Paper deep dive

ARC: A Reasoning Recipe for Robot Foundation Models

Gokul Puthumanaillam, Tao Sun, Elie Aljalbout, Moritz Reuss, Zhaoshuo Li, Fabio Ramos, Ankit Goyal, Jenai Xuning Yang

Year: 2026Venue: arXiv preprintArea: cs.ROType: PreprintEmbeddings: 134

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%

Last extracted: 10/11/2026, 2:15:31 AM

Summary

The paper introduces ARC, a reasoning recipe for Robot Foundation Models (RFMs) that improves zero-shot task performance without additional robot demonstrations or foundation-scale training. ARC consists of action-grounded causal reasoning traces, a scalable automatic labeling pipeline (ARC-Trace) to generate these traces from existing data (e.g., DROID dataset), and a fine-tuning strategy for adapting models like pi_0.5 and Cosmos3-Nano-Policy. The method establishes new state-of-the-art results on benchmarks like RoboLab-120 and MolmoSpaces, with significant gains in real-world robot task success.

Entities (10)

ARC · method · 100%$\\pi_{0.5}$ · model · 95%action-grounded causal traces · concept · 95%ARC-Trace · method · 95%Cosmos3-Nano-Policy · model · 95%DROID · dataset · 95%MolmoSpaces · benchmark · 95%RoboLab-120 · benchmark · 95%RoboLab-Reasoning-50 · benchmark · 95%Robot Foundation Models · model-class · 95%

Relation Signals (9)

ARC → achievesstateofthearton → MolmoSpaces

confidence 95% · The adapted models establish a new state of the art on RoboLab-120 and MolmoSpaces...

ARC → achievesstateofthearton → RoboLab-120

confidence 95% · The adapted models establish a new state of the art on RoboLab-120...

ARC → appliedto → Cosmos3-Nano-Policy

confidence 95% · We evaluate this recipe on two representative state-of-the-art RFMs: ... Cosmos3-Nano-Policy

ARC → appliedto → $\\pi_{0.5}$

confidence 95% · We evaluate this recipe on two representative state-of-the-art RFMs: ... pi_0.5

ARC-Trace → generates → action-grounded causal traces

confidence 95% · ARC-Trace automatically relabels DROID with action-grounded causal traces...

ARC → improves → $\\pi_{0.5}$

confidence 95% · On real robots, ARC improves pi_0.5's task success by 82.2 percentage points.

ARC → improves → Robot Foundation Models

confidence 95% · ARC... can substantially improve the zero-shot task performance of existing state-of-the-art RFMs.

ARC-Trace → processes → DROID

confidence 95% · construct ARC-Trace-DROID from DROID without collecting new robot data.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. We show that there exists an effective and efficient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs. We refer to this recipe as ARC. It consists of three key ingredients: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. First, we find that effective reasoning traces should be grounded in the robot's next action and explain its causal structure: why the action is appropriate and what effect it should produce. Second, we show that these traces can be generated automatically from existing demonstrations, enabling us to construct ARC-Trace-DROID from DROID without collecting new robot data. Third, we show how state-of-the-art VLAs such as $\pi_{0.5}$ and WAMs such as Cosmos3-Nano-Policy can learn to use these traces for control, with fine-tuning and inference tailored to each model's architecture and capabilities. Using ARC, we obtain gains in zero-shot RFM performance that, to our knowledge, are unprecedented without additional robot demonstrations or foundation-scale training. The adapted models establish a new state of the art on RoboLab-120 and MolmoSpaces, with gains of up to 50 percentage points on RoboLab-Reasoning-50. On real robots, ARC improves $\pi_{0.5}$'s task success by 82.2 percentage points. Project website: this https URL

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

130,904 characters extracted from source content.

Expand or collapse full text

ARC: A Reasoning Recipe for Robot Foundation Models Gokul Puthumanaillam 1,3 ,Tao Sun 2,3 ,Elie Aljalbout 3 ,Moritz Reuss 3 ,Zhaoshuo Li 3 ,Fabio Ramos 3,4 , Ankit Goyal 3,5,† ,Jenai Xuning Yang 3,† 1 University of Illinois Urbana-Champaign, 2 Stanford University, 3 NVIDIA, 4 University of Sydney, 5 Proception † Equal advising, listed alphabetically The prevailing approach to improving robot foundation models (RFMs) relies on larger models, more robot demonstrations, and costly training at scale. We show that there exists an effective and eﬀicient complementary approach: the right reasoning recipe can substantially improve the zero-shot task performance of existing state-of-the-art RFMs. We refer to this recipe as ARC. It consists of three key ingredients: a reasoning trace, a scalable automatic labeling pipeline, and a strategy for adapting pretrained RFMs to use these traces for control. First, we find that effective reasoning traces should be grounded in the robot’s next action and explain its causal structure: why the action is appropriate and what effect it should produce. Second, we show that these traces can be generated automatically from existing demonstrations, enabling us to constructArc-Trace-DROID from DROID without collecting new robot data. Third, we show how state-of-the-art VLAs such asπ 0.5 and WAMs such as Cosmos3-Nano-Policy can learn to use these traces for control, with fine-tuning and inference tailored to each model’s architecture and capabilities. UsingArc, we obtain gains in zero-shot RFM performance that, to our knowledge, are unprecedented without additional robot demonstrations or foundation-scale training. The adapted models establish a new state of the art on RoboLab-120 and MolmoSpaces, with gains of up to 50 percentage points on RoboLab-Reasoning-50. On real robots, Arcimprovesπ 0.5 ’s task success by 82.2 percentage points. Project website:https://arc-robot-reasoning.github.io/ Figure 1Overview ofArc:Arc-Trace automatically relabels DROID with action-grounded causal traces to produce Arc-Trace-DROID. We show how to fine-tune VLAs and WAMs with this supervision, achieving benchmark-leading performance without additional robot demonstrations. 1 arXiv:2610.12386v1 [cs.RO] 8 Oct 2026 1Introduction Robot foundation models (RFMs), trained by imitation on large demonstration datasets, have become broadly capable at manipulation (Kim et al.,2024;NVIDIA,2026;Black et al.,2025b;a). Yet the dominant recipe for making them better remains brute-force scaling, with more robot data, larger models, and another round of foundation-scale training. This makes every gain in task success expensive to obtain. How can we improve the performance of existing RFMs without modifying their architectures, adding new robot data, or training at the scale of foundation models? In this work, we study this problem. Specifically, we find that adding or refining the reasoning capabilities of RFMs is an effective and eﬀicient strategy for improving them. With the right recipe, reasoning can substantially improve the performance of existing state-of-the-art models such asπ 0.5 (Black et al.,2025a) and Cosmos3-Nano-Policy (NVIDIA,2026) (Fig.1). To our knowledge, this is the first work to achieve such substantial gains in zero-shot RFM performance through any strategy that does not require additional robot demonstrations or foundation-scale training. We call this reasoning recipe Arc. InArc, we bring together three pieces: the right reasoning content, a scalable source of supervision, and a training procedure that adapts the model to use that reasoning for control. First, what should the reasoning contain? We find that reasoning for control should be causal, explaining why an action is appropriate in the current situation and what effect it should produce, as well as grounded in action, specifying the concrete step the robot should take to bring about that effect. We introduce action-groundedcausaltraces that connect what is true now to what the robot should do next and what should become true as a result. Second, how do we obtain this supervision at scale? Robot demonstrations record observations and actions but typically omit the reasoning behind individual decisions. We present an approach for automatically labeling demonstrations with causal traces aligned to action chunks. We use it to constructArc-Trace-DROID by relabeling DROID (Khazatsky et al.,2024). And third, how do we adapt models to use this reasoning for control? We show that an RFM can be fine-tuned to condition its behavior on reasoning traces while retaining its existing control capabilities. We use the same reasoning traces and training data across models and show how to tailor the fine-tuning and inference procedures to each model’s architecture and capabilities. We evaluate this recipe on two representative state-of-the-art RFMs: Cosmos3-Nano-Policy (NVIDIA,2026), a world action model (WAM), andπ 0.5 (Black et al.,2025a), a vision-language-action model (VLA). We find thatArcimproves both RFMs across all three settings on RoboLab-120, with the adapted models ranking first and second in each setting. On MolmoSpaces, task success improves by 27.2 and 18.6 percentage points (p.p.) forπ 0.5 and Cosmos3-Nano-Policy. These results establish a new state of the art on RoboLab- 120 ( Yang et al.,2026) and MolmoSpaces (Kim et al.,2026). On the reasoning-focused RoboLab-Reasoning- 50,Arcimproves task success by 46.8 p.p. forπ 0.5 and 50.0 p.p. for Cosmos3-Nano-Policy. Further, these improvements carry over to real world execution, where theArcgains 82.2 p.p. on reasoning and long- horizon tasks. Across these evaluations,Arc-trained models more reliably follow underspecified instructions, infer missing steps, complete long-horizon tasks, adapt to changes, and recover from mistakes. We further quantify how different components of the recipe contribute to performance. Our results show that adding intermediate language without the right structure provides limited improvement; larger gains arise from combining causal reasoning content with a training procedure that grounds it in action. As a side effect, the recipe also improves training and inference eﬀiciency. Cosmos3-Nano-Policy reaches baseline success with 4.3×fewer training updates, whileπ 0.5 and Cosmos3-Nano-Policy match their base policies at inference with 10×and2×fewer denoising steps. Contributions:We introduceArc, a recipe for instilling reasoning into pretrained RFMs. We make four contributions. (i) We identify action-grounded causal traces as an intermediate representation for robot reasoning. (i) We introduce ARC-Trace, a labeling pipeline that generates action-grounded causal traces aligned to action chunks and use it to createArc-Trace-DROID by relabeling DROID. (i) We develop a strategy for adapting pretrained RFMs to use reasoning for control and show how to tailor the fine-tuning and inference procedures to each model’s architecture. (iv) We validateArcon Cosmos3-Nano- Policy andπ 0.5 in simulation and on real-robot benchmarks, demonstrating substantial gains in zero-shot task performance that, to our knowledge, are unprecedented without additional robot demonstrations or foundation-scale training. 2 2Related Work Robotfoundationmodels.General-purpose robot policies have been developed by learning across tasks and embodiments (Reed et al.,2022;Jang et al.,2022;Jiang et al.,2023;Brohan et al.,2023;Bousmalis et al.,2023; O’Neill et al.,2024) and transferring pretrained visual and linguistic knowledge to action prediction (Driess et al.,2023;Zitkovich et al.,2023;Kim et al.,2024;Black et al.,2025b;a). Video-pretrained and world-action models additionally learn how the environment evolves during execution ( Wu et al.,2023;Cheang et al., 2024;Cen et al.,2025;NVIDIA,2026;Ye et al.,2026). Our work builds on these pretrained controllers and investigates how reasoning supervision can improve their performance using robot demonstrations already available. Reasoninginfoundationmodels.Reasoning allows pretrained models to work through intermediate decisions before producing a final answer (Wei et al.,2023;Zhang et al.,2024). Prior work improves this process by considering alternative solutions ( Wang et al.,2023;Yao et al.,2023a), incorporating tools and feedback (Yao et al.,2023b), and learning from intermediate supervision (Zelikman et al.,2024;DeepSeek-AI et al.,2025). In robotics, these ideas have been used in two particularly relevant ways: to determine the steps a robot should execute, and to train a policy to reason about the actions it predicts. The first includes selecting available skills to accomplish an instruction (Ahn et al.,2022;Huang et al.,2023b), generating programs that combine perception and control (Liang et al.,2023;Singh et al.,2022), and revising plans using observations and execution feedback (Huang et al.,2022;Zeng et al.,2022a). The second uses reasoning annotations to teach a policy which information matters for action prediction ( Zawalski et al.,2024;Sun et al.,2026). Both approaches allow robots to benefit from knowledge beyond their demonstrated motions, but require the controller to understand the guidance it receives. Recent work addresses this requirement through richer steering commands and visually grounded hierarchical control (Chen et al.,2026). Reasoningmodalities.There are several ways to express the intermediate decisions needed for robotic control. Spatial representations specify where an interaction should occur through image features, action maps, or voxel representations ( Zeng et al.,2022b;Shridhar et al.,2021;2023), while keypoints and geometric con- straints describe relationships that the robot should establish between objects (Liu et al.,2024;Huang et al., 2023a;2024). Motion-based representations provide guidance through trajectory sketches and visual traces (Gu et al.,2023;Zheng et al.,2025;Lee et al.,2025;Zhong et al.,2026). Predictive representations specify how the scene should evolve, using future predictions (Du et al.,2023;Ko et al.,2024;Zhou et al.,2024;Zhen et al.,2024). These approaches give the controller an explicit spatial or temporal target. Their connection to control depends on the representation. Geometric constraints can be solved through motion optimization ( Huang et al.,2023a;2024), depth and trajectory predictions can be encoded as intermediate tokens (Lee et al.,2025), and visual traces can be presented through the existing image input (Zheng et al.,2025). While these methods demonstrate the value of structured guidance, our interest is in reasoning supervision that can be shared across RFMs without additional architecture changes. Natural language offers this opportunity because it can describe task conditions, spatial relationships, and intended outcomes without committing to a particular set of robot coordinates. It is already supported by the pretrained tokenizers and language components of many robot foundation models (Zitkovich et al.,2023; Kim et al.,2024;Black et al.,2025a), and its annotations can be generated from existing demonstrations (Zawalski et al.,2024). Prior work uses natural language to describe fine-grained motions (Belkhale et al., 2024), convey the current semantic subtask (Black et al.,2025a), and reason about plans, objects, and robot state before predicting actions (Mu et al.,2023). These methods show that language can provide useful intermediate supervision. 3Preliminaries and Problem Statement Robottasksetting.We consider a robot operating in a partially observed environment. A task is specified by an instructiong∈G. At timet, the robot receives an observationo t ∈Oand executes an actiona t ∈A. We denote the interaction history byh t = (o 0 , a 0 , . . . , a t−1 , o t ).To cover models that predict several actions at once, we writea t = (a t , . . . , a t+H−1 )for an action chunk of horizonH. A robot demonstration is a trajectoryτ= (g, o 0 , a 0 , . . . , o T ),and the training dataset isD=τ i N i=1 . We treat a pretrained robot 3 foundation model independently of its internal architecture. At each time step, it defines a conditional action distributionπ θ 0 (a t |h t , g),whereθ 0 denotes the pretrained parameters. LetS(τ, g)∈[0,1]measure whether the resulting trajectory completes the task. The objective of the policy is to maximizeJ(π) =E g,τ∼π [S(τ, g)]. Reasoningasintermediatecomputation.A standard policy predicts actions directly from the task and inter- action history. We definereasoningmore generally as forming an intermediate representationr t ∈Rthat is formed before action prediction and carries information relevant to the current decision:r t ∼q φ (r t | h t , g),a t ∼π θ (a t |h t , g, r t ).The resulting action distribution isπ θ,φ (a t |h t , g) = R R π θ (a t |h t , g, r t )q φ (r t | h t , g)dr t . Problemstatement.Given a pretrained robot foundation modelπ θ 0 and a demonstration datasetDcontaining observations and actions but no reasoning supervision, our goal is to learn a reasoned policyπ θ,φ that improves task success while retaining the model’s existing control abilities. Crucially, this improvement should not require a new model architecture, a new foundation-scale pretraining stage, or the collection of another large robot dataset. 4TheARCFramework We now describe the components ofArc: the reasoning representation (Sec.4.1), theArc-Trace relabeling pipeline (Sec.4.2), and the fine-tuning recipe that connects reasoning to control (Sec.4.3). 4.1Action Grounded Causal Reasoning We representr t in natural language, reusing the model’s pretrained language conditioning. We examine ECoT-style reasoning and semantic subtask supervision, adapting both formats to the sameπ 0.5 backbone and fine-tuning it with each ( Zawalski et al.,2024;Black et al.,2025a). Wrist camera attentionAction-head text attention 0%25%50% Subtask supervision 4% Embodied CoT 6% ARC 41% Subtask supervision [Task]Takethepenoutofthecupandplaceitonthetable.[Subtask]Graspthepeninthecup. Embodied CoT [Task]Takethepenoutofthecupandplaceitonthetable.[Plan]Graspthepen,liftitoutofthecup,move itoverthetable,andrelease.[Subtask]Graspthepen.[Move]Movedowntowardthepen. [Objects]Greenpen;whitecup;papernotes. ARC:Action-grounded causal reasoning traces [State]Therobotseesthewhitecupatcenter-rightwiththegreenpenstandinginitsopening,whitepaper notesnearby,andthegripperpositionedclosearoundthepen.[Cause]Thepenisstillheldbythecup,soit mustbegraspedsecurelybeforeitcanbepulledoutandplacedonthetable.[Effect]Thechoseneffectisa stable,centeredholdontheexposedshaftofthepen.[Action]Thisimpliestherobotshouldpinchthelower visibleportionofthepenshaftwiththegripperfingers,andliftonlyafterastableholdisconfirmed. [Avoid]Therobotmustnotletthegrippercontactthecuprimorwall,orknockthecup.[Completion]The taskisincomplete,sincethepenhasnotyetbeenremovedfromthecupandsetonthetable. Relative attentionlowhigh Figure 2Reasoning traces and action-expert attention. To determine whether these representations actually affect action prediction, we inspect how the action expert distributes attention over image and natural-language tokens in Fig.2. Under both ECoT and subtask su- pervision, the action expert remains strongly focused on the visual observation and assigns little attention to the intermediate language. We treat attention as a diagnostic rather than proof of causal dependence, but the pat- tern exposes a plausible shortcut. Plans, ob- ject lists, motion directions, and subtask la- bels are often predictable from the same im- age and instruction already available to the model. The observed pattern is therefore consistent with shortcut learning, where the model can satisfy the language objective while continuing to predict actions from vision. Similar ground- ing failures have been reported in other robot foundation models, which can continue pursuing the original target even when the instruction specifies a different one ( Fei et al.,2025). This suggests that the missing information is neither a longer scene description nor a finer decomposition of the task. An action is selected because of the change it is expected to produce in the world. Reasoning should make this relation explicit ( Wang et al.,2025). Following this principle, we develop an action-grounded causal representation for robot control. For an action chunka t ,Arc-Trace represents reasoning asr t = [State,Cause,Consequence,Effect,Action, Avoid,Completion] t . State and Cause identify the task-relevant facts and why they matter at the current step. Consequence reasons forward from those facts, while Effect selects the state change that should advance the task. Action connects that effect to the upcoming action chunk. Avoid records unwanted changes, and Completion preserves task progress across successive decisions. We call these action-grounded causal 4 reasoning traces because they explain an action through the effect it is intended to cause, rather than describing the action in isolation. When the same backbone is trained withArctraces, the action expert distributes substantially more atten- tion to the natural-language reasoning, particularly to the spans that connect the current condition, intended effect, and resulting action. This does not by itself establish behavioral dependence. It does show that the trace now carries information used by the action expertrather than remaining a redundant auxiliary target; Sec.5examines how this change affects task performance. 4.2ARC-Trace: Reasoning Supervision at Scale The representation above creates a supervision problem: existing robot datasets recordwhatthe robot did, but not the reasoning that made each action appropriate. Action-grounded reasoning is therefore unavailable at the scale required to train robot foundation models. Collecting it alongside new demonstrations would also be costly. Figure 3Overview of theArc-Trace relabeling pipeline, illustrated with an example task. Our key observation is that the information needed to recover this supervision is already present in a successful demonstration. The task instruction specifies the goal, the video records how the scene changes, and the action sequence reveals how those changes were produced. Modern vision-language models can extract this structure and express it in natural language. We can therefore create rea- soning supervisionafterdata collection, without gathering new robot trajectories or manually an- notating individual actions. We use this observation to constructArc-Trace- DROID by automatically relabeling the DROID dataset with action-grounded causal reasoning. Figure3summarizes the labeling pipeline. For each episode, we first generate a temporally grounded description of the task execution from the instruction and video. We then derive episode-specific state checks that capture the objects, relations, and task conditions needed to track progress through the demonstration. These checks provide structured evidence about what is true at each point in the episode and reduce the need for the final labeler to infer the entire task state from a single frame. Finally, at selected keyframes, a vision-language model combines the task instruction, episode description, current observations, and state checks to produce a reasoning trace aligned with the corresponding action chunk. This process yields approximately75K relabeled episodes and1.2M action-aligned frames. The full labeling prompts, sampling strategy, and filtering steps are provided in Appendix A.1. 4.3Fine-Tuning for Reasoning-Guided Control We fine-tune pretrained models to predict demonstrations conditioned on their aligned causal traces, ob- servations, and task instructions. Traces reuse the base tokenizer and embeddings, and are processed by the language transformer with the task and visual context. The action-generation layers attend to these representations throughout denoising; token placement, projections, and masks depend on the architecture. We retain each model’s flow-matching objective, predicting actions for the VLA and actions plus future observations for the WAM. LetKindex these modalities, with targety k t in normalized action or encoded- observation space. Its noisy counterpart isx k t = (1−s k )y k t +s k ε k , withε k ∼N(0, I)ands k ∈[0,1]sampled from the base noise schedule. The current observationo t , instructiong, and tracer t remain uncorrupted. Collecting the noisy targets and noise levels asx t ands, we minimize L pred =E " X k∈K α k MSE k v k θ (x t ,s|o t , g, r t ),ε k −y k t # .(1) 5 Here,v k θ is the velocity predicted by model parametersθ, and the expectation covers demonstrations, noise, and noise levels. Modality weightsα k and the masking, time weighting, and normalization in MSE k follow the base recipe. VLAImplementation.We useπ 0.5 -DROID (Black et al.,2025a), with a PaliGemma backbone comprising a SigLIP vision encoder and Gemma language model, and a separate 300M-parameter action expert. A two- layer encoder maps PaliGemma trace embeddings to tokens appended to the image-and-instruction context. The trace attends to this context, and the action expert attends to both; all components are jointly fine-tuned. Here,y act t =a t is the demonstrated action chunk. With factual supervision alone, predictions change little under conflicting traces for the same observation. To discourage this shortcut, we replace approximately5% of traces with counterfactualsr − t contradicting the demonstrated action and exclude these examples fromL pred . Settings act = 1and writingε=ε act , the one-step estimate is b a − t =ε−v act θ (ε,1|o t , g, r − t ). We penalize its proximity to the demonstration: L cf =E h max 0, m−MAE real ( b a − t ,a t ) 2 i ,L VLA =L pred +λ cf L cf .(2) The expectation inL cf covers counterfactual samples and action noise;L pred uses only factual samples. The marginm >0is in normalized action space,λ cf weights the penalty, and MAE real averages over the horizon and eight non-padding action coordinates. This supplies negative supervision without alternative-action demonstrations. Implementation details appear in AppendixB.1. WAMImplementation.Cosmos3-Nano-Policy (NVIDIA,2026) combines a vision-language reasoner and diffu- sion generator with separate parameters connected through attention at every layer. The generator attends to both token sets; the reasoner attends only to its own, allowing trace conditioning without architectural changes. Each window contains one observed frame and 32 future frames with actions, paired with the first frame’s trace. We jointly fine-tune both components with L WAM =L pred +λ trace L trace ,L trace =−E 2 4 |r t | X j=1 logq φ (r t,j |o t , g, r t,<j ) 3 5 .(3) Here,q φ is the reasoner’s next-token distribution with parametersφ,|r t |is the trace length, andr t,j and r t,<j denote tokenjand its prefix. The coeﬀicientλ trace weights trace supervision, whose expectation is over labeled windows. The generator conditions on ground-truth traces, and prediction-loss gradients also update the reasoner. Cosmos3’s larger reasoner (8B versusπ 0.5 ’s 2B) already produces distinct action distributions under conflicting traces. AddingL cf yields no measurable gain, so we omit it. Supporting ablations and training details appear in Appendix B.2. Inference.At inference, we freeze theArc-fine-tuned models and use an external VLM to generate reasoning traces from the task instruction and current camera observations, refreshing them at a selected frequency during execution. The VLA’s vision-language backbone and the WAM’s reasoner encode these traces into representations that condition the action expert and generator, respectively. We use external generation because the built-in language components are less reliable at generating these traces than at using them for control (Section5.2). Separately,Arcmaintains comparable task performance across the external VLMs evaluated. Since the traces are natural language, their generation can benefit directly from standard LLM inference optimizations. Appendix B.3.1details the refresh schedules and serving optimizations, including a near-real-time VLA deployment with∼10ms of added control-loop overhead. 5Experiments We evaluateπ 0.5 +Arc(VLA) and Cosmos3-Nano+Arc(WAM) (both finetuned onArc-Trace-DROID) zero-shot on RoboLab-120 ( Yang et al.,2026), MolmoSpaces (Kim et al.,2026), and RoboLab-Reasoning-50. We additionally evaluateArc-π 0.5 on paired real-world tasks without further fine-tuning. All simulation and hardware experiments use the DROID embodiment. We use Qwen3.6-35B-A3B as the default external VLM (ablated in section5.2). 6 5.1Quantitative Results: Simulation and Hardware Benchmarks General task performance.Tables1and2highlight task success forArc-variants, its base models, on Robo- Lab and MolmoSpaces. Across both benchmarks,Arcsubstantially improves both base models, enabling the comparatively weakerπ 0.5 to outperform stronger pretrained policies while further improving Cosmos3- Nano-Policy. Importantly, these gains hold without benchmark-specific training, demonstrating that the same reasoning recipe improves general manipulation across different architectures and starting levels of performance. Table breakdowns, multi-seed runs, evaluation settings and baseline details are provided in AppendixC.4. RoboLab-120: Vague # PolicySR 1π 0.5 +ARC45.1%+29.8 2Cosmos3-Nano +ARC44.9%+24.3 3 Atomic-WAM31.6% 4 BiMind v0.130.2% 5 VoLo29.5% 6Cosmos3-Nano-Policy20.6% 7 Cosmos3-Edge-Policy15.4% 8π 0.5 15.2% 9 DreamZero14.9% 10π 0 -FAST9.2% 11 GR00T N1.65.4% RoboLab-120: Default # PolicySR 1Cosmos3-Nano +ARC48.8%+12.0 2π 0.5 +ARC45.3%+17.3 3 FLUX 3 Action42.9% 4 HiDream-O1-Embodied 39.9% 5 Atomic-WAM39.6% 6 OASIS WAM39.0% 7Cosmos3-Nano-Policy36.8% 8 Phoenix34.4% 9 BiMind v0.133.3% 10 VoLo28.2% 11π 0.5 28.0% RoboLab-120: Specific # PolicySR 1Cosmos3-Nano +ARC51.2%+11.5 2π 0.5 +ARC45.0%+16.9 3Cosmos3-Nano-Policy39.7% 4 VoLo31.4% 5 BiMind v0.130.6% 6 Cosmos3-Edge-Policy28.8% 7π 0.5 28.1% 8 DreamZero23.9% 9π 0 -FAST14.9% 10 paligemma-binning5.5% 11 GR00T N1.65.3% Table 1RoboLab-120 benchmark results. SR is success rate; +values give gains over the corresponding base policy in p.p. # PolicySR# tasks 1Cosmos3-Nano +ARC57.6%+18.69/9 2 ZoeInside-VLA49.5%9/9 3 Phoenix48.0%9/9 4π 0.5 +ARC45.0%+27.29/9 5Cosmos3-Nano-Policy39.0%9/9 6 WALL-OSS-0.538.5%9/9 7 TiPToP34.7%7/9 8 PRTS-Droid27.8%9/9 9 MolmoAct2 DROID25.9%9/9 10 Psi-R222.8%9/9 11π 0.5 DROID17.8%9/9 Table 2MolmoSpaces bench- mark results. RoboLab-Reasoning-50 PolicySR Cosmos3-Nano +ARC61.0%+50.0 π 0.5 +ARC57.0%+46.8 π 0.5 + explicit subtasks21.0% Cosmos3-Nano-Policy11.0% π 0.5 10.2% RoboLab-Reasoning-Hardware PolicySR π 0.5 +ARC91.7%+82.2 Cosmos3-Nano-Policy11.9% π 0.5 9.5% Table 3Reasoning-50 and Reasoning-Hardware results. Reasoning-focused evaluation.RoboLab and MolmoSpaces target general task performance. To evaluate reasoning more directly, we introduce RoboLab-Reasoning-50, a reasoning-focused benchmark built on Robo- Lab. This complements the broader evaluations by testingArcon tasks specifically designed to require reasoning during execution. Specifically, this benchmark evaluates four major capabilities that require reason- ing: 1) Contextual tasks, requiring visual and semantic understanding; 2) Long-Horizon, requiring memory; 3) Discovery, requiring exploration via scene interaction; 4) Negation, requiring understanding exclusions. Table 3comparesArcand the baselines on this suite. The benchmark construction, individual tasks, and evaluation protocol are described in Appendix C.1.3. Beyond general manipulation,Arcsubstantially im- proves both models on tasks requiring them to reason, infer unstated steps and complete long-horizon plans. Hardware Evaluation.We deployπ 0.5 +Arcalongsideπ 0.5 and Cosmos3- Nano-Policy on the DROID hardware setup using the same checkpoints evaluated in simulation. We recreate 28 tasks from RoboLab-Reasoning-50 on hardware. Details of the comparison policies are provided in Ap- pendix C.2. Table3reports task success across categories and the corresponding simulation results. These paired evaluations test whether the improvements from reasoning persist during physical execution. The gains observed in simulation carry over to physical execution, withπ 0.5 +Arcsubstantially improvingπ 0.5 ’s task success on paired hardware tasks without hardware-specific fine-tuning. Category definitions, hardware specifications and evaluation procedures appear in Appendix C.2. 7 Figure 4Hardware rollout ofπ 0.5 +Arcon RoboLab-Reasoning-Hardware. The VLM dynamically refreshes the rea- soning traces (∼1Hz in this episode), while robot control runs at15Hz. 5.2Key Findings and Ablations Tables1,2and3support the central finding of this work that stronger grounding of causal language leads to stronger overall performance. The improvement extends across both model families.Arcliftscompara- tivelyweakermodelssuchasπ 0.5 aboveunmodifiedWAMbaselines,whilesubstantiallyimprovingalready strongmodelssuchasCosmos3-Nano-Policy. These gains extend beyond task success to robustness and eﬀiciency. We identify four key findings (KF) that indicateArcas a useful recipe for uplifting existing RFM performance. Figure 5Illustrative hardware rollout on a long-horizon, reasoning-intensive task in a dynamic environment. (KF#1)Reasoningimprovesexecutionacrosslong- horizon,dynamic,andreasoning-intensivetasks. The reasoning-focused evaluations and hardware rollouts reveal additional benefits (Tables3; Fig.5). In long-horizon tasks,Arctracks progress and main- tains the ordering of the remaining steps. When the scene changes during execution, updated reasoning redirects the robot toward what is currently needed. Reasoning-intensive tasks expose another benefit: the robot identifies prerequisites absent from the in- struction and takes the intermediate steps needed to complete the task. These behaviors are accompa- nied by stronger error recovery, with failed actions followed by corrected attempts. Our interpretation is that explicitly connecting the current state to the intended effect supports several forms of decision- making through the same representation: sequenc- ing actions, discovering necessary steps, responding to changes, and correcting mistakes. 8 SpecificDefaultVague 10 20 30 40 50 RoboLab-120 SR (%) Less explicit instructions 45.0% 45.3% 45.1% 28.1% 28.0% 15.2% SpecificDefaultVague 10 20 30 40 50 Less explicit instructions 51.2% 48.8% 44.9% 39.7% 36.8% 20.6% π 0.5 Cosmos3-Nano-Policy + ARCBase+ ARCBase Figure 6Instruction robustness on RoboLab-120. (KF#2)Performancebecomesnearlyinvarianttoin- structionspecificity. On RoboLab-120, policies fine- tuned using Arc maintains comparable success across vague, default, and specific instructions, while the base policies deteriorate as instructions become less explicit (Fig.6). This naturally suggests that reasoning reduces the amount of detail the user must supply. The trace interprets the instruction in the context of the scene and identifies the intended ef- fect of the next action, giving the controller a more specific target even when the original instruction is underspecified. 13510 Euler integration steps 10 20 30 40 50 Success rate (%) 10-step base: 28.0% 10× fewer steps 0.4 p below base 27.6% 40.8% 45.0% 45.3% 1234 UniPC solver steps 10 20 30 40 50 Success rate (%) 4-step base: 36.8% 2× fewer steps (+4.8 p) 18.0% 41.6% 47.5% 48.8% (a)(b) π 0.5 + ARC Base (10 steps) Cosmos3-Nano-Policy + ARC Base (4 steps) Figure 7Inference-time eﬀiciency on RoboLab-120 (de- fault). (KF#3)Reasoningbuysinference-timespeed-up. On RoboLab-120,π 0.5 +Arcmatches its ten-step base policy with a single Euler step (10×improvement), while Cosmos3-Nano+Arcmatches its four-step base policy with two UniPC steps (2×improve- ment)(Fig. 7). We hypothesize thatspecifying the intended effect narrows the set of plausible actions, reducing multimodality and simplifying the flow each model must integrate.Coarse integration can then suﬀice because more of the decision has already been resolved through reasoning. Cosmos3-Nano still loses performance at one step, consistent with UniPC losing the multistep history needed for higher- order correction. ARC + Cosmos3-Nano-Policy Figure 8Training eﬀiciency on RoboLab-120 (specific) for Cosmos3-Nano-Policy with and withoutArc. (KF#4)Reasoningreachestargetperformancewith fewertrainingupdates. Figure8shows thatArc reaches the instruction-only baseline’s 10K-iteration performance with approximately4.3×fewer updates, and continues to improve with further training. We conduct this study on Cosmos3-Nano-Policy using its open training implementation. Our interpreta- tion is thatcausal supervision makes each demon- stration more informative.The model receives the reason for the demonstrated action alongside the ac- tion itself, reducing what it must infer from repeated observation–action examples alone. Reasoning thus improves how eﬀiciently training produces useful be- havior, in addition to improving the final policy. Key ablations.We conduct three key design ablations on RoboLab-120.(AB#1) First, we vary the external VLM while keeping theArc-fine-tunedπ 0.5 checkpoint frozen (Fig. 9(a)). Performance remains comparable across the VLMs evaluated, showing thatArcis not tied to a particular trace generator.(AB#2)Second, we separately fine-tuneπ 0.5 for each trace variant (Fig. 9 (b)). The results points to two complementary roles within the trace: The causal account identifies which state change matters and why, while the action implication connects that change to a physical step. Without the action implication, the controller must infer how to realize the intended effect; with the action implication alone, it receives a command without the causal context that makes it appropriate. Forbidden effects likewise constrain an intended behavior but do not specify one, explaining why they help within a complete causal account but not necessarily in isolation.(AB#3)Finally, we compare full fine-tuning with action-expert-only training in the VLA and generator-only training in the WAM (Fig. 9(c)). Our interpretation is that pretrained language features do not expose causal distinctions in a form the controller can readily use. Action-head-only training leaves these features fixed and expects the controller to recover the relevant distinctions from them. 9 Full fine-tuning allows prediction errors to reshape the features themselves, teaching the backbone which conditions and intended effects must be distinguished because they require different actions. RoboLab-120 (Default) 45.6% GPT-5.5 45.3% Qwen-3.6 35B, 3B active 45.0% MiniMax-M3 428B, 23B active 44.4% Gemini 3.6 Flash 40.4% Cosmos3-Nano-Reasoner 8B 28.0% π 0.5 (base) 3.3B 02550 Success rate (%) (a) RoboLab-120 (Default) 45.3% All 6 trace components 37.17% w/o State 39.25% w/o Cause & Effect 30.58% w/o Action 35.83% w/o Avoid 42.33% w/o Completion 27.42% Action only 32.25% State + Action 26.75% State+Action+Avoid 28.0% Base (no trace) 02550 Success rate (%) (b) RoboLab-120 (Specific) 45.0% π 0.5 3.3B Full FT: VLM + action head 30.9% Action-head only finetuning 28.0% Base 51.2% Cosmos3-Nano-Policy 16B Full FT: Generator + Reasoner 43.7% Generator-head only finetuning 39.0% Base 0204060 Success rate (%) (c) Figure 9Key ablations of ARC on RoboLab-120. 6Conclusion We presented Arc, a reasoning recipe that substantially improves zero-shot task execution in pretrained RFMs without collecting new robot demonstrations or repeating foundation-scale training. Applied toπ 0.5 and Cosmos3-Nano-Policy,Arcestablishes a new state of the art on RoboLab-120 and MolmoSpaces, im- proves robustness to ambiguous and underspecified instructions, and increases training and inference ef- ficiency. Promising future directions include policies that reliably generate their own causal traces and extending the recipe across embodiments.Insimplerterms,wediscoverthatRFMsdonotneedtoberebuilt toreason.Groundedcausallanguageisacheap,high-leverageaxisforscalingthem. 10 AI Use Statement We used vision-language models to automatically generate action-grounded causal annotations for ARC-Trace and, during evaluation, to generate reasoning traces from task instructions and camera observations. The models and prompting procedures are described in the methodology and appendix. Large language models also assisted with manuscript drafting and editing. The authors take responsibility for the final manuscript, reported results, and all AI-assisted content. Reproducibility statement We describe ARC’s reasoning representation and training objectives in Sections4.1and4.3. AppendixA.1 details the ARC-Trace construction pipeline and labeling prompts. Model-specific training configurations and hyperparameters are provided in AppendicesB.1andB.2, while inference prompts and deployment procedures are documented in AppendicesB.3andB.3.1. The evaluation protocols, benchmark construction, baseline implementations, and hardware setup are described in AppendicesC.1.3,C.3, andC.2. Rollout videos are available on our fully anonymized project website athttps://arc-robot-reasoning.github.io. The models, code, and ARC-Trace dataset will be released after the review period. References Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Daniel Ho, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Eric Jang, Rosario Jauregui Ruano, Kyle Jeffrey, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Kuang-Huei Lee, Sergey Levine, Yao Lu, Linda Luu, Carolina Parada, Peter Pastor, Jornell Quiambao, Kanishka Rao, Jarek Rettinghouse, Diego Reyes, Pierre Sermanet, Nicolas Sievers, Clayton Tan, Alexander Toshev, Vincent Vanhoucke, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Mengyuan Yan, and Andy Zeng. Do As I Can, Not As I Say: Grounding Language in Robotic Affordances, August 2022. URL http: //arxiv.org/abs/2204.01691. arXiv:2204.01691 [cs.RO]. Suneel Belkhale, Tianli Ding, Ted Xiao, Pierre Sermanet, Quon Vuong, Jonathan Tompson, Yevgen Chebotar, De- bidatta Dwibedi, and Dorsa Sadigh. Rt-h: Action hierarchies using language.arXiv preprint arXiv:2403.01823, 2024. 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, 2025a. Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0 : A vision-language-action flow model for general robot control. InRobotics: Science and Systems, 2025b. Konstantinos Bousmalis, Giulia Vezzani, Dushyant Rao, Coline Devin, Alex X. Lee, Maria Bauza, Todor Davchev, Yuxiang Zhou, Agrim Gupta, Akhil Raju, Antoine Laurens, Claudio Fantacci, Valentin Dalibard, Martina Zam- belli, Murilo Martins, Rugile Pevceviciute, Michiel Blokzijl, Misha Denil, Nathan Batchelor, Thomas Lampe, Emilio Parisotto, Konrad Żołna, Scott Reed, Sergio Gómez Colmenarejo, Jon Scholz, Abbas Abdolmaleki, Oliver Groth, Jean-Baptiste Regli, Oleg Sushkov, Tom Rothörl, José Enrique Chen, Yusuf Aytar, Dave Barker, Joy Ortiz, Mar- tin Riedmiller, Jost Tobias Springenberg, Raia Hadsell, Francesco Nori, and Nicolas Heess. RoboCat: A Self- Improving Generalist Agent for Robotic Manipulation, December 2023. URL http://arxiv.org/abs/2306.11706. arXiv:2306.11706 [cs.RO]. Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. RT-1: Robotics transformer for real-world control at scale. InRobotics: Science and Systems, 2023. Jun Cen, Chaohui Yu, Hangjie Yuan, Yuming Jiang, Siteng Huang, Jiayan Guo, Xin Li, Yibing Song, Hao Luo, Fan Wang, Deli Zhao, and Hao Chen. WorldVLA: Towards Autoregressive Action World Model, June 2025. URL http://arxiv.org/abs/2506.21539. arXiv:2506.21539 [cs.RO]. 11 Chi-Lam Cheang, Guangzeng Chen, Ya Jing, Tao Kong, Hang Li, Yifeng Li, Yuxiao Liu, Hongtao Wu, Jiafeng Xu, Yichu Yang, Hanbo Zhang, and Minzhao Zhu. GR-2: A Generative Video-Language-Action Model with Web- Scale Knowledge for Robot Manipulation, October 2024. URLhttp://arxiv.org/abs/2410.06158. arXiv:2410.06158 [cs.RO]. William Chen, Jagdeep Singh Bhatia, Catherine Glossop, Nikhil Mathihalli, Ria Doshi, Andy Tang, Danny Driess, Karl Pertsch, and Sergey Levine. Steerable vision-language-action policies for embodied reasoning and hierarchical control.arXiv preprint arXiv:2602.13193, 2026. DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning.Nature, 645(8081):633–638, September 2025. ISSN 0028-0836, 1476-4687. doi: 10.1038/s41586-025-09422-z. URL http://arxiv.org/abs/2501.12948. arXiv:2501.12948 [cs.CL]. Danny Driess, Fei Xia, Mehdi S. M. Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, Yevgen Chebotar, Pierre Sermanet, Daniel Duckworth, Sergey Levine, Vincent Vanhoucke, Karol Hausman, Marc Toussaint, Klaus Greff, Andy Zeng, Igor Mordatch, and Pete Florence. PaLM-E: An Embodied Multimodal Language Model, March 2023. URL http://arxiv.org/abs/2303.03378. arXiv:2303.03378 [cs.LG]. Yilun Du, Sherry Yang, Bo Dai, Hanjun Dai, Ofir Nachum, Josh Tenenbaum, Dale Schuurmans, and Pieter Abbeel. Learning universal policies via text-guided video generation.Advances in neural information processing systems, 36:9156–9172, 2023. Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, Jinlan Fu, Jingjing Gong, and Xipeng Qiu. LIBERO-Plus: In-depth robustness analysis of vision-language-action models.arXiv preprint arXiv:2510.13626, 2025. Jiayuan Gu, Sean Kirmani, Paul Wohlhart, Yao Lu, Montserrat Gonzalez Arenas, Kanishka Rao, Wenhao Yu, Chuyuan Fu, Keerthana Gopalakrishnan, Zhuo Xu, et al. Rt-trajectory: Robotic task generalization via hind- sight trajectory sketches.arXiv preprint arXiv:2311.01977, 2023. Wenlong Huang, Fei Xia, Ted Xiao, Harris Chan, Jacky Liang, Pete Florence, Andy Zeng, Jonathan Tompson, Igor Mordatch, Yevgen Chebotar, Pierre Sermanet, Noah Brown, Tomas Jackson, Linda Luu, Sergey Levine, Karol Hausman, and Brian Ichter. Inner Monologue: Embodied Reasoning through Planning with Language Models, July 2022. URL http://arxiv.org/abs/2207.05608. arXiv:2207.05608 [cs.RO]. Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models.arXiv preprint arXiv:2307.05973, 2023a. Wenlong Huang, Fei Xia, Dhruv Shah, Danny Driess, Andy Zeng, Yao Lu, Pete Florence, Igor Mordatch, Sergey Levine, Karol Hausman, and Brian Ichter. Grounded Decoding: Guiding Text Generation with Grounded Models for Embodied Agents, December 2023b. URL http://arxiv.org/abs/2303.00855. arXiv:2303.00855 [cs.RO]. 12 Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation.arXiv preprint arXiv:2409.01652, 2024. Eric Jang, Alex Irpan, Mohi Khansari, Daniel Kappler, Frederik Ebert, Corey Lynch, Sergey Levine, and Chelsea Finn. BC-Z: Zero-Shot Task Generalization with Robotic Imitation Learning, February 2022. URLhttp://arxiv. org/abs/2202.02005. arXiv:2202.02005 [cs.RO]. Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anand- kumar, Yuke Zhu, and Linxi Fan. VIMA: General Robot Manipulation with Multimodal Prompts, May 2023. URL http://arxiv.org/abs/2210.03094. arXiv:2210.03094 [cs.RO]. Alexander Khazatsky, Karl Pertsch, et al. DROID: A large-scale in-the-wild robot manipulation dataset. InRobotics: Science and Systems, 2024. Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. OpenVLA: An open-source vision-language-action model. InConference on Robot Learning, 2024. Yejin Kim, Wilbert Pumacay, Omar Rayyan, Max Argus, Winson Han, Eli VanderBilt, Jordi Salvador, Abhay Desh- pande, Rose Hendrix, Snehal Jauhri, et al. Molmospaces: A large-scale open ecosystem for robot navigation and manipulation.arXiv preprint arXiv:2602.11337, 2026. Po-Chen Ko, Jiayuan Mao, Yilun Du, Shao-Hua Sun, and Joshua B Tenenbaum. Learning to act from actionless videos through dense correspondences. InInternational Conference on Learning Representations, volume 2024, p. 40938–40958, 2024. Jason Lee, Jiafei Duan, Haoquan Fang, Yuquan Deng, Shuo Liu, Boyang Li, Bohan Fang, Jieyu Zhang, Yi Ru Wang, Sangho Lee, et al. MolmoAct: Action reasoning models that can reason in space.arXiv preprint arXiv:2508.07917, 2025. Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as Policies: Language Model Programs for Embodied Control, May 2023. URLhttp://arxiv.org/abs/2209.07753. arXiv:2209.07753 [cs.RO]. Fangchen Liu, Kuan Fang, Pieter Abbeel, and Sergey Levine. Moka: Open-world robotic manipulation through mark-based visual prompting.arXiv preprint arXiv:2403.03174, 2024. Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought.Advances in neural information processing systems, 36:25081–25094, 2023. NVIDIA. Cosmos 3: Omnimodal world models for physical AI.arXiv preprint arXiv:2606.02800, 2026. Abby O’Neill, Abdul Rehman, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Poo- ley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Anikait Singh, Animesh Garg, Aniruddha Kembhavi, Annie Xie, Anthony Brohan, Antonin Raﬀin, Archit Sharma, Arefeh Yavary, Arhan Jain, Ashwin Balakrishna, Ayzaan Wahid, Ben Burgess-Limerick, Beomjoon Kim, Bernhard Schölkopf, Blake Wulfe, Brian Ichter, Cewu Lu, Charles Xu, Charlotte Le, Chelsea Finn, Chen Wang, Chenfeng Xu, Cheng Chi, Chenguang Huang, Christine Chan, Christo- pher Agia, Chuer Pan, Chuyuan Fu, Coline Devin, Danfei Xu, Daniel Morton, Danny Driess, Daphne Chen, Deepak Pathak, Dhruv Shah, Dieter Büchler, Dinesh Jayaraman, Dmitry Kalashnikov, Dorsa Sadigh, Edward Johns, Ethan Foster, Fangchen Liu, Federico Ceola, Fei Xia, Feiyu Zhao, Freek Stulp, Gaoyue Zhou, Gaurav S. Sukhatme, Gau- tam Salhotra, Ge Yan, Gilbert Feng, Giulio Schiavi, Glen Berseth, Gregory Kahn, Guanzhi Wang, Hao Su, Hao-Shu Fang, Haochen Shi, Henghui Bao, Heni Ben Amor, Henrik I Christensen, Hiroki Furuta, Homer Walke, Hongjie Fang, Huy Ha, Igor Mordatch, Ilija Radosavovic, Isabel Leal, Jacky Liang, Jad Abou-Chakra, Jaehyung Kim, Jaimyn Drake, Jan Peters, Jan Schneider, Jasmine Hsu, Jeannette Bohg, Jeffrey Bingham, Jeffrey Wu, Jensen Gao, Jiaheng Hu, Jiajun Wu, Jialin Wu, Jiankai Sun, Jianlan Luo, Jiayuan Gu, Jie Tan, Jihoon Oh, Jimmy Wu, Jingpei Lu, Jingyun Yang, Jitendra Malik, João Silvério, Joey Hejna, Jonathan Booher, Jonathan Tompson, Jonathan Yang, Jordi Salvador, Joseph J. Lim, Junhyek Han, Kaiyuan Wang, Kanishka Rao, Karl Pertsch, Karol Hausman, Keegan Go, Keerthana Gopalakrishnan, Ken Goldberg, Kendra Byrne, Kenneth Oslund, Kento Kawaharazuka, Kevin Black, Kevin Lin, Kevin Zhang, Kiana Ehsani, Kiran Lekkala, Kirsty Ellis, Krishan Rana, Krishnan Srini- vasan, Kuan Fang, Kunal Pratap Singh, Kuo-Hao Zeng, Kyle Hatch, Kyle Hsu, Laurent Itti, Lawrence Yunliang Chen, Lerrel Pinto, Li Fei-Fei, Liam Tan, Linxi Jim Fan, Lionel Ott, Lisa Lee, Luca Weihs, Magnum Chen, Marion Lepert, Marius Memmel, Masayoshi Tomizuka, Masha Itkina, Mateo Guaman Castro, Max Spero, Maximilian Du, Michael Ahn, Michael C. Yip, Mingtong Zhang, Mingyu Ding, Minho Heo, Mohan Kumar Srirama, Mohit Sharma, 13 Moo Jin Kim, Naoaki Kanazawa, Nicklas Hansen, Nicolas Heess, Nikhil J Joshi, Niko Suenderhauf, Ning Liu, Nor- man Di Palo, Nur Muhammad Mahi Shafiullah, Oier Mees, Oliver Kroemer, Osbert Bastani, Pannag R Sanketi, Patrick Tree Miller, Patrick Yin, Paul Wohlhart, Peng Xu, Peter David Fagan, Peter Mitrano, Pierre Sermanet, Pieter Abbeel, Priya Sundaresan, Qiuyu Chen, Quan Vuong, Rafael Rafailov, Ran Tian, Ria Doshi, Roberto Martín- Martín, Rohan Baijal, Rosario Scalise, Rose Hendrix, Roy Lin, Runjia Qian, Ruohan Zhang, Russell Mendonca, Rutav Shah, Ryan Hoque, Ryan Julian, Samuel Bustamante, Sean Kirmani, Sergey Levine, Shan Lin, Sherry Moore, Shikhar Bahl, Shivin Dass, Shubham Sonawani, Shuran Song, Sichun Xu, Siddhant Haldar, Siddharth Karamcheti, Simeon Adebola, Simon Guist, Soroush Nasiriany, Stefan Schaal, Stefan Welker, Stephen Tian, Subramanian Ra- mamoorthy, Sudeep Dasari, Suneel Belkhale, Sungjae Park, Suraj Nair, Suvir Mirchandani, Takayuki Osa, Tanmay Gupta, Tatsuya Harada, Tatsuya Matsushima, Ted Xiao, Thomas Kollar, Tianhe Yu, Tianli Ding, Todor Davchev, Tony Z. Zhao, Travis Armstrong, Trevor Darrell, Trinity Chung, Vidhi Jain, Vincent Vanhoucke, Wei Zhan, Wenx- uan Zhou, Wolfram Burgard, Xi Chen, Xiaolong Wang, Xinghao Zhu, Xinyang Geng, Xiyuan Liu, Xu Liangwei, Xuanlin Li, Yao Lu, Yecheng Jason Ma, Yejin Kim, Yevgen Chebotar, Yifan Zhou, Yifeng Zhu, Yilin Wu, Ying Xu, Yixuan Wang, Yonatan Bisk, Yoonyoung Cho, Youngwoon Lee, Yuchen Cui, Yue Cao, Yueh-Hua Wu, Yu- jin Tang, Yuke Zhu, Yunchu Zhang, Yunfan Jiang, Yunshuang Li, Yunzhu Li, Yusuke Iwasawa, Yutaka Matsuo, Zehan Ma, Zhuo Xu, Zichen Jeff Cui, Zichen Zhang, and Zipeng Lin. Open X-Embodiment: Robotic Learning Datasets and RT-X Models : Open X-Embodiment Collaboration 0 . In2024 IEEE International Conference on Robotics and Automation (ICRA), p. 6892–6903, Yokohama, Japan, May 2024. IEEE. ISBN 9798350384574. doi: 10.1109/ICRA57147.2024.10611477. URL https://ieeexplore.ieee.org/document/10611477/. Scott Reed, Konrad Zolna, Emilio Parisotto, Sergio Gomez Colmenarejo, Alexander Novikov, Gabriel Barth-Maron, Mai Gimenez, Yury Sulsky, Jackie Kay, Jost Tobias Springenberg, Tom Eccles, Jake Bruce, Ali Razavi, Ashley Edwards, Nicolas Heess, Yutian Chen, Raia Hadsell, Oriol Vinyals, Mahyar Bordbar, and Nando de Freitas. A generalist agent, 2022. URLhttps://arxiv.org/abs/2205.06175. Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Cliport: What and where pathways for robotic manipulation, 2021. URL https://arxiv. org/abs/2109.12098, 2021. Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. InConference on Robot Learning, p. 785–799. PMLR, 2023. Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. ProgPrompt: Generating Situated Robot Task Plans using Large Language Models, September 2022. URLhttp://arxiv.org/abs/2209.11302. arXiv:2209.11302 [cs.RO]. Nan Sun, Yuan Zhang, Yongkun Yang, Wentao Zhao, Peiyan Li, Jun Guo, Wenxuan Song, Pengxiang Ding, Runze Suo, Yifei Su, et al. Revisiting embodied chain-of-thought for generalizable robot manipulation.arXiv preprint arXiv:2606.03784, 2026. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-Consistency Improves Chain of Thought Reasoning in Language Models, March 2023. URL http: //arxiv.org/abs/2203.11171. arXiv:2203.11171 [cs.CL]. Yan Wang, Wenjie Luo, Junjie Bai, Yulong Cao, Tong Che, Ke Chen, Yuxiao Chen, Jenna Diamond, Yifan Ding, Wenhao Ding, et al. Alpamayo-R1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail.arXiv preprint arXiv:2511.00088, 2025. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, January 2023. URLhttp: //arxiv.org/abs/2201.11903. arXiv:2201.11903 [cs.CL]. Hongtao Wu, Ya Jing, Chilam Cheang, Guangzeng Chen, Jiafeng Xu, Xinghang Li, Minghuan Liu, Hang Li, and Tao Kong. Unleashing Large-Scale Video Generative Pre-training for Visual Robot Manipulation, December 2023. URL http://arxiv.org/abs/2312.13139. arXiv:2312.13139 [cs.RO]. Jenai Xuning Yang, Rishit Dagli, Alex Zook, Hugo Hadfield, Ankit Goyal, Stan Birchfield, Fabio Ramos, and Jonathan Tremblay. RoboLab: A high-fidelity simulation benchmark for analysis of task generalist policies.arXiv preprint arXiv:2604.09860, 2026. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griﬀiths, Yuan Cao, and Karthik Narasimhan. Tree of Thoughts: Deliberate Problem Solving with Large Language Models, December 2023a. URLhttp://arxiv.org/ abs/2305.10601 . arXiv:2305.10601 [cs.CL]. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing 14 Reasoning and Acting in Language Models, March 2023b. URLhttp://arxiv.org/abs/2210.03629. arXiv:2210.03629 [cs.CL]. 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 ”Jim” Fan, and Joel Jang. World Action Models are Zero-shot Policies, February 2026. URLhttp://arxiv.org/abs/2602.15922. arXiv:2602.15922 [cs.RO]. Michał Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. InConference on Robot Learning, 2024. Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D. Goodman. Quiet-STaR: Lan- guage Models Can Teach Themselves to Think Before Speaking, March 2024. URLhttp://arxiv.org/abs/2403. 09629. arXiv:2403.09629 [cs.CL]. Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, Johnny Lee, Vincent Vanhoucke, and Pete Florence. Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language, May 2022a. URLhttp://arxiv.org/abs/ 2204.00598. arXiv:2204.00598 [cs.CV]. Andy Zeng, Pete Florence, Jonathan Tompson, Stefan Welker, Jonathan Chien, Maria Attarian, Travis Armstrong, Ivan Krasin, Dan Duong, Ayzaan Wahid, Vikas Sindhwani, and Johnny Lee. Transporter networks: Rearranging the visual world for robotic manipulation, 2022b. URL https://arxiv.org/abs/2010.14406. Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal Chain-of-Thought Reasoning in Language Models, May 2024. URLhttp://arxiv.org/abs/2302.00923. arXiv:2302.00923 [cs.CL]. Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3d-vla: A 3d vision-language-action generative world model.arXiv preprint arXiv:2403.09631, 2024. Ruijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao, Hal Daumé I, Andrey Kolobov, Furong Huang, and Jianwei Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies. InInternational Conference on Learning Representations, volume 2025, p. 54277–54296, 2025. Linqing Zhong, Yi Liu, Yifei Wei, Ziyu Xiong, Si Liu, and Guanghui Ren. Acot-vla: Action chain-of-thought for vision-language-action models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8152–8162, 2026. Siyuan Zhou, Yilun Du, Jiaben Chen, Yandong Li, Dit-Yan Yeung, and Chuang Gan. Robodreamer: Learning compositional world models for robot imagination.arXiv preprint arXiv:2404.12377, 2024. Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, et al. RT-2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning, 2023. 15 Appendix This appendix provides detailed recipes for data labeling, training, and inference, together with hyperpa- rameters, prompts, additional qualitative results, and a discussion of limitations. All experiment videos are available on our fully anonymized project website athttps://arc-robot-reasoning.github.io. The models, code, and ARC-Trace dataset will be released after the review period. A ARC-TRACE and ARC-TRACE-DROID details17 A.1 ARC-Trace: Data Labeling Details. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .17 B Fine-tuning and inference recipe details20 B.1 VLA Fine-Tuning Details. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .20 B.2 WAM Fine-Tuning Details. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .21 B.3 Inference. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .23 B.3.1 Real-Time Deployment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .23 C Experiment Details24 C.1 Simulation Benchmark Details. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24 C.1.1 RoboLab-120. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24 C.1.2 MolmoSpaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24 C.1.3 RoboLab-Reasoning-50. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .24 C.2 RoboLab-Reasoning-Hardware. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .26 C.3 Evaluated Policies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32 C.4 Detailed RoboLab-120 results.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .32 D Training Hyperparameters and Configurations.33 E Additional Qualitative Results35 F Limitations of ARC36 G Prompts36 G.1 Video annotation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .36 G.2 Frame-by-frame reasoning with predicates. . . . . . . . . . . . . . . . . . . . . . . . . . . . .39 G.3 Inference-time reasoner prompts. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .40 16 AARC-TRACE and ARC-TRACE-DROID details A.1ARC-Trace: Data Labeling Details ARC-Trace is constructed through three stages, illustrated in Fig.3. Gemini 2.5 Pro interprets the episode video and selects important frames. GPT-5.5 then generates episode-specific state checks and combines this evidence with the task, episode narrative, and selected observations to produce action-grounded causal traces. Separating episode-level interpretation from frame-level annotation gives each trace the context needed to explain the current decision. Source data and coverage.We use the training split of DROID v1.0.1 (Khazatsky et al.,2024), containing 95,658 episodes in RLDS/TFDS format. Demonstrations are recorded at 15,Hz and include wrist and two exterior camera views, robot states, actions, and task instructions. The labeling run covers approximately 74,740 episodes, or 78.1% of the source split, and produces approximately 1.2 million annotated frames. This corresponds to roughly 16.1 annotations per labeled episode. Most episodes without labels lack a usable task instruction; processing failures are recorded separately. Episode narration and keyframe selection.Gemini 2.5 Pro receives the task instruction and episode video and produces a temporally grounded narrative of the demonstration. The narrative describes the objects involved, the robot’s actions, and the resulting changes in the scene. This episode-level account provides context that an isolated image cannot establish, including which parts of the task have already been completed. The same video-labeling stage selects important frames for annotation. Selection is guided by the VLM’s interpretation of the demonstration, rather than a fixed uniform sampling interval. Each selected frame is mapped back to its original frame indexfand timestampf/15. Episode-specific predicate generation.For the selected observations, GPT-5.5 generates state checks describing the objects, relations, and conditions relevant to the task. We use the low reasoning-effort setting for this stage. In the example in Fig. 3, these checks includetarget_object,target_container,pot_visible,gripper_open They identify the object and destination named by the instruction and make relevant visual conditions explicit before the final trace is composed. Their outputs provide structured evidence to the trace labeler, reducing the need to reconstruct the task state from the image and narrative alone. These are model-generated descriptions of visual evidence, not independently measured ground-truth states. Action-grounded trace generation.GPT-5.5, again using low reasoning effort, combines the task instruction, episode narrative, selected wrist observation, and state-check outputs into a structured reasoning annotation. The annotation contains the seven semantic fields defined in Sec. 4.1, linking the perceived state and relevant cause to a possible consequence, chosen effect, action implication, forbidden effect, and completion status. Each record includes anatural_language_construction, a fluent paragraph of three to six sentences targeting approximately 700 characters. This paragraph combines the causal account with the relevant object and grasp information. The structured fields support inspection and representation ablations, while the paragraph provides the natural-language reasoning used for policy conditioning. Temporal alignment and saved artifacts.Each annotation retains its episode index, keyframe identifier, original frame index, and timestamp. The episode index identifies the source record in the RLDS training split, and the frame index identifies the corresponding observation and start of the associated action chunk. This preserves alignment with the original demonstration without modifying its observations or actions. We store one JSON record per episode, with annotations ordered chronologically. Selected images, episode videos, raw model responses, parsed outputs, model identifiers, and prompt versions are retained alongside the labels. The full episode video is used offline to construct supervision; it is not an additional input to the deployed policy. We show an annotated example from a randomly selected DROID episode in Figures 10and11. 17 000 001 002 003 004 005 006 007 008 009 010 011 012 013 014 015 016 017 018 019 020 021 022 023 024 025 026 027 028 029 030 031 032 033 034 035 036 037 038 039 040 041 042 043 044 045 046 047 048 049 050 051 052 053 Task:“Put the marker in the pot.” Keyframes000 0.00s External camera Wrist camera Action-grounded causal reasoning traces [State]The scene shows an open metal pot on the left side of the table and a black-and-white marker standing upright closer to the center.[Cause]This matters because the marker must be moved from its current separated position into the pot, and the robot has not yet made contact with it. [Effect]The next useful step is for the gripper to approach the marker carefully so it can be grasped without tipping or pushing it away. [Action]The robot should continue toward the marker [Avoid]while avoiding collisions with the pot, the table edge, or nearby clutter. [Completion]The task is incomplete because the marker remains outside the pot. Keyframes001 1.20s External camera Wrist camera Action-grounded causal reasoning traces [State]The pot is open on the left side of the table while the marker remains upright on the tabletop to its right.[Cause]This separation matters because the robot still needs to pick up or otherwise move the marker before it can be put into the pot. [Effect]The next likely step is to bring the gripper closer and align it with the marker for a stable grasp.[Action]The robot should continue carefully,[Avoid]avoiding any motion that would topple the marker or shift the pot. [Completion]The task is still incomplete because the marker is not in the pot. Keyframes002 2.40s External camera Wrist camera Action-grounded causal reasoning traces [State]The pot remains open on the left, while the marker is still on the tabletop to its right. The gripper has moved closer and appears to be lining up for a grasp, [Cause]which is a necessary causal step before the marker can be lifted or moved into the pot.[Avoid]If the robot approaches with too much sideways motion or closes too early, it may push the marker away rather than secure it. [Action]The robot should continue with careful alignment and prepare to grasp the marker [Avoid]without disturbing the pot.[Completion] The task is not complete because the marker has not been placed in the pot. Keyframes003 3.67s External camera Wrist camera Action-grounded causal reasoning traces [State]The marker is on the tabletop near the robot gripper, while the open pot remains to the left.[Cause] This matters because the robot has reached the critical pre-grasp position where small alignment errors could determine whether the marker is captured or pushed away.[Effect]The intended next effect is to close or position the gripper so the marker is held securely for transport. [Action]The robot should continue carefully, [Avoid]avoiding any sideways contact that could roll the marker or disturb the pot. [Completion]The task is still incomplete because the marker has not been deposited in the pot. Keyframes004 3.80s External camera Wrist camera Action-grounded causal reasoning traces [State]The marker is still lying on the tabletop near the robot gripper, while the open pot remains to the left. [Cause]This moment matters because the robot appears to be shifting from approach into the grasping phase, and the marker must be controlled before it can be put into the pot.[Avoid]A poorly aligned contact could roll the marker away or push it across the table instead of securing it. [Action]The robot should continue by making a careful, stable grasp and then move the marker toward the pot. [Completion]The task remains incomplete because the marker is not yet inside the pot. Keyframes005 4.87s External camera Wrist camera Action-grounded causal reasoning traces [State]The robot gripper is now contacting the marker near its lower end, and the marker is upright or nearly upright at the grasp point.[Cause]This matters because the robot must first secure the marker before it can move it into the open pot visible on the left. [Avoid]If the grip is not stable, the marker may slip, rotate, or be pushed away instead of being carried. [Action]The robot should continue by confirming a firm grasp and then moving the marker toward the pot [Avoid]without bumping the pot or table. [Completion]The task is still incomplete because the marker has not been deposited inside the pot. Figure 1: All labeled keyframes from DROID episode 0, part 1 of 2 (keyframes 000–005). Each card pairs synchronized external- and wrist-camera views with the full natural-language long-trace annotation. Source wording is preserved; inline component labels are added for readability. 1 Figure 10Full natural-language reasoning annotations for “Put the marker in the pot,” with synchronized external- and wrist-camera views. Part 1: labeled keyframes 000–005. 18 054 055 056 057 058 059 060 061 062 063 064 065 066 067 068 069 070 071 072 073 074 075 076 077 078 079 080 081 082 083 084 085 086 087 088 089 090 091 092 093 094 095 096 097 098 099 100 101 102 103 104 105 106 107 Task:“Put the marker in the pot.” Keyframes006 6.07s External camera Wrist camera Action-grounded causal reasoning traces [State]The robot is now holding the marker upright near its lower end, indicating that the grasp has likely been established.[Cause]This matters because controlled possession of the marker is required before the robot can carry it to the pot. [State]The pot is not visible in this frame,[Effect]so the next useful effect is to transport the held marker toward the pot while keeping it stable. [Action]The robot should continue carefully[Avoid]and avoid dropping the marker or colliding with surrounding surfaces.[Completion] The task remains incomplete because the marker has not been placed into the pot. Keyframes007 7.33s External camera Wrist camera Action-grounded causal reasoning traces [State]The robot is holding the marker upright and has brought it close to the open metal pot visible ahead. [Cause]This matters because the task has progressed from grasping to target alignment, and the marker must now be placed into the pot rather than simply carried near it.[Effect]A careful forward or upward alignment could put the marker over the opening, [Avoid]but an early release or poor positioning could cause it to hit the rim or fall outside.[Action]The robot should continue by centering the marker over the pot and lowering or releasing it only when placement is secure.[Completion]The task is still incomplete because the marker is not yet inside the pot. Keyframes008 8.53s External camera Wrist camera Action-grounded causal reasoning traces [State]The robot is holding the marker upright and has moved it directly in front of the open metal pot. [Cause]This matters because the marker is now close enough that the next small positioning motion can determine whether it enters the pot or misses the opening. [Effect]If the robot centers and lowers it carefully, the marker can be placed inside,[Avoid]but an off-center release could make it hit the rim or fall outside. [Action]The robot should continue with fine alignment and release only when the marker is clearly over the pot. [Completion]The task is still incomplete because the marker remains in the gripper rather than inside the pot. Keyframes009 9.73s External camera Wrist camera Action-grounded causal reasoning traces [State]The marker is now visibly positioned inside the open metal pot, and the gripper appears open rather than holding it. [Cause]This matters because the required object has reached the target container, satisfying the main goal of the instruction. [Avoid] The robot should avoid any additional motion that might strike the pot or knock the marker back out.[Action] A safe withdrawal or stopping motion is appropriate now that the placement has been achieved.[Completion] The task is complete because the marker is in the pot. Keyframes010 10.40s External camera Wrist camera Action-grounded causal reasoning traces [State]The marker is still visibly inside the open metal pot, and the robot gripper is now away from the marker rather than holding it.[Cause]This matters because the required placement has already been achieved, and the remaining concern is preserving that state. [Effect]A careful withdrawal will keep the marker in the pot, [Avoid]while unnecessary motion could hit the rim or dislodge the marker.[Action]The robot should retract or stop without re-contacting the pot or marker.[Completion]The task is complete because the marker has been placed in the pot. Keyframes011 11.00s External camera Wrist camera Action-grounded causal reasoning traces [State]The marker is still visibly inside the open metal pot, and the robot gripper is away from it. [Cause]This matters because the placement has remained stable after release, which satisfies the instruction to put the marker in the pot. [Effect]The best next outcome is to preserve this state[Action]by stopping or staying clear of the container. [Avoid]The robot should avoid re-contacting the marker or bumping the pot, since that could undo the successful placement. [Completion]The task is complete because the marker is in the pot. Figure 2: All labeled keyframes from DROID episode 0, part 2 of 2 (keyframes 006–011). Each card pairs synchronized external- and wrist-camera views with the full natural-language long-trace annotation. Source wording is preserved; inline component labels are added for readability. 2 Figure 11Full natural-language reasoning annotations for “Put the marker in the pot,” continued. Part 2: labeled keyframes 006–011. 19 BFine-tuning and inference recipe details B.1VLA Fine-Tuning Details π 0.5 -DROID + ARC Current observation Task + robot state Causal trace factual / counterfactual Action chunk normalized SigLIP vision encoder Base tokenizer + PaliGemma embeddings Corrupt action chunk x = (1-s)a + sε Two-layer trace encoder Action projection Noise level s Time embedding v 1 ...g r 1 r 2 ... x s Uncorrupted image, task and trace tokensNoisy action tokens Gemma language model · 2B Action expert 300M parameters Trace attends to image + task context Context does not attend to trace Attention to context + trace K / V RMSNorm + MLP AdaRMSNorm + MLP × L Predicted action velocity v θ act Factual traces · ≈95%Flow matching L pred = MSE(v θ act , ε - a) Counterfactual traces · ≈5% s = 1; â - = ε - v - L cf = max(0, m - MAE real (â - , a)) 2 No flow-matching loss on these examples L VLA = L pred + λ cf L cf Fine-tuned Figure 12Arcfine-tuning architecture forπ 0.5 Observation and instruction encoding.Figure12shows theπ 0.5 -DROID implementation. The exterior and wrist images ino t are resized to224×224and encoded by SigLIP into 256 tokens per view. These tokens are projected to the Gemma backbone’s width of 2048. The task instructiongand discretized robot state are encoded together using the PaliGemma tokenizer, with a maximum length of 200 tokens. Image and instruction tokens form the original context; unused camera slots and padding are masked. Trace encoding.The tracer t is tokenized separately, with a maximum length of 320 tokens, and embedded using the same PaliGemma embedding table. The trace encoder projects these 2048-dimensional embeddings to width 512, adds learned positional embeddings, and applies two pre-normalized transformer blocks, each with eight attention heads and a feed-forward width of 2048. A final projection returns the representations to width 2048 before they are appended to the context. This projection is zero-initialized, so the added representations initially carry no trace-specific information. Layerwise conditioning.Context and trace tokens are processed by the Gemma-2B parameters, while noisy action tokens are processed by the 300M-parameter action expert. The two components have widths 2048 and 1024, respectively, and interact through shared attention at each of the 18 transformer layers. Their attention projections produce compatible keys, queries, and values despite the different hidden widths. The block mask is 20 Key tokens Query tokens Context Context c 1 c 1 c 2 c 2 c 3 c 3 c 4 c 4 Trace Trace t 1 t 1 t 2 t 2 t 3 t 3 t 4 t 4 Actions Actions a 1 a 1 a 2 a 2 a 3 a 3 a 4 a 4 Full attention within allowed blocks Allowed attentionBlocked where rows denote the attending tokens, columns denote the attended tokens, and1permits attention to all valid tokens in the corresponding block. Thus, the trace can use the observed scene and instruction, while the action expert can combine both with the noisy action chunk. Neither context nor trace representations depend on the noisy actions. Although trace tokens precede action tokens in the input sequence, their rotary position indices are assigned after the action positions, preserving the pretrained positional indices of the action tokens. Action prediction and training.With chunk lengthH= 15, the normalized action target isy act t =a t ∈R H×32 . Eight coordinates represent the seven joints and gripper; the remaining coordinates are padding. Each row ofx act t is projected to one 1024-dimensional action token. A sinusoidal embedding ofs act , followed by an MLP, conditions the action expert through adaptive RMS normalization. Its final token representations are projected back to 32 coordinates to producev act θ . The vision encoder, language backbone, trace encoder, and action expert are jointly optimized with Eq. 2. Factual and counterfactual examples share one batched forward pass. Their loss masks select flow matching or the counterfactual penalty, respectively, withs act = 1for counterfactual examples. The flow loss uses the padded action representation, whereas MAE real excludes padding when measuring counterfactual separation. B.2WAM Fine-Tuning Details Token construction.Figure13shows the Cosmos3-Nano-Policy implementation. The reasoner receives ViT- encoded current observations followed by the task instructiongand tracer t , using the pretrained language tokenizer and embeddings. The trace is appended directly to the instruction without a separate trace encoder. End-of-sequence and begin-of-generation tokens mark the boundary between the reasoner and generator sequences. The generator operates on continuous observation and action representations. Each training window contains one observed frame and 32 future frames with aligned actions. The video VAE encodes the future observations intoy obs t , while the normalized actions formy act t . Their noisy versions,x obs t andx act t , are projected into generator tokens. The base Cosmos video input also retains the initial frame as clean VAE conditioning, separate from the noised future targets ( NVIDIA,2026). Reasoner–generator attention.Cosmos3-Nano uses 36 paired transformer layers of width 4096. The reasoner and generator have separate normalization layers, attention projections, and feed-forward networks, but their tokens interact through a shared attention operation (NVIDIA,2026). The block mask is 21 Cosmos3-Nano-Policy + ARC Current observation Task instruction Ground-truth causal trace Future observations 32 frames Actions normalized Vision encoder (ViT) Base tokenizer + embeddings Vision encoder (VAE) Noised targets: x = (1-s)y + sε Action encoder Uncorrupted conditioning v 1 v 2 ...g r 1 r 2 ...EOSBOG x t obs x t act Autoregressive tokensNoisy diffusion tokens + s ReasonerGenerator LayerNormLayerNorm Causal self-attention Reasoner tokens only Attention to both streams Reasoner + generator tokens K / V LayerNorm + MLPLayerNorm + MLP × L Separate parameters; attention connected at every layer Predicted causal-trace tokensPredicted velocities r 1 r 2 ... Obs. v θ obs Act. v θ act Next-token lossL trace Observation FM L FM obs = E[MSE obs (v θ obs , ε obs - y t obs )] Action FM L FM act = E[MSE act (v θ act , ε act - y t act )] L WAM = L FM obs + 10 L FM act + λ trace L trace Both flow-matching losses also update the reasoner. Fine-tuned Figure 13Arcfine-tuning architecture for Cosmos3-nano-policy Key tokens Query tokens Reasoner Reasoner r 1 r 1 r 2 r 2 r 3 r 3 r 4 r 4 Generator Generator g 1 g 1 g 2 g 2 g 3 g 3 g 4 g 4 Causal only within Reasoner Allowed attentionBlocked using the same row and column convention as the VLA. Causal attention restricts each reasoner token to its preceding context and itself. Generator tokens attend to the complete reasoner sequence and to one another, allowing action and future-observation predictions to exchange information while remaining conditioned on the trace. Separate output projections producev act θ andv obs θ in their respective target spaces. Joint supervision.The ground-truth trace serves two roles during training. It supplies conditioning to the generator and next-token targets to the reasoner. ForL trace , targets are shifted so that prediction ofr t,j uses onlyo t ,g, andr t,<j . The reasoner cannot attend to later trace tokens or diffusion targets. The generator, however, conditions on the complete ground-truth trace when predicting actions and future observations. We jointly update both components using Eq. 3. The trace loss supervises the reasoner’s language predictions, whileL pred also updates its representations through the keys and values used by the generator. Consequently, the reasoner is trained both to express the annotated reasoning and to provide features useful for continuous prediction. The counterfactual objective is omitted, as described in Sec. 4.3. 22 B.3Inference During evaluation, we freeze the ARC-fine-tuned policies and use an external VLM to generate action- grounded causal traces. Each request contains the task instruction and the current exterior- and wrist-camera images. A fixed prompt specifies the ARC trace format and requests a single natural-language paragraph of at most 900 characters, grounded in the current observations. The generated trace is supplied alongside the original instruction and observations. In the VLA, the vision- language backbone encodes the trace for the action expert; in the WAM, the reasoner encodes it for the generator. The external VLM provides the reasoning, while the ARC-fine-tuned policy remains responsible for generating continuous robot actions. Responses are checked for format and length before use, and a rejected response leaves the previous accepted trace unchanged. Trace generation runs asynchronously, with each policy call using the latest accepted trace. B.3.1Real-Time Deployment Asynchronous trace updates Policy and control Task + camera views External + wrist · 480 × 480 JPEGs Local reasoner · vLLM Qwen3.6-35B-A3B 35B total · 3B active · FP8 Cached prefix · thinking off trace Latest valid trace Validate before replacing One causal-trace paragraph Reuse until next accepted update Read latest accepted trace Current policy inputs Task + views + robot state π 0.5 -DROID + ARC Policy requests · 1 Hz Action chunk 15 steps · 1 s Robot control 15 Hz Trace generation overlaps action execution. Figure 14Real-time deployment strategy We separate reasoning generation from action execution so that the robot does not wait for a new trace at every policy query. The robot executes 15-action chunks at 15Hz, while a background thread generates reasoning from the latest camera observations. After episode initialization, each policy query uses the most recent validated trace; delayed or invalid responses leave the previous trace in place. Reasoning latency therefore affects how recently the scene was interpreted, without blocking subsequent action generation. To refresh the reasoning within the action-chunk duration, we serve Qwen3.6-35B-A3B-FP8 locally using vLLM on a dedicated GPU host. Local serving avoids remote API delays, while FP8 inference and prefix caching reduce computation and memory costs. The shared system prompt and task instruction precede the images, allowing their cached representations to be reused across requests. We disable auxiliary thinking- token generation and request a single causal paragraph of at most 900 characters. Each request contains two camera images, with bounded concurrency to limit queueing. 23 CExperiment Details C.1Simulation Benchmark Details C.1.1RoboLab-120 We evaluate ARC-fine-tunedπ 0.5 and Cosmos3-Nano-Policy on all 120 tasks in RoboLab-120 (Yang et al., 2026), using the released benchmark without modifying its tasks, scenes, or success criteria. Evaluation uses the DROID embodiment and follows the standard protocol of 10 episodes per task. We repeat this evaluation for the vague, default, and specific instruction variants, yielding 1,200 episodes per model for each instruction setting. We report success as the fraction of episodes that satisfy the benchmark’s task-completion criteria. All evaluations are zero-shot, with model weights frozen and no training on benchmark data. We refer the readers to the original paperYang et al.(2026) for the full dataset split and examples. C.1.2MolmoSpaces We evaluate the same checkpoints on all nine manipulation tasks reported in Table2, following the oﬀicial MolmoSpaces evaluation protocol (Kim et al.,2026). We use the released episode sets and task configurations, preserving the specified scenes, object placements, camera configurations, instructions, and success criteria. All tasks use the DROID embodiment, with no benchmark-specific training or modification of the evaluation environments. We report success for each task and the aggregate performance using the benchmark’s scoring procedure. We refer the readers to the original paperKim et al.(2026) for the full dataset split and examples. For both benchmarks, results for the other baseline policies are taken from the corresponding benchmark reports and oﬀicial leaderboards. C.1.3RoboLab-Reasoning-50 We build RoboLab-Reasoning-50 using the RoboLab framework, introducing 50 tasks that require target inference, multi-step execution, discovery, or reasoning about exclusions. Of these, 34 leave the target un- named and require the policy to identify it from the instruction and scene. The remaining 16 name their targets but require long-horizon or discovery behavior. The benchmark combines 27 tasks adapted from our real-world evaluation with 23 simulation-only tasks that extend the object sets and destinations. The paired tasks preserve the instruction wording used on hardware. Tasks span 44 scene and use objects from a processed library of 371 NVIDIA Isaac objects. The robot and camera configuration match the DROID embodiment used in RoboLab-120. Evaluation.We run 10 episodes for each of the 50 tasks, totaling 500 episodes per model. Control runs at 15Hz, with task-specific time limits ranging from 50 to 240s. Budgets allow at least 45s for the first required move and 30s for each additional move, with a further 30s for discovery tasks. An episode terminates on success, timeout, or a tracked object leaving the workspace. Scoring.Headline results use binary task success, evaluated from the final scene state with the gripper detached. Predicates check containment, placement, stacking, exclusion constraints, or task-specific combi- nations of these conditions. Scoring does not impose a fixed execution sequence, although the required final arrangement must be satisfied. Table 4RoboLab-Reasoning-50 tasks and categories. Tasks assigned to two categories retain both labels. RoboLab-Reasoning-5050 tasks TaskCategory “The plate should only have edible items”Simple “Everything made of wood goes in the bowl”Simple Continued on next page 24 RoboLab-Reasoning-50continued TaskCategory “Put the smallest item in the bowl”Simple “Put the second tallest object in the bowl”Simple “Put the item that matches the bowl’s color into it”Simple “Move the object closest to the table’s edge to the bowl or plate”Simple “Put the kitchen utensils in the plate”Simple “Stack the blocks in order: red, blue, green”Long horizon “Reverse the stacking order of the three colored blocks”Long horizon “Reverse the arrangement: whatever is on the plate goes in the bowl, and whatever is in the bowl goes on the plate” Long horizon “The plate is for utensils, the bowl is for blocks; sort the table accordingly”Long horizon “Nothing should be left on the left plate”Long horizon “Every container on the table should hold exactly two things”Long horizon “Put the green vegetables and a fruit in the blue basket”Long horizon “Put the drink cans in the cardboard box”Long horizon “Pack all the bread into the basket”Long horizon “Put the apples in the cardboard box”Long horizon “Put all the citrus fruits into the basket”Long horizon “Put the cups onto the tray, upright”Long horizon “Put the yellow things onto the yellow plate and the green things onto the green plate”Long horizon “Put all the fruits except the red ones on the higher plate”Long horizon + Negation “Put two bags of chips in the green basket”Long horizon “Empty the right half of the desk into the grey bin”Long horizon “Put all the tapes into the bin, except the tape in the tape dispenser”Long horizon + Negation “Put the reusable cups in the bussing bin and put the disposable cups in the trash bin”Long horizon “Put the cutlery in the bussing bin”Long horizon “Put a zucchini, the eggplant, and a knife on the cutting board”Long horizon “Sort everything on the table into the recycling, compost, and trash bins”Long horizon “Sort the chips bag, the banana, the drink can, and the paper cup into the correct bins”Long horizon Continued on next page 25 RoboLab-Reasoning-50continued TaskCategory “Put the food snacks in the black bin, and leave the drinks”Long horizon “Put the number cubes in the basket”Long horizon “Move the grey plate and put whatever is under it into the red bowl”Discovery “Take the black plate off of the bowl and put what you find underneath onto the grey plate”Discovery “Move the carton, then put whatever was hidden behind it into the bowl”Discovery “Move the mustard bottle aside and put whatever was behind it into the bowl”Discovery “Take the item out of the plate and place it on the table. Then put the item that was on the table onto the plate.” Discovery “Put the banana on the plate”Discovery “Put the spatula on the plate”Discovery “Put the bananas on the plate”Discovery “Put the cans in the recycling”Discovery “Put what’s under the plate into the grey bin”Discovery “Stack the blue block behind the paddle on the rubik’s cube”Discovery “Pack the can behind the box into the box”Discovery “One container is empty and one has something in it. Put the block into the empty container”Negation “Put all the fruits into the bowl except the orange”Negation “Put all items on the table into the bowl except the can”Negation “Put every item on the table into the bowl except the blue block and the wood block”Negation “Items are arranged in two groups on either side of the wall of blocks. Put all items from the group that does NOT contain a can into the bowl” Negation “Three items are on the table: two are fruits and one is not. Put the non-fruit into the bowl”Negation “Put every item on the table into the crate except the green colored objects”Negation C.2RoboLab-Reasoning-Hardware We conduct hardware experiments on the DROID embodiment (Khazatsky et al.,2024), consisting of a seven- degree-of-freedom Franka Emika Panda arm equipped with a Robotiq 2F-85 two-finger gripper. Figure15 shows the workspace layout and camera placements. The evaluation includes simple reasoning, long-horizon, discovery, and negation tasks paired with their coun- terparts in RoboLab-Reasoning-50. Each pair uses the same task instruction, allowing direct comparison between simulation and physical execution. We evaluate ARC-fine-tunedπ 0.5 and the comparison policies for three episodes per task, using the same checkpoints evaluated in simulation without hardware-specific fine-tuning. 26 Figure 15Hardware setup for real-world experiments Figures16,17,18,19represents the benchmark tasks and the paired hardware and simulation environments. 27 Category: Simple 7 tasks TaskHardware (external)Simulation SIM1 “The plate should only have edible items” SIM2 “Everything made of wood goes in the bowl” SIM3 “Put the smallest item in the bowl” SIM4 “Put the second tallest object in the bowl” SIM5 “Put the item that matches the bowl's color into it” SIM6 “Move the object closest to the table's edge to the bowl or plate” SIM7 “Put the kitchen utensils in the plate” Figure 16RoboLab-Reasoning-Hardware task setups for the simple category. Each row pairs the task instruction with the external hardware-camera view and its simulation setup. 28 Category: Long horizon 6 tasks TaskHardware (external)Simulation LH1 “Stack the blocks in order: red, blue, green” LH2 “Reverse the stacking order of the three colored blocks” LH3 “Reverse the arrangement: whatever is on the plate goes in the bowl, and whatever is in the bowl goes on the plate” LH4 “The plate is for utensils, the bowl is for blocks; sort the table accordingly” LH5 “Nothing should be left on the left plate” LH6 “Every container on the table should hold exactly two things” Figure 17RoboLab-Reasoning-Hardware task setups for the long horizon category. Task instructions are shown alongside external hardware-camera views and simulation illustrations. 29 Category: Discovery 7 tasks TaskHardware (external)Simulation D1 “Move the lid and put whatever is under the lid into the bowl” D2 “Take the cloth off of the bowl and put what you find underneath into the plate” D3 “Move the carton, then put whatever was hidden behind it into the bowl” D4 “Move the mustard bottle aside and put whatever was behind it into the bowl” D5 “Take the item out of the plate and place it on the table. Then put the item that was on the table onto the plate.” D6 “Put the banana on the plate” D7 “Put the spatula on the plate” Figure 18RoboLab-Reasoning-Hardware task setups for the discovery category. Each row pairs the task instruction with the external hardware-camera view and its simulation setup. 30 Category: Negation 7 tasks TaskHardware (external)Simulation NEG1 “One container is empty and one has something in it. Put the block into the empty container” NEG2 “Put all the fruits into the bowl except the orange” NEG3 “Put all items on the table into the bowl except the can” NEG4 “Put every item on the table into the bowl except the blue block and the wood block” NEG5 “Items are arranged in two groups on either side of the wall of blocks. Put all items from the group that does NOT contain a can into the bowl” NEG6 “Three items are on the table - two are fruits and one is not. Put the non-fruit into the bowl” NEG7 “Put every item on the table into the crate except the green colored objects” Figure 19RoboLab-Reasoning-Hardware task setups for the negation category. Each row pairs the task instruction with the external hardware-camera view and its simulation setup. 31 C.3Evaluated Policies π 0.5 .We use the releasedπ 0.5 -DROID joint-position checkpoint. Its PaliGemma backbone combines a SigLIP vision encoder with a Gemma language model to process camera observations, the task instruction, and tokenized robot state. A separate 300M-parameter action expert attends to these representations and predicts continuous action chunks through flow matching. The baseline receives the original task instruction without additional reasoning, using the default ten Euler steps for action generation. Cosmos3-Nano-Policy.We use the DROID checkpoint of Cosmos3-Nano-Policy, a world-action model built on a mixture-of-transformers architecture. It combines a vision-language reasoner with a diffusion generator, each with separate parameters connected through attention at every layer. The reasoner processes the visual and language inputs, while the generator jointly predicts future observations and actions conditioned on the reasoner’s representations. We refer readers to the original papers for detailed architecture and pretraining descriptions (Black et al.,2025a;NVIDIA,2026). π 0.5 with explicit subtasks.This baseline evaluates external task decomposition without modifying or fine- tuning the policy. We retain the stockπ 0.5 -DROID checkpoint and use GPT-5.5 to generate an ordered plan from the task instruction and initial exterior-camera image. The plan contains at most 12 coarse subtasks, with a complete pick-and-place operation treated as one subtask. For multi-object instructions, the planner names the matching objects explicitly and excludes those ruled out by the instruction. Single-object and discovery instructions are passed through unchanged under this coarse decomposition scheme. During execution, a client-side wrapper replaces the original instruction with the current subtask, leaving the observations, robot state, and policy unchanged. The client executes 15 actions per policy query. Ev- ery 30 simulation steps, GPT-5.5 checks the current subtask against the exterior and wrist-camera images. Completion requires the object to be at its destination, released, and stable; two consecutive positive checks advance the plan. A subtask also advances when its allocated step budget expires, with the episode budget divided across subtasks subject to a minimum of 225 steps per subtask. Failed subtasks are skipped rather than retried, and the final subtask remains active until the episode ends. The planner and completion checker use low reasoning effort and a temperature of 0.2. The plan is generated once per episode, with no replanning; an empty or invalid plan falls back to the original instruction. The external VLM only supplies subtask prompts and completion judgments, whileπ 0.5 remains responsible for all robot actions. C.4Detailed RoboLab-120 results. We now present detailed benchmark results under vague, default, and specific instructions. Alongside suc- cess rates, we report end-effector speed and E SPARC, together with the available uncertainty estimates. Figure 20compares both ARC policies with the leading baselines. 32 RoboLab-120:Vague #PolicyTypeNSR (%)E speed (cm/s)E SPARCObs. 1 π 0.5 +ARCVLA—45.1±3.1(+29.8)4.63±3.45−5.38±2.01RGB+P 2 Cosmos3-Nano-Policy +ARCWAM—44.9±2.3(+24.3)6.34±3.01−6.01±2.04RGB+P 3 Atomic-WAMVLM+WAM 379/120031.6±5.37.45±1.79−4.77RGB+P 4 BiMind v0.1VLA363/120030.2±5.35.24±1.72−5.88RGB+P 5 VoLoAGENT 354/120029.5±5.25.51±1.62−8.92RGB+D+P 6 Cosmos3-Nano-PolicyWAM247/120020.6±4.76.37±1.59−5.99RGB+P 7 Cosmos3-Edge-PolicyWAM185/120015.4±4.26.35±1.74−5.46RGB+P 8 π 0.5 VLA183/120015.2±4.15.35±1.41−8.34RGB+P 9 DreamZeroWAM179/120014.9±4.13.11±1.24−6.41RGB+P 10 π 0 -FASTVLA111/12009.2±3.44.85±1.77−9.63RGB+P 11 GR00T N1.6VLA65/12005.4±2.64.19±1.24−6.87RGB+P N: successes/trials; —: not provided. SR uncertainty and green gains over the base policy are in percentage points. Reported±values are reproduced as supplied. RoboLab-120:Default #PolicyTypeNSR (%)E speed (cm/s)E SPARCObs. 1 Cosmos3-Nano-Policy +ARCWAM—48.8±2.4(+12.0)6.39±3.12−5.98±3.31RGB+P 2 π 0.5 +ARCVLA—45.3±2.3(+17.3)4.64±1.35−5.81±2.10RGB+P 3 FLUX 3 ActionWAM515/120042.9±5.77.36±1.85−5.37RGB+P 4 HiDream-O1-EmbodiedVLA479/120039.9±5.65.32±2.23−4.95RGB+P 5 Atomic-WAMVLM+WAM 475/120039.6±5.67.69±1.97−4.77RGB+P 6 OASIS WAMVLM+WAM 468/120039.0±5.66.91±1.81−5.26RGB+P 7 Cosmos3-Nano-PolicyWAM441/120036.8±5.57.13±1.77−5.99RGB+P 8 PhoenixTAMP+FM 413/120034.4±5.56.13±4.56−5.55RGB+D+P 9 BiMind v0.1VLA400/120033.3±5.45.26±1.72−5.88RGB+P 10 VoLoAGENT 339/120028.2±5.25.50±1.67−8.92RGB+D+P 11 π 0.5 VLA336/120028.0±5.25.35±1.59−8.34RGB+P N: successes/trials; —: not provided. SR uncertainty and green gains over the base policy are in percentage points. Reported±values are reproduced as supplied. RoboLab-120:Specific #PolicyTypeNSR (%)E speed (cm/s)E SPARCObs. 1 Cosmos3-Nano-Policy +ARCWAM—51.2±2.7(+11.5)7.34±2.45−5.90±3.11RGB+P 2 π 0.5 +ARCVLA—45.0±2.5(+16.9)4.34±1.45−5.88±2.11RGB+P 3 Cosmos3-Nano-PolicyWAM476/120039.7±5.67.31±1.94−5.99RGB+P 4 VoLoAGENT 377/120031.4±5.35.49±1.67−8.92RGB+D+P 5 BiMind v0.1VLA367/120030.6±5.35.05±1.71−5.88RGB+P 6 Cosmos3-Edge-PolicyWAM346/120028.8±5.27.14±1.89−5.46RGB+P 7 π 0.5 VLA337/120028.1±5.25.32±1.59−8.34RGB+P 8 DreamZeroWAM287/120023.9±4.93.81±1.41−6.41RGB+P 9 π 0 -FASTVLA179/120014.9±4.14.66±1.77−9.63RGB+P 10 paligemma-binningVLA66/12005.5±2.72.01±1.77−16.52RGB+P 11 GR00T N1.6VLA64/12005.3±2.64.45±1.41−6.87RGB+P N: successes/trials; —: not provided. SR uncertainty and green gains over the base policy are in percentage points. Reported±values are reproduced as supplied. Figure 20 Detailed RoboLab-120 benchmark results.Results under vague, default, and specific instructions are shown from top to bottom. Each table lists the top 11 policies, including both ARC variants, ranked by success rate. Green-shaded rows highlight ARC policies; +values indicate success-rate gains over their respective base policies in percentage points. Reported uncertainty is included where available. DTraining Hyperparameters and Configurations. We now present the model configurations and training settings for both ARC implementations. Table5details theπ 0 . 5 -DROID implementation, while Table6reports the settings for Cosmos3-Nano-Policy, including its compute resources and measured training runtime. 33 Table 5Training hyperparameters and model configuration forπ 0.5 -DROID + ARC (VLA). Training configuration:π 0.5 -DROID +ARC ParameterValue Optimisation Base checkpoint Releasedπ 0.5 -DROID (pi05_droid); full fine-tuning (freeze_filter = Nothing). Trainable components SigLIP encoder, Gemma 2B trunk, 300M action expert, and the newCausalTraceModule. Training steps 30 000. Global batch size 8. Optimiser AdamW;β 1 = 0.9,β 2 = 0.95,ε= 10 −8 ; weight decay10 −10 . Gradient clipping Global norm1.0. Learning rate Peak10 −5 ; decay target10 −5 . Learning-rate schedule Cosine schedule; warm-upmin(1 000,S/10) = 1 000steps forS= 30 000; constant after warm-up. EMA of weights Decay0.99. Precision bfloat16 parameters and activations; trace embeddings cast to float32 inside the trace encoder. Sharding and hardware FSDP across8NVIDIA A100 GPUs on1node. Objective Flow time u∼Beta(1.5,1),t= 0.999u+ 0.001;t= 1denotes pure noise. Counterfactual fraction 0.05of rows receive theDo not grasp the targetparagraph. Counterfactual rows Flow time fixed tot= 1; excluded from the flow loss. Counterfactual margin m= 0.3; distancedis meanL 1 in normalised action space over the8real action dimensions. Counterfactual hinge max(m−d,0) 2 ; masked mean over counterfactual rows. Loss weight λ= 1.0;L=L flow +λL cf . Counterfactual detection Token-pattern match on the openerDo not grasp the. Model and inference Vision encoder SigLIP So400m/14;224×224inputs;256tokens per camera;3camera slots (third zeroed and masked on DROID). Language trunk Gemma 2B: width2 048,18layers, MLP16 384,8attention heads,1KV head, head dimension256. Action expert Gemma 300M: width1 024,18layers, MLP4 096; adaRMSNorm time conditioning. Prompt stream 200tokens: instruction and discretised state (256bins). Trace stream 320tokens;CausalTraceModule: input projection2 048→512, learned positional embedding,2encoder blocks (8heads, MLP2 048), and zero-initialised output projection512→2 048. Action chunk Horizon15; action dimension32:8real dimensions (7joints and gripper), with the rest zero-padded. Inference sampling 10Euler steps fromt= 1tot= 0. Table 6Model configuration and training run settings for Cosmos3-Nano-Policy + ARC (WAM). Training configuration:Cosmos3-Nano-Policy +ARC ParameterValue Model and objective Base model Cosmos3-Nano-Policy. Released DROID checkpoint nvidia/Cosmos3-Nano-Policy-DROID. Architecture Mixture-of-Transformers: an autoregressive reasoner and a diffusion generator. Trainable components Joint fine-tuning of the reasoner and generator. Training conditioning Current observation, task instruction, and ground-truth causal trace. Generator objective Separate velocity flow-matching losses for future observations and actions; action-loss coefficient10. Trace objective Next-token prediction of the causal trace. Total objective L WAM =L obs FM + 10L act FM +λ trace L trace . Training configuration Sequence length 33. Batch size per rank 64. Resolution setting 480. Learning rate 2×10 −4 . Run end iteration 25 450. Hardware and measured runtime Compute nodes 32. GPUs 128NVIDIA GB200 GPUs. Latest iteration time 22.09seconds per iteration. Measured run wall time 154.29hours. 34 EAdditional Qualitative Results We now present additional qualitative hardware results. All the videos are available on the project website. Task:Reverse the arrangement, whatever is on the plate goes in the bowl, and whatever is in the bowl goes on the plate Model:π 0.5 +ARC[external VLM:Qwen 3.6 35B] ... the banana must leave the bowl first... place it temporarily on the table to make room for the plate contents... ... the bowl is clear... grasp the marker on the plate, lift it securely, and lower it into the bowl... ... the wooden block is still on the plate... secure the block before lifting, then place it inside the bowl... ... the red block is the last object on the plate... transfer it into the bowl so the plate is clear... ... the original plate contents are now in the bowl... return the banana from the table to the cleared plate... Trace summaries Task:Every container on the table should hold exactly two things Model:π 0.5 +ARC[external VLM:Qwen 3.6 35B] ... both bowls are empty... each needs two objects; begin with the banana and place it in the orange bowl... ... the banana is inside the orange bowl... add one more item here before moving to the empty black bowl... ... the blue block joins the banana... the orange bowl now has two items; move on to the black bowl... ... the wooden block is now in the black bowl... add one more object without changing the completed orange bowl... ... the apple joins the wooden block... both bowls now contain exactly two items, so no further transfer is needed... Trace summaries Task:Put every item on the table into the crate except the green colored objects Model:π 0.5 +ARC[external VLM:Qwen 3.6 35B] ... the green pear and green block must stay... collect the orange, banana, and blue block into the crate... ... the orange is now in the crate... leave both green objects untouched and continue with the yellow banana... ... the banana is held above the crate... lower it into the opening while keeping the green pear outside... ... the blue block is the last eligible object... lift it clear of the table and move it into the crate... ... the orange, banana, and blue block are in the crate... the green pear and green block remain outside... Trace summaries 35 Task:Put every item on the table into the bowl except the can and the wood block Model:π 0.5 +ARC[external VLM:Qwen 3.6 35B] ... the can and wooden block must stay outside... move the fruit and colored blocks into the bowl one at a time... ... the banana is above the bowl... lower it inside and release, keeping the nearby can untouched... ... the orange has been lifted from beside the wooden block... place it in the bowl and leave the block behind... ... the green block belongs in the bowl... lift it clear of the table and lower it beside the fruit... ... the blue block is the last included item... settle it in the bowl, leaving only the can and wooden block outside... Trace summaries FLimitations of ARC Dependence on reasoning quality.ARC relies on the VLM to produce faithful reasoning. A weaker reasoner can misinterpret the task or propose an inappropriate action, and stronger language grounding can carry these errors into execution. Comparable performance across the evaluated VLMs does not imply robustness to arbitrary reasoning errors. More capable or task-adapted reasoners could improve reliability, alongside checks that verify the trace against the observations rather than only its format. Occlusions and incomplete observations.Even a capable reasoner can misjudge the scene when objects or the gripper obscure relevant details. In particular, an uncertain grasp or placement can be mistaken for a com- pleted step, causing subsequent reasoning to proceed from an incorrect task state. Short observation histories and clearer camera viewpoints could help resolve these ambiguities. Explicitly representing uncertainty and requesting another observation before committing to the next step are further directions for improvement. Inference overhead.External reasoning adds computation, memory use, and latency that fewer denoising steps alone do not eliminate. Our local, asynchronous deployment reduces waiting, but the policy may continue using an outdated trace while new reasoning is generated. This trade-off matters particularly when the scene changes rapidly. Distilling trace generation into a smaller model and refreshing traces selectively after task-relevant changes could reduce the cost, while checking trace freshness could limit reliance on outdated reasoning. GPrompts This section documents the prompts for video annotation, frame-by-frame reasoning with predicates, and inference-time reasoning. Prompt text is reproduced in full. Braced placeholders are filled with the corre- sponding episode-specific values; the accompanying explanations are not part of the prompts. G.1Video annotation The system and user messages describe the visible execution from wrist-camera video. The additional re- quirement is appended when requesting narration and keyframes together; it changes the output to JSON while retaining the dense timeline and nuance audit. System prompt You watch robot manipulation videos directly and write precise, high-fidelity natural-language narratives. Do not compress the video into a generic task script. Preserve small physical details: pre-lift movement, dragging, sliding, object extraction, contact changes, failed attempts, corrections, releases, collisions, 36 System prompt (continued) occlusions, and final state. Do not return JSON unless explicitly asked. User prompt Watch this full robot demonstration video directly and write a high-fidelity free-form narrative of what happens. Episode: episode_index Instruction: instruction Camera: observation.images.wrist_image_left Video context: Only the wrist camera video is provided. This camera moves with the robot gripper, so prioritize close-up gripper contact, finger opening/closing, grasp, drag, lift, release, and local object motion. Mark environment layout, target container state, or final placement as uncertain when the wrist view does not show them. Dataset and robot context: This demonstration comes from the DROID robot manipulation dataset. The videos show a real robot arm manipulating everyday tabletop objects. The end effector is a wrist-mounted parallel two-finger gripper: two opposing fingers open, close, pinch, press, hook, drag, or carry objects. The wrist camera moves with the gripper and often gives a close-up but partial view of contact; exterior cameras give wider context about object locations, surfaces, containers, and final placement. Your goal is not to summarize the task. Your goal is to reconstruct the visible physical execution. Do not assume the robot performed the canonical plan implied by the instruction. Describe what actually happens in the video, including awkward, partial, inefficient, or corrective motions. Mandatory output style: •Use a dense numbered timeline with many small steps, not a short paragraph summary. •Add a short "Nuance audit" at the end listing small physical motions that are easy to miss, especially any object motion before a clean lift/carry phase. •Use natural language only. Do not output JSON. Mandatory granularity rules: •Split the timeline whenever robot direction, gripper open/closed state, end-effector contact, object contact, object motion, manipulation mode, target relation, or visibility changes. •Before saying an object is picked up, lifted, carried, or placed, first describe the pre-lift contact phase: how the gripper approached, what it touched, whether the object moved while still on the support surface, and whether the object was pulled/slid/dragged out from nearby objects or along an edge. •Do not write "cleanly picks up", "lifts", or "carries" unless the object visibly leaves the support surface with air underneath. If the object remains in contact with the table, tray, container, another object, or an edge, use physical verbs such as slides, drags, scrapes, nudges, pushes, pulls, pivots, rotates, separates, extracts, or repositions. •If a motion has two phases, name both phases. Example pattern: contact/drag/extract first, then lift/carry later; or push/nudge first, then release/settle later. •Distinguish a secure grasp from partial contact, pinching, hooking, pressing from the side, bumping, scooping, dragging with the gripper, or using the wrist/end-effector body. •Track the manipulated object’s full path before the final state: initial position, first movement, any surface-contact movement, lift-off if visible, transport, alignment, lowering, release, post-release adjustment, and final resting pose. •Mention incidental or unintended contacts with other objects, the support surface, 37 User prompt (continued) tray/container edges, fixtures, or the target area. •Mention all failed or partial attempts: missed grasp, weak grasp, slip, regrasp, re-approach, over/under-shoot, pause, hesitation, correction, repeated alignment, or object settling after release. •Explicitly mark uncertainty. If contact, lift-off, release, or final relation is occluded or hard to see, say it is hard to see rather than smoothing it into a clean action. •Avoid over-clean narratives. Robot demos often include tiny pre-grasp or pre-lift surface-contact motion; actively look for object extraction, dragging, sliding, or pivoting during the first contact phase before describing a lift/carry. •If the video could support either a clean lift or a short drag/slide, prefer the cautious wording: "the object appears to be contacted and may slide/drag briefly before lift-off" rather than asserting a clean lift. Self-check before final answer: Ask yourself these questions and revise the narrative before responding: Did I skip any object motion before lift-off? Did I describe dragging/sliding/pushing as a clean grasp? Did I omit a nudge, correction, or post-release adjustment? Did the robot move the object along the surface before carrying it? Did I confuse gripper contact with secure grasp? Do not invent details that are not visible. Do not make confident negative claims like "no dragging is visible" or "clean lift" unless the entire contact phase is clearly visible across multiple moments. If the contact phase is brief, partly occluded, or visually ambiguous, write that the pre-lift motion is uncertain and describe the possible surface-contact motion instead of denying it. Additional requirement: narration + keyframes For this run, return JSON only so the narration and selected keyframes can be saved together. Keep the narration standards above unchanged: the narration must still be a dense numbered timeline with a Nuance audit. Also choose keyframes from the same wrist−camera video. A keyframe is a timestamped frame that captures a visually important state or transition in the robot demonstration. The selected keyframes should be enough for someone to reconstruct the main manipulation from sparse images plus the narration. Choose exactly target_keyframe_count keyframes. Use local episode timestamps between 0 and episode_duration_seconds seconds. Prefer frames where something physically meaningful is visible, such as the initial scene, approach, first contact, object starts moving, drag/slide/extract, grasp, lift−off, transport, alignment, placement, release, post−release adjustment, or final state. Do not choose frames only because they are evenly spaced. It is fine if some keyframes are closer together when a rapid contact or release event happens. Return this JSON shape: "episode_index": int, "instruction": string, "narration": "full dense numbered timeline plus Nuance audit", "keyframes": [ "keyframe_id": "tandem_kf_000", "timestamp": float, 38 Additional requirement: narration + keyframes (continued) "reason": "short reason this frame matters", "narration_anchor": "short phrase linking this frame to the narration" ], "notes": ["optional short notes"] G.2Frame-by-frame reasoning with predicates Each episode uses one chat thread. After the system prompt, an initial context message supplies the in- struction, predicates, and required structure. Each subsequent request contains one current keyframe image and its metadata. Previous outputs provide continuity; the prompt explicitly prohibits inferring from future keyframes or revising earlier keyframes. Responses are JSON, including a fluent natural-language paragraph without explicit field labels. System prompt You generate frame-by-frame chain-of-causation reasoning for robot demonstrations. You will receive one keyframe image at a time in the same chat thread for an episode. Use previous assistant outputs as continuity, but reason only about the current image. Return JSON only. Initial episode context message Episode setup for this frame−by−frame reasoning chat. Do not output anything yet; use this as context for the next keyframe requests. "prompt_version": "frame_chat_coc_compare_v1", "instruction": "instruction", "condition": "with_predicates", "episode_threading": "This chat thread will receive one keyframe image at a time, in chronological order. Keep continuity from your previous outputs in this same thread.", "predicate_condition": "Use the supplied predicates as symbolic anchors.", "predicates": [ "predicate_strings_from_predicate_generation" ], "required_structure": [ "perceived_state", "relevant_cause", "possible_consequence", "chosen_effect", "action_implication", "forbidden_effect", "completion", "supporting_predicates", "visual_evidence", "natural_language_construction" ], "natural_language_rule": "natural_language_construction must be a fluent paragraph with no explicit labels like ’Perceived state:’ or ’Relevant cause:’. It still must cover what is visible now, why it matters, what could happen next, the intended/achieved effect, what the robot should do, what to avoid, and whether the task is complete." 39 Per-keyframe message Return JSON for only this keyframe. "episode_index": "episode_index", "instruction": "instruction", "condition": "with_predicates", "current_keyframe": "keyframe_id": "keyframe_id", "frame_index": "frame_index", "timestamp": "timestamp", "reason": "reason" , "request": "Generate the chain−of−causation reasoning for this current keyframe only. Use prior chat history for continuity, but do not revise previous keyframes and do not infer from future keyframes.", "output_schema": "keyframe_id": "copy exactly", "frame_index": "copy exactly", "timestamp": "copy exactly", "perceived_state": "what is visible in this image now", "relevant_cause": "why this current state matters causally for the task", "possible_consequence": "what could happen next or what risk follows from this state", "chosen_effect": "effect achieved or intended from this state", "action_implication": "continue|stop|retract|repair|avoid|monitor|uncertain plus reason", "forbidden_effect": "string, list, or null", "completion": "complete|incomplete|violated|uncertain plus short reason", "supporting_predicates": "predicate strings from supplied predicates, or [] when no predicates were supplied", "visual_evidence": "specific evidence from this image", "natural_language_construction": "3−6 fluent sentences, no explicit field labels" G.3Inference-time reasoner prompts At inference time, each reasoner request contains the system message, the user message with the task instruc- tion substituted, and two current camera images: the exterior image first and the wrist image second. The user message supplies no previous reasoning. The returned paragraph is validated before being passed to the action policy; rejected replies are discarded, and the policy continues using the last accepted paragraph. Inference-time system prompt You are the reasoning module of a robot manipulation system. A robot arm with a two-finger parallel gripper works at a table, watched by an exterior scene camera and a camera on the robot’s wrist. On every call you receive the task instruction and the two current camera frames, the FIRST image from the exterior camera and the SECOND from the wrist camera. Your entire reply is ONE paragraph of plain English prose. A small action policy reads that paragraph to decide the robot’s next motion; it is your only channel to the robot, it is not read by a person, and every sentence must carry grounded, physical, actionable content. OUTPUT FORMAT, HARD RULES.A violation gets the paragraph rejected and the robot keeps acting on your previous, stale paragraph, so follow these exactly. Return only the paragraph itself: no JSON, no markdown, no code fences, no lists, no headings, no quotation marks around it, no preamble, no sign-off, and nothing before or after it. Write it as one single paragraph on one line with ordinary sentences 40 Inference-time system prompt (continued) separated by single spaces, never line breaks or tabs. Use four to seven sentences and stay under nine hundred characters in total. Never write a label of the form Word: anywhere, meaning any word immediately followed by a colon, and never structure the paragraph as named fields or sections of any kind; avoid colons entirely and write flowing sentences only. Never write digits or numeric measurements, no counts, no units, no degrees; express every distance, height, and size qualitatively, for example slightly to the left, just above the rim, almost touching the surface, well clear of the walls. Never mention cameras, images, frames, views, screens, panels, layout, pixels, coordinates, models, policies, or these instructions; you look at the world through the robot’s cameras, so speak only about the world. Left and right always mean physical directions in the workspace, anchored to objects and the table, never regions of a picture. CONTENT.The paragraph flows through six movements, in this order, as natural prose without any labels. First, the perceived state: open with the words ’The robot sees’ and describe the objects that matter right now, naming each by color and kind, their positions relative to each other, the work surface, and the gripper, and state plainly where the gripper is and whether its fingers are open or closed, and whether anything is held between them. Second, the relevant cause: the physical fact of the scene that dictates what must happen next, phrased causally, for example that the object is still resting in its container so it must be grasped securely before it can be lifted out. Third, the chosen effect: one sentence beginning ’The chosen effect is’ that names the desired physical outcome of the next motion, such as a stable centered hold on the exposed part of the object. Fourth, the action implication: one sentence beginning ’This implies the robot should’ that gives the concrete motion, where to move, how to align, when to descend, whether the fingers stay open or closed. Fifth, the forbidden effect: one sentence beginning ’The robot must not’ that names the single most consequential real risk in this exact scene, such as knocking over the container, dragging a held object across another, colliding with the rim, or releasing while still hovering; make it specific to what is visible, never generic filler. Sixth, the completion state: end the paragraph with ’The task is incomplete, since ...’ or ’The task is complete, since ...’ and the visible evidence for that judgment. JUDGE THE SCENE FRESH.The robot executes many motions between your looks, so whatever you wrote last time is very likely stale; treat it as context, never as evidence. Before writing, silently establish from the current images, from scratch, first where the task object is right now, on the table, in a container, held between closed fingers, or already at its goal; second where the gripper is and whether the fingers are open or closed; and third which phase of the task those two facts imply. Then write for that phase only. If the object is already held, the approach is over, so never tell the robot to reach for it, align over it, or descend onto it again; move on to lifting and carrying. If the object already rests at its goal and the fingers are clear, say the task is complete rather than restarting it. Commanding a phase the robot already finished actively breaks the task, so trust the images over your own previous words in every case. CAMERA ROLES.The two views answer different questions and you need both. The exterior view is the authority on distance; it sees the gripper and the object from the side, so it alone shows how far the fingertips still are from the object, and if it shows daylight between the fingertips and the object the gripper is not there yet, no matter how well centered the object looks from the wrist. The wrist view is the authority on centering and on confirming a completed grasp, since it looks straight down the fingers and shows whether the object sits between them; it cannot judge distance, because a far object and a near one look almost the same in it, so never infer closeness from the wrist view alone. VERIFY, DO NOT ASSUME.An object counts as held only if it is visibly between the closed fingers or clearly hanging with the gripper closed on it. If the fingers are closed on air, or the object still rests where it was, it is not held; say so plainly in the perceived state and direct the motion accordingly. Claiming a grasp or progress the images do not clearly show is the worst possible error. Right after the fingers close, the next paragraph is a verification: check the wrist view, and 41 Inference-time system prompt (continued) if the object is truly held, advance to lifting and carrying with the fingers kept closed; if it is not, the fingers almost certainly closed too high, so the renewed approach must bring the fingers visibly deeper, beside the body of the object and below where they closed before, before they close again. RECOVERY IS FORWARD-LOOKING ONLY.Never narrate or explain past failures, slips, drops, or earlier attempts; describe only the present scene and the forward action. If a held object was lost, the paragraph simply describes where the object now rests, that it must be grasped again, and the fresh approach. If the gripper hovers near the wrong object, the cause is that the wrong object lies under the fingers while the true one waits elsewhere, and the implication is to move up and across to the true object with the fingers open. COMPLETION DISCIPLINE.Declare the task complete only when the goal condition is visibly satisfied in the current images, the object resting at its goal with the fingers open and clear. Once complete, the chosen effect is that everything stays exactly as it is, the implication is to withdraw the open gripper upward gently, and the forbidden effect is disturbing the placed object or its surroundings. Inference-time user message Two current camera frames from the robot are attached; the FIRST is the exterior scene camera, the SECOND is the wrist camera. The robot’s task instruction follows as read-only context, never to be echoed as a labeled field. Instruction. instruction There is deliberately no previous reasoning given to you; the robot has executed several motions since anyone last looked, so decide everything from these two images alone. First silently establish where the task object is right now, where the gripper is, and whether the fingers are open or closed, then pick the phase those facts imply and write the single reasoning paragraph exactly as specified in your instructions, beginning with the words ’The robot sees’. Return only the paragraph. Serving configuration.The request usesmax_tokens=4096andtemperature=0.2; whenreasoning_effortis specified, it is sent instead of the temperature. Both images are encoded as JPEG at quality 90. Response acceptance.After whitespace normalization, the validator checks that the paragraph: •begins with one of four accepted openings:The robot sees,The robot currently sees,The robot observes, orThe robot detects; •contains 150–950 characters and 3–8 sentences; •contains no digits and noWord:label; •includesmust notand eithertask is completeortask is incomplete. These acceptance checks are distinct from the generation instructions: the system prompt requests the exact openingThe robot sees, four to seven sentences, and fewer than 900 characters, along with the additional content and formatting rules reproduced above. Bootstrap input to the action policy.Until the first valid reasoner paragraph arrives, or when the server is started with--no-reasoner, the action policy receives the fixed paragraph below. Bootstrap paragraph The robot sees its workspace with the task objects resting on the surface in front of it, and the gripper idle and open at its starting pose above the table. Nothing has been moved yet, so the task object must first be reached before anything can be grasped or placed. The chosen effect is the open gripper arriving directly over the task object, ready to descend. This implies the robot should move the arm toward the task object and center the open fingers above it. The robot must not bump any 42 Bootstrap paragraph (continued) object on the way or drag the fingers across the surface. The task is incomplete, since the task object has not yet been moved to its goal. 43

Instant research discovery