0.5 misaligns the graduated cylinder with the funnel and spills liquid, whereas InternW0 pours more steadily. Second, π 0.5 degrades sharply from subtask 8 onward (46.7%→ 33.3%→ 6.7%), where the small transparent funnel, flask, and stopper must be precisely grasped and inserted, while InternW0 sustains 66.7% through subtask 12 and only drops at the final button-pressing stage (26.7%), which demands millimeter-level precision. Fast-WAM performs poorly on this task: without pretraining, it suffers from severe state confusion across the visually similar stages, repeatedly re-executing or skipping steps, and never progresses beyond the first three subtasks. Its results, which trail InternW0 by a large margin on every task, confirm that the gains of InternW0 stem from pretraining and the predictive objective rather than from the architecture alone. To evaluate whether force-aware active hybrid force–position interaction enables reliable dexter- ous manipulation and compliant tool use, we use the quantitative pipetting task as a challenging testbed, with the results reported in Table 7. It requires a 20-DoF dexterous hand to achieve stable 16 Table 7: Subtask success rates and overall task progress rates on Quantitative Pipetting task. #Subtaskπ 0.5 Fast-WAMInternW0 1Pickup & reorientation93.346.786.7 2Tip attachment40.013.366.7 3Liquid aspiration33.313.366.7 4Liquid dispensing33.313.360.0 5Tip ejection & return33.36.746.7 Progress rate46.718.765.3 grasping and precise in-hand reorientation of the pipette through coordinated multi-finger move- ments. Moreover, tip attachment and tip ejection and return require human-like active hybrid force–position interaction, in which the robot regulates contact forces while maintaining the de- sired tool pose during contact. Among the evaluated methods, InternW0 achieves the strongest overall performance, obtaining the highest success rates on four of the five subtasks and the high- est overall progress rate of 65.3%. Its advantages are particularly pronounced on the contact- and force-sensitive subtasks, including tip attachment, liquid aspiration, liquid dispensing, and tip ejec- tion. Although π 0.5 performs better on pickup reorientation, InternW0 substantially outperforms the other methods in the subsequent manipulation stages. These results suggest that reliable quan- titative pipetting benefits from the combination of high-DoF dexterous manipulation and active hybrid force–position interaction. 휋 !.# Fast-WA MInternW0 pipetting MoF Figure 6: Qualitative comparison on the pouring step of MoF Experiments (top) and the tip- attachment step of Quantitative Pipetting (bottom). InternW0 aligns the tool before executing the contact-rich action, whereas the baselines act with visible misalignment. Failure Case Analysis. Figure 6 visualizes the two most representative failure modes. In the pour- ing step of MoF Experiments, InternW0 keeps the mouth of the graduated cylinder centered over the funnel while tilting, so the liquid enters the flask cleanly; π 0.5 starts tilting slightly off-center and part of the liquid runs down the outside of the flask, while Fast-WAM tilts the cylinder far from the funnel and pours directly onto the platform. In the tip-attachment step of Quantitative Pipetting, through force-aware active hybrid position–force interaction, InternW0 accommodates an initially misaligned pipette by adapting its pose online in response to contact forces and yielding compli- antly to external forces, thereby aligning the tip with the target opening and successfully lifting the pipette. In contrast, π 0.5 continues to apply downward force during insertion until the pipette slips out of its grasp, whereas Fast-WAM exhibits a substantial positional error at the insertion site, bending and deforming the pipette body. 17 6 Efficiency For a physical world model, predictive accuracy alone is insufficient: predictions must also become available in time to influence ongoing interaction. This makes timeliness a model-level requirement, rather than merely a systems consideration. InternW0 addresses this requirement by separating the computational timescales of prediction and control. A high-capacity Video-DiT maintains longer-horizon predictive context at a slower timescale, while a lightweight Action-DiT updates actions more frequently from the latest observa- tion and available predictive context. Together with the observation-conditioned context-routing interface introduced in Section 2.1, this design keeps expensive future prediction off the latency- critical action path without removing its influence on control. 6.1 Critical-Path Latency and Policy Update Rate 501002005001k2k Latency per update (ms) ↓ Motus Fast-WAM InternW0 (a) Latency 1866.10 190.00 60.73 051015 Update rate (Hz) ↑ Speedup (b) Throughput 0.540.10× 5.261.00× 16.473.13× One action-generation update · update rate = 1 / latency Figure 7: Critical-path inference efficiency. We evaluate inference efficiency using the same action-generation protocol across all com- pared methods.For InternW0, we dis- tinguish the critical action path from the slower predictive path. The critical-path la- tency includes current-observation encoding, observation-conditioned context routing, and Action-DiT denoising using the latest available predictive context. Video-plan generation is scheduled asynchronously and is therefore not included in this critical-path measurement. Conceptually, under asynchronous execution, the rate of closed-loop policy updates is deter- mined by the latency of the fast action path, whereas Video-DiT inference belongs to the slower predictive path and does not directly block each action update. All latency measurements are performed on the same RTX 5090D hardware under the same numerical precision setting. Results and Analysis. As shown in Figure 7, InternW0 reduces the critical-path action-generation latency to 60.73 ms, corresponding to a maximum model-side policy update rate of 16.47 Hz. This is a 3.13× speedup over Fast-WAM and substantially reduces latency relative to Motus. The practical significance of this reduction is not only higher throughput. A shorter critical path decreases the interval between receiving a new physical observation and producing an action up- date conditioned on that observation. In closed-loop interaction, this directly affects how quickly the policy can respond when the actual environment begins to deviate from its predicted evolution. 6.2 Asynchronous Prediction and Closed-Loop Execution Following the asynchronous planner–executor principle introduced in AHA-WAM (Cai et al., 2026a), InternW0 decouples low-frequency world prediction from high-frequency action genera- tion. In InternW0, this principle is instantiated through the pretrained Video-DiT and Action-DiT together with the observation-conditioned video-context interface described in Section 2.1. The Video-DiT periodically produces a longer-horizon predictive plan and exposes its layerwise K/V representations as reusable predictive context. Between video-plan updates, the Action-DiT does not rerun the full video predictor. Instead, each action update reads the latest available video context and adapts it using the newly observed visual state through the chunk K/V editor before generating the next action chunk. When a refreshed video plan becomes available, it replaces the previous predictive context for subsequent action updates. 18 This organization moves the expensive video-generation process away from the latency-critical control path. The Video-DiT determines how often long-horizon predictive context can be re- freshed, whereas the Action-DiT determines how quickly new observations can influence control. The two frequencies serve different purposes: prediction maintains longer-horizon physical con- text, while the fast action path maintains responsiveness to the realized state of the environment. Importantly, asynchronous execution does not imply that the predictive plan is treated as fixed between updates. Observation-conditioned K/V routing continuously changes how the cached plan is exposed to each action chunk. The system can reuse expensive predictive computation while remaining sensitive to discrepancies between the predicted future and the state that unfolds. Critical-path acceleration. Beyond asynchronous scheduling, the deployment pipeline reduces the cost of the fast action path through complementary implementation optimizations. Graph-level acceleration is applied to the Action-DiT and context-processing modules to reduce launch and ex- ecution overhead, while selective compilation is used for the Video-DiT prefill path. Computations invariant across denoising steps are hoisted outside the inner sampling loop, and redundant tensor transfers and repeated state traversal are removed. These optimizations do not alter the model architecture or inference precision. Instead, they reduce systems overhead around the same video-context routing and action-generation computation, fur- ther separating the latency of responsive control from the cost of long-horizon prediction. Efficiency as a property of physical intelligence. Taken together, asynchronous scheduling, reusable predictive context, observation-conditioned routing, and critical-path acceleration allow InternW0 to retain an expressive video predictor without making its full generation latency the bottleneck of every action update. The resulting interaction loop continuously observes, updates, and acts while predictive computation proceeds at its own timescale. This closes the loop with the design motivation of the InternW series: a useful physical world model should not only predict accurately and produce decision-relevant representations, but should make that information available before it becomes stale. This requirement is particularly important in dynamic and contact-rich interactions, where the value of a prediction depends not only on what it predicts, but also on whether the agent can react to the physical evidence that arrives afterward. 7 Summary and Future Work In this report, we presented InternW0, which establishes a foundation for physical world modeling by combining visual and force perception with language, proprioception, prediction, and action. Its duplex interaction, enabled by asynchronous, multi-frequency processing supports continuous sensing, prediction, action, and correction, and makes InternW0 work much efficiently. On main- stream simulation benchmarks of robot manipulation, InternW0 has shown its highly competitive performance. Furthermore, we apply InternW0 to real-world scientific experiments that require dexterous, precise, and safe manipulation. Future work will extend this foundation in the following directions. From Observation to Intervention Understanding. InternW0 integrates force feedback with vi- sion and body state, helping robots determine whether objects are grasped securely, contact is es- tablished, or an operation has reached the desired state. Future models should incorporate addi- tional physical modalities, including tactile sensing, torque, proprioception, and contact signals. They should learn not only what happens, but how interventions change subsequent states, dis- tinguishing passive evolution from action-induced effects and updating hidden-state estimates as new feedback arrives. This capability is essential for scientific procedures whose outcomes depend on timing, force, temperature, mixing, and operation order (Schölkopf et al., 2021). 19 Broader Modalities and Specialized Experts. Future InternW models should incorporate ad- ditional modalities, including depth, 3D geometry, audio, and other task-specific observations, through modality-specific encoders and decoders. These experts should exchange information through a shared physical representation while preserving modality-specific temporal and spatial structure. Realizing this extension requires multimodal data with reliable timestamps and spatial calibration. The data pipeline should filter sensor failures, corrupted intervals, implausible tra- jectories, and annotation errors, normalize units and coordinate conventions, and record validity masks for missing channels. Prediction Across Scales and Horizons. Physical processes combine fast local dynamics with slower, multi-stage changes. Future models should connect these scales through hierarchical representations, memory, and adaptive temporal resolution. InternW0’s asynchronous, multi- frequency design provides a basis for long-horizon prediction and replanning while maintaining high-frequency control. Reliability will require handling partial observability, accumulated error, and intervention-dependent dynamics through uncertainty-calibrated prediction, continual state correction, and evaluation across multiple timescales (Kim et al., 2019). One Physical Foundation for Many Environments. The InternW world model series should transfer across robots, instruments, laboratories, and physical settings while allowing rapid spe- cialization. Building on InternW0’s heterogeneous data pipeline and domain-specific action inter- faces, future models should combine shared physical representations with adaptable interfaces for sensors, actuators, embodiments, and action spaces. Parameter-efficient adaptation and limited demonstrations should enable specialization to new robots and protocols without relearning basic physical knowledge. Evaluation should measure transfer performance and the data and interaction required for adaptation. Closed-Loop Operation in the Real World. A physical world model must predict, act, observe the outcome, and update its next decision. InternW0’s duplex and asynchronous architecture pro- vides a foundation for background planning alongside responsive action generation. Real-world prediction errors can reveal missing data, calibration problems, and environment-specific failure modes. Future systems should use these signals for continual improvement while preserving ex- isting capabilities (Kirkpatrick et al., 2017). Evaluation should include sustained task success, dis- turbance recovery, transfer, latency, and reliable action correction, especially in multi-stage wet-lab procedures. Lightweight World Models for Embedded Intelligence. Future InternW world models should target approximately 4B parameters and ultimately fewer than 1B, enabling deployment on cus- tomized small devices and embedded physical agents so as to work in various physical envi- ronments. Distillation, quantization, compact latent representations, efficient action heads, and hardware-aware inference will be important. Compression should preserve multimodal state es- timation, intervention-sensitive prediction, high-frequency action generation, and recovery from unexpected contact, while accounting for memory, energy, and perception costs. Combined with rapid specialization, lightweight InternW models could serve as local world models for diverse robots and scientific instruments. Contributors The following are the contributors of this article: Jisong Cai, Yao Mu, Ganlin Yang, Zhe Cao, Zhangzheng Tu, Xing Gao, Kailin Li, Xinyu Zhan, Lixin Yang, Yangkun Zhu, Haoxiang Ma, Ming Zhou, Qiaojun Yu, Yufei Xue, Liqun He, Yifei Yao, Yifan Zhu, Long Ling, Bingqi Jiang, Haoyu Guo, Xueyue Zhu, Bowen Zhou, Bin Zhao, Tianfan Xue, Chunhua Shen, and Weinan Zhang. 20 References Lei Bai, Jiaqi Cao, Chiyu Chen, Guanzhou Chen, Kai Chen, Guangran Cheng, Erfei Cui, Xuanlang Dai, Shengyuan Ding, Shangheng Du, et al. Intern-S2-Preview: Scientific agentic foundation model. arXiv preprint arXiv:2608.13505, 2026. doi: 10.48550/arXiv.2608.13505. URL https:// arxiv.org/abs/2608.13505. Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, et al. Motus: A unified latent action world model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 35101– 35113, 2026. Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fu- sai, Lachy Groom, Karol Hausman, Brian Ichter, et al. π 0 : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. Qingwen Bu, Jisong Cai, Li Chen, Xiuqi Cui, Yan Ding, Siyuan Feng, Shenyuan Gao, Xindong He, Xu Huang, Shu Jiang, et al. Agibot world colosseo: A large-scale manipulation platform for scalable and intelligent embodied systems. arXiv preprint arXiv:2503.06669, 2025. Jisong Cai, Long Ling, Shiwei Chu, Zhongshan Liu, Jiayue Kang, Zhixuan Liang, Wenjie Xu, Yinan Mao, Weinan Zhang, Xiaokang Yang, et al. Aha-wam: Asynchronous horizon-adaptive world- action modeling with observation-guided context routing. arXiv preprint arXiv:2606.09811, 2026a. Rui Cai, Jun Guo, Xinze He, Piaopiao Jin, Jie Li, Bingxuan Lin, Futeng Liu, Wei Liu, Fei Ma, Kun Ma, et al. Xiaomi-robotics-0: An open-sourced vision-language-action model with real-time exe- cution. arXiv preprint arXiv:2602.12684, 2026b. Ronghan Chen, Yandan Yang, Zuojin Tang, Dongjie Huo, Tong Lin, Haoning Wu, Haoyun Liu, Yuzhi Chen, Lulu Zheng, Botai Yuan, et al. Abot-m0. 5: Unified mobility-and-manipulation world action model. arXiv preprint arXiv:2607.00678, 2026a. Tianxing Chen, Zanxin Chen, Baijun Chen, Zijian Cai, Yibin Liu, Zixuan Li, Qiwei Liang, Xianliang Lin, Yiheng Ge, Zhenyu Gu, et al. Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation. arXiv preprint arXiv:2506.18088, 2025a. Tianxing Chen, Yue Chen, Zixuan Li, Junyuan Tang, Kailun Su, Haoran Lu, Weijie Wan, Baijun Chen, Songling Liu, Haowen Yan, et al. Robodojo: A unified sim-and-real benchmark for com- prehensive evaluation of generalist robot manipulation policies. arXiv preprint arXiv:2607.04434, 2026b. Xinyi Chen, Yilun Chen, Yanwei Fu, Ning Gao, Jiaya Jia, Weiyang Jin, Hao Li, Yao Mu, Jiangmiao Pang, Yu Qiao, et al. Internvla-m1: A spatially guided vision-language-action framework for generalist robot policy. arXiv preprint arXiv:2510.13778, 2025b. Xinyuan Chen, Haoyu Guo, Shi Guo, Bingqi Jiang, Chunhua Shen, Xing Shen, Tianfan Xue, Yufei Xue, Mulin Yu, Weinan Zhang, et al. A definition and roadmap for world models. arXiv preprint arXiv:2607.06401, 2026c. Yizhou Chen, Mark Van der Merwe, Andrea Sipos, and Nima Fazeli. Visuo-tactile transformers for manipulation. In Proceedings of The 6th Conference on Robot Learning, volume 205 of Proceedings of Machine Learning Research, pages 2026–2040. PMLR, 2023. URL https://proceedings.mlr.press/ v205/chen23d.html. Jun Guo, Qiwei Li, Peiyan Li, Zilong Chen, Nan Sun, Yifei Su, Heyun Wang, Yuan Zhang, Xinghang Li, and Huaping Liu. Unified 4d world action modeling from video priors with asynchronous denoising. arXiv preprint arXiv:2604.26694, 2026. 21 David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. URL https://arxiv.org/abs/1803.10122. Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In Proceedings of the 36th Interna- tional Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2555–2565. PMLR, 2019. URL https://proceedings.mlr.press/v97/hafner19a.html. Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse control tasks through world models. Nature, 640:647–653, 2025. doi: 10.1038/s41586-025-08744-2. URL https://w.nature.com/articles/s41586-025-08744-2. Nicklas Hansen, Hao Su, and Xiaolong Wang. TD-MPC2: Scalable, robust world models for con- tinuous control. In The Twelfth International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2310.16828. InkStone, Shanghai AI Laboratory. InkStone: Ai research workspace. https://discovery-home. intern-ai.org.cn/home/. Accessed: September 22, 2026. Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al. π 0.5 : a vision-language-action model with open-world generalization. arXiv preprint arXiv:2504.16054, 2025. Tao Jiang, Tianyuan Yuan, Yicheng Liu, Chenhao Lu, Jianning Cui, Xiao Liu, Shuiqi Cheng, Jiyang Gao, Huazhe Xu, and Hang Zhao. Galaxea open-world dataset and g0 dual-system vla model. arXiv preprint arXiv:2509.00576, 2025. URL https://arxiv.org/abs/2509.00576. Leslie Pack Kaelbling, Michael L. Littman, and Anthony R. Cassandra.Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101(1–2):99–134, 1998. doi: 10.1016/S0004-3702(98)00023-X. URL https://w.sciencedirect.com/science/article/pii/ S000437029800023X. Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-tuning vision-language-action models: Opti- mizing speed and success. arXiv preprint arXiv:2502.19645, 2025. Taesup Kim, Sungjin Ahn, and Yoshua Bengio. Variational temporal abstraction. Advances in neural information processing systems, 32, 2019. James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcom- ing catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114 (13):3521–3526, 2017. Jason Lee, Jiafei Duan, Haoquan Fang, Yuquan Deng, Shuo Liu, Boyang Li, Bohan Fang, Jieyu Zhang, Yi Ru Wang, Sangho Lee, Winson Han, Wilbert Pumacay, Angelica Wu, Rose Hendrix, Karen Farley, Eli VanderBilt, Ali Farhadi, Dieter Fox, and Ranjay Krishna. Molmoact: Action reasoning models that can reason in space, 2025. URL https://arxiv.org/abs/2508.07917. Michelle A. Lee, Yuke Zhu, Krishnan Srinivasan, Parth Shah, Silvio Savarese, Li Fei-Fei, Animesh Garg, and Jeannette Bohg. Making sense of vision and touch: Self-supervised learning of mul- timodal representations for contact-rich tasks. In 2019 International Conference on Robotics and Automation (ICRA), 2019. URL https://arxiv.org/abs/1810.10191. Fuhao Li, Wenxuan Song, Han Zhao, Jingbo Wang, Pengxiang Ding, Donglin Wang, Long Zeng, and Haoang Li. Spatial forcing: Implicit spatial representation alignment for vision-language- action model. In International Conference on Learning Representations, volume 2026, pages 132324– 132345, 2026a. Lin Li, Qihang Zhang, Yiming Luo, Shuai Yang, Ruilin Wang, Fei Han, Mingrui Yu, Zelin Gao, Nan Xue, Xing Zhu, et al. Causal world modeling for robot control. arXiv preprint arXiv:2601.21998, 2026b. 22 Yang Li, Zhaxizhuoma Zhaxizhuoma, Hongru Jiang, Junjie Xia, Hongquan Zhang, Jinda Du, Yun- song Zhou, Jia Zeng, Ce Hao, Jieji Ren, et al. Forcevla2: Unleashing hybrid force-position control with force awareness for contact-rich manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8911–8920, 2026c. Weixin Liang, LILI YU, Liang Luo, Srini Iyer, Ning Dong, Chunting Zhou, Gargi Ghosh, Mike Lewis, Wen tau Yih, Luke Zettlemoyer, and Xi Victoria Lin. Mixture-of-transformers: A sparse and scalable architecture for multi-modal foundation models. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=Nu6N69i8SB. Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. Advances in Neural Information Processing Systems, 36:44776–44791, 2023. Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. Kevin M. Lynch and Frank C. Park.Modern Robotics:Mechanics, Planning, and Con- trol.Cambridge University Press, 2017.URL https://modernrobotics.northwestern.edu/ nu-gm-book-resource/3-4-wrenches/. Philipp Moritz, Robert Nishihara, Stephanie Wang, Alexey Tumanov, Richard Liaw, Eric Liang, Melih Elibol, Zongheng Yang, William Paul, Michael I Jordan, et al. Ray: A distributed frame- work for emergingAI applications. In 13th USENIX symposium on operating systems design and implementation (OSDI 18), pages 561–577, 2018. NVIDIA, Johan Bjorck, Nikita Cherniadev Fernando Castañeda, Xingye Da, Runyu Ding, Linxi "Jim" Fan, Yu Fang, Dieter Fox, Fengyuan Hu, Spencer Huang, Joel Jang, Zhenyu Jiang, Jan Kautz, Kaushil Kundalia, Lawrence Lao, Zhiqi Li, Zongyu Lin, Kevin Lin, Guilin Liu, Edith Llon- top, Loic Magne, Ajay Mandlekar, Avnish Narayan, Soroush Nasiriany, Scott Reed, You Liang Tan, Guanzhi Wang, Zu Wang, Jing Wang, Qi Wang, Jiannan Xiang, Yuqi Xie, Yinzhen Xu, Zhen- jia Xu, Seonghyeon Ye, Zhiding Yu, Ao Zhang, Hao Zhang, Yizhou Zhao, Ruijie Zheng, and Yuke Zhu. GR00T N1: An open foundation model for generalist humanoid robots. In ArXiv Preprint, March 2025. M. H. Raibert and J. J. Craig. Hybrid position/force control of manipulators. Journal of Dynamic Systems, Measurement, and Control, 103(2):126–133, 1981. doi: 10.1115/1.3139652. URL https: //doi.org/10.1115/1.3139652. Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109(5):612–634, 2021. Oriane Siméoni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seung Eun Yi, Michael Ramamonjisoa, Francisco Massa, Daniel HAZIZA, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julien Mairal, Herve Jegou, Patrick Labatut, and Piotr Bojanowski. DINOv3. Transactions on Machine Learning Re- search, 2026. ISSN 2835-8856. URL https://openreview.net/forum?id=2NlGyqNjns. Featured Certification. Rui Song, Yang Xu, Zhangzheng Tu, Hongzhou Wang, Hongchen Tan, and Huchuan Lu. Visual- textual information-driven tactile data generation method. IEEE Transactions on Image Processing, 2026. GigaBrain Team, Angen Ye, Axiang Sun, Can Jin, Chenxi Cheng, Chong Shi, Dengke Shang, Dingqian Zhang, Guan Huang, Guangqiang Wang, et al. Gigabrain-0.7: Scaling embodied foundation models to emergent capabilities with a three-system architecture. arXiv preprint arXiv:2608.15875, 2026. 23 Yang Tian, Yuyin Yang, Yiman Xie, Zetao Cai, Xu Shi, Ning Gao, Hangxu Liu, Xuekun Jiang, Zherui Qiu, Feng Yuan, et al. Interndata-a1: Pioneering high-fidelity synthetic data for pre-training gen- eralist policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 976–985, 2026. Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. Siyin Wang, Junhao Shi, Zhaoyang Fu, Xinzhe He, Feihong Liu, Chenchen Yang, Yikang Zhou, Zhaoye Fei, Jingjing Gong, Jinlan Fu, et al. World action models: The next frontier in embodied ai. arXiv preprint arXiv:2605.12090, 2026a. Yuran Wang, Siqiao Huang, Mingleyang Li, Chenhao Zhang, Jiaqi Liang, Weiyang Jin, Yue Chen, Xuemin Chi, Donghao Zhou, Qize Yu, et al. Openwam: An open, modular exploration towards systematic world-action model pretraining. arXiv preprint arXiv:2609.07398, 2026b. Shihan Wu, Xuecheng Liu, Shaoxuan Xie, Pengwei Wang, Xinghang Li, Bowen Yang, Zhe Li, Kai Zhu, Hongyu Wu, Yiheng Liu, et al. Robocoin: An open-sourced bimanual robotic data collection for integrated manipulation. arXiv preprint arXiv:2511.17441, 2025. Lishan Yang, Wenxuan Song, Xi Wang, Pingyue Sheng, Zheng Fang, Ziyang Zhou, Junjie He, Haodong Yan, Jiayi Chen, Nan Sun, et al. 4d-wam: Infusing spatiotemporal awareness into world action models through trajectory fields. arXiv preprint arXiv:2608.08023, 2026a. Yandan Yang, Shuang Zeng, Tong Lin, Xinyuan Chang, Dekang Qi, Junjin Xiao, Haoyun Liu, Rong- han Chen, Yuzhi Chen, Dongjie Huo, et al. Abot-m0: Vla foundation model for robotic manipu- lation with action manifold learning. arXiv preprint arXiv:2602.11236, 2026b. Seonghyeon Ye, Yunhao Ge, Kaiyuan Zheng, Shenyuan Gao, Sihyun Yu, George Kurian, Suneel Indupuru, You Liang Tan, Chuning Zhu, Jiannan Xiang, Ayaan Malik, Kyungmin Lee, William Liang, Nadun Ranawaka, Jiasheng Gu, Yinzhen Xu, Guanzhi Wang, Fengyuan Hu, Avnish Narayan, Johan Bjorck, Jing Wang, Gwanghyun Kim, Dantong Niu, Ruijie Zheng, Yuqi Xie, Jimmy Wu, Qi Wang, Ryan Julian, Danfei Xu, Yilun Du, Yevgen Chebotar, Scott Reed, Jan Kautz, Yuke Zhu, Linxi Fan, and Joel Jang. World action models are zero-shot policies. arXiv preprint arXiv:2602.15922, 2026. URL https://arxiv.org/abs/2602.15922. Jiawen Yu, Hairuo Liu, Qiaojun Yu, Jieji Ren, Ce Hao, Haitong Ding, Guangyu Huang, Guofan Huang, Yan Song, Panpan Cai, et al. Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation. Advances in Neural Information Processing Systems, 38:93409–93439, 2026. Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-WAM: Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026. URL https://arxiv. org/abs/2603.16666. Jinglei Zhang, Jiankang Deng, Chao Ma, and Rolandos Alexandros Potamias. Hawor: World-space hand motion reconstruction from egocentric videos. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1805–1815. IEEE, 2025. Jinliang Zheng, Jianxiong Li, Zhihao Wang, Dongxiu Liu, Xirui Kang, Yuchun Feng, Yinan Zheng, Jiayin Zou, Yilun Chen, Jia Zeng, et al. X-vla: Soft-prompted transformer as scalable cross- embodiment vision-language-action model. In International Conference on Learning Representa- tions, volume 2026, pages 60580–60606, 2026. 24