acceptodds
Under review as a conference paper at ICLR 2027

ARC: A Reasoning Recipe for Robot Foundation Models

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 . 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 from DROID without collecting new robot data. Third, we show how state-of-the-art VLAs such as 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 's task success by 82.2 percentage points.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot foundation models, Embodied reasoning, Robot manipulation, Robot learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_R3RQOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARC, a reasoning-based fine-tuning recipe for pretrained robot foundation models. The key idea is to introduce action-grounded causal reasoning traces that connect the current state, intended effects, and robot actions. The authors develop an automatic annotation pipeline to generate these traces from existing DROID demonstrations and use the resulting supervision to fine-tune π0.5 and Cosmos3-Nano-Policy.

Critical strengths

ARC achieves substantial improvements on both π0.5 and Cosmos3-Nano-Policy across multiple benchmarks, particularly on reasoning-intensive tasks and underspecified instructions. The real-robot evaluation further demonstrates that these improvements are not limited to simulation. These results provide convincing evidence that reasoning-conditioned control can substantially improve existing robot policies.

The authors introduce an automatic relabeling pipeline that augments existing DROID demonstrations with action-aligned causal reasoning traces. The ability to improve pretrained robot policies using existing data and relatively lightweight fine-tuning makes the approach practical

Critical weaknesses

The idea of using reasoning to improve robotic controal has been explored in prior work like ECoT, however the paper does not compare it with ECoT task success rate. Also the counterfacutual loss for π0.5 to prevent the policy from ignoring reasoning traces is not directly ablated (with and without this objective)

Questions for authors

Could the authors report the performance of π0.5 + ARC with and without the counterfactual objective?

Could the authors report the downstream task success of the ECoT-style fine-tuned baseline already introduced in Figure 2, rather than comparing only attention statistics?

3 comments

We thank the reviewer for their thoughtful assessment and for recognizing ARC's substantial improvements across simulation and real-world benchmarks. We are encouraged by their appreciation of the action-grounded reasoning formulation and the practical benefits of improving pretrained robot policies using existing demonstrations. We appreciate the two suggestions and have addressed both through additional experiments in the updated manuscript. We summarize the results below.

Q1 and Weakness 1. Ablation of the counterfactual objective

We thank the reviewer for pointing out this missing ablation and apologize for its omission in the original submission. We have now added an explicit ablation of the counterfactual objective for π₀.₅ + ARC in Appendix B.1.1 (Page 22) and reproduce the results here for convenience.

Under the same training and evaluation settings, we obtain the following results across all three instruction settings on RoboLab-120:

MethodVagueDefaultSpecific
π₀.₅ (base)15.2%28.0%28.1%
π₀.₅ + ARC w/o counterfactuals36.1%34.2%37.8%
π₀.₅ + ARC45.1%45.3%45.0%

As discussed in Section 4.3, the counterfactual objective in Equation (2) is designed to discourage the action expert from ignoring reasoning traces. It introduces contradictory traces during training and penalizes action predictions that remain too close to the demonstrated actions despite the conflicting reasoning. The new ablation directly measures the contribution of this objective to downstream task success and complements the existing analysis of reasoning-token attention and trace-component ablations.

Q2 and Weakness 2. Downstream task success of the ECoT-style baseline

We appreciate this suggestion and agree that reporting downstream success for the ECoT-style baseline introduced in Figure 2 would provide a useful comparison. We have now extended the evaluation to include task success for this baseline and added the results to Appendix C.4.1 (Page 35).

Given the limited rebuttal timeline, we conducted this additional evaluation on a diverse subset of 40 tasks from RoboLab-120. The tasks were selected to cover different manipulation and reasoning requirements, providing a broader comparison across task types. We have included the exact task list and evaluation details in the appendix for transparency and reproducibility.

Importantly, the ECoT-style baseline uses the same π₀.₅ backbone and demonstration dataset as ARC, as described in Section 4.1. This allows us to examine how the choice of reasoning representation influences downstream control while keeping the underlying policy architecture and training data consistent.

On the evaluated subset, we obtain the following task-success results:

MethodSuccess rate
π₀.₅ (base)21.1%
π₀.₅ + ECoT-style25.1%
π₀.₅ + ARC49.8%

These additional results complement Figure 2, which examines how the action expert attends to intermediate reasoning under different supervision formats.

Thank you for the detailed response and additional experiments. My concerns are mostly addressed so I raise to weak acceptance

Thank you for taking the time to consider our responses and additional experiments. We are glad that we could address most of your concerns, and we sincerely appreciate your feedback and continued support for our work.

Reviewer_eUJFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ARC, a way to add language reasoning to existing robot foundation models without collecting new robot data. The authors define an "action-grounded causal" trace, use Gemini and GPT-5.5 to automatically label about 75K DROID episodes with such traces, and fine-tune two different models on them: Pi0.5 and Cosmos3-Nano-Policy. At test time an external VLM writes the traces from the current camera images. Both models improve a lot over their base checkpoints on RoboLab-120 and MolmoSpaces, on a new reasoning-focused suite (RoboLab-Reasoning-50), and on a set of real-robot tasks.

Critical strengths
  • The recipe is simple and works on two different model families. The same labels lift both a VLA and a WAM, and the gains hold on benchmarks the authors did not design (RoboLab-120 and MolmoSpaces), not just on their own suite. The robustness to vague instructions was the most interesting result for me: with ARC, Pi0.5 stays at roughly the same success rate from specific to vague instructions, while the base model drops by almost half. It also helps that an open-weight VLM works about as well as GPT-5.5 as the trace generator, so the system does not depend on a closed API.

  • The relabeled DROID dataset is a useful contribution. If it is released as promised, others can build on it, and the appendix gives enough detail (prompts, keyframe selection, training configs) to reproduce the pipeline.

Critical weaknesses
  • It is not fully clear how much of the gain comes from the causal trace itself rather than from having a strong VLM look at the scene about once per second. Most baselines are plain policies with no VLM in the loop, and the one VLM-based baseline (Pi0.5 with explicit subtasks) is fairly weak: the plan is made once and some instructions are passed through unchanged. A fairer comparison would be the same VLM, at the same refresh rate, giving stock Pi0.5 a subtask or a more explicit instruction. Likewise, the claim that causal traces beat ECoT or subtask labels is mainly supported by the attention analysis, not by task success. I don't think this undermines the paper, but it limits how strongly the main claim about the reasoning content can be stated.

  • Some parts of the evaluation make the size of the gains hard to judge. Baseline numbers come from leaderboards while ARC is run in the authors' own setup. The inference prompt includes hand-written notes about the controller's failure modes, and the paper does not say which tasks were used to tune it.

Questions for authors
  • Were the base Pi0.5 and Cosmos models re-evaluated in your own harness, or are those numbers copied from the leaderboard?

  • How was the inference prompt developed, and was any of it tuned on the evaluation tasks?

  • The efficiency claims (fewer denoising steps, fewer training updates) may take into account the cost of running the external VLM and of labeling the data.

5 comments

Comment (1/2)

We thank the reviewer for their thoughtful assessment and for recognizing ARC's strong performance across model families, particularly its robustness to underspecified instructions and its improvements on independently developed benchmarks. We also appreciate their recognition of ARC-Trace-DROID as a useful contribution to the community. We address the concerns below and highlight the additional experiments and clarifications included in the revised manuscript.

Weakness 1: How much of the gain comes from the causal trace itself?

We thank the reviewer for raising this question. The paper includes several experiments that directly examine the contributions of external VLM planning, causal reasoning supervision, and policy adaptation. We highlight three complementary comparisons below.

First, Table 3 explicitly evaluates an external VLM-in-the-loop baseline without ARC fine-tuning. As detailed in Appendix C.3, the original π₀.₅ policy is guided by a strong VLM (GPT-5.5), which generates a subtask plan and actively monitors execution through visual feedback, advancing through subtasks based on observed progress. This provides closed-loop VLM guidance throughout execution. On RoboLab-Reasoning-50, this baseline improves success from 10.2% to 21.0%, whereas π₀.₅ + ARC achieves 57.0%, exceeding the externally guided baseline by 36.0 percentage points. This substantial difference demonstrates that ARC's improvements extend well beyond those achieved by providing the original policy with strong external VLM planning.

Second, Figure 9(b) directly investigates the contribution of causal reasoning through separately fine-tuned policies with different trace representations. Full ARC achieves 45.30% success on RoboLab-120 (default). Removing Cause & Effect while retaining the remaining components, including explicit action guidance, reduces success to 39.25%. Action-only supervision achieves 27.42%, comparable to the 28.00% base policy, while adding State to Action improves success to only 32.25%. These results demonstrate that richer action descriptions alone are insufficient to reproduce ARC's gains, and that the causal components make an additional contribution to downstream performance.

Third, Figure 9(a) demonstrates that ARC maintains comparable performance across different external reasoners, indicating that its improvements are consistent across VLM choices. Figure 9(c) further establishes the importance of policy adaptation, with full fine-tuning substantially outperforming action-head-only or generator-only fine-tuning.

Regarding the comparison with ECoT, we thank the reviewer for suggesting a downstream task-success evaluation to complement the attention analysis in Figure 2. We have now added this experiment to Appendix C.4.1 (Page 35). Given the rebuttal timeline, we evaluate a diverse subset of 40 RoboLab-120 tasks spanning 33 scenes, covering a broad range of manipulation and reasoning requirements. The complete task list and evaluation protocol are provided in the appendix.

On this subset, the original π₀.₅ achieves 21.1% success, while ECoT-style fine-tuning improves success to 25.1%. In comparison, π₀.₅ + ARC achieves 49.8%, outperforming ECoT-style supervision by 24.7 percentage points. Both methods use the same π₀.₅ backbone, underlying demonstration dataset, and full fine-tuning setup, allowing a direct comparison of their reasoning supervision formats. These additional results provide behavioral evidence that ARC's action-grounded causal traces offer substantial advantages over ECoT-style reasoning supervision.

Q1 and Weakness 2. Benchmark evaluation and provenance of baseline results

We thank the reviewer for raising this question and appreciate the opportunity to clarify the evaluation protocol. The baseline results for RoboLab-120 and MolmoSpaces are taken from their respective official benchmark reports and leaderboards, as already stated in Appendix C.1. These benchmarks were developed and evaluated by research teams independent of the ARC evaluation, and the reported baseline results include the original π₀.₅ and Cosmos3-Nano-Policy checkpoints.

In particular, the original RoboLab benchmark paper (Yang et al., 2026) and its official leaderboard report evaluations of both π₀.₅ and Cosmos3-Nano-Policy using the benchmark's standardized protocol. Similarly, the MolmoSpaces benchmark (Kim et al., 2026) maintains an official leaderboard that includes these models and their task-level performance. These reference results are therefore independently established benchmark measurements rather than baselines constructed or selectively evaluated for ARC.

Comment (2/2):

For our ARC-fine-tuned models, we use the released benchmark environments and official evaluation protocols without modifying their tasks, scenes, or success criteria. We believe that comparison against the independently established benchmark results provides a meaningful and reproducible assessment of ARC's improvements. We have further clarified the provenance of the baseline scores and our evaluation procedure in the revised manuscript (Page 25).

Q2. Development of the inference prompt and evaluation-task tuning

We apologize for the confusion regarding the inference prompt and have clarified its scope in the revised manuscript (Page 46). The hardware-specific notes highlighted by the reviewer are used only for real-robot deployment. They are not generally required. Additionally, the simulation evaluations use the general ARC reasoning prompt, without additional instructions. The hardware notes account for practical differences in the physical robot setup, particularly the positioning of the external and wrist cameras, gripper geometry, and the resulting visibility and manipulation constraints. These details help the external reasoner correctly interpret observations from the deployed embodiment.

Importantly, these describe hardware-specific constraints. As described in the appendix, they are configurable deployment parameters that users can adapt to their own robot embodiment, camera placement, and gripper configuration.

We have explicitly distinguished the general inference prompt from the hardware-specific deployment notes in the updated manuscript (Page 46) to avoid further ambiguity.

Q3. Efficiency claims and external VLM overhead

We thank the reviewer for this comment and apologize for the confusion regarding our efficiency claims. The results in Figures 7 and 8 specifically measure the efficiency of the underlying action policy, in terms of denoising steps required for action generation and training updates required to reach a given success rate. These experiments are intended to evaluate the improvements in policy-level efficiency, rather than end-to-end system latency, which also includes external VLM inference.

We have clarified this distinction in the revised manuscript (Page 9, paragraph: KF#3) and included an additional latency analysis to provide further context on the computational overhead of external reasoning (Page 10, paragraph: End-to-end latency).

Thanks for the response and the new experiments. The ECoT comparison is helpful, though I'd like to know if ECoT also used the same external VLM at test time. The prompt and latency clarifications resolve those questions. I still don't think the explicit-subtask baseline is a matched comparison (one-shot plan, some instructions passed unchanged), and the hardware questions weren't addressed. Overall still a solid paper, so I'll keep my weak accept.

Reviewer_d1ScOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ARC, a reasoning recipe for improving pretrained robot foundation models without collecting additional robot demonstrations. It introduces action-grounded causal reasoning traces that connect the current state, intended effects, and subsequent actions, along with an automatic annotation pipeline built on the DROID dataset. By fine-tuning existing VLA and WAM models with these traces and using external VLMs for inference-time reasoning, ARC achieves substantial improvements across simulation benchmarks and real-robot tasks. The experiments also demonstrate improved robustness to instruction variations and greater training and inference efficiency.

Critical strengths

1.ARC achieves substantial improvements across RoboLab-120, MolmoSpaces, and reasoning-intensive benchmarks. The gains are particularly impressive on RoboLab-Reasoning-50 and real-robot tasks, demonstrating the practical potential of the proposed recipe.
2.The paper identifies a meaningful limitation of existing reasoning supervision for robot control and proposes action-grounded causal traces that explicitly connect the current state, intended effects, and subsequent actions.
3.The framework improves both π0.5 and Cosmos3-Nano-Policy, suggesting that structured reasoning supervision can benefit different classes of pretrained robot foundation models without requiring new robot demonstrations.
4.The ARC-Trace pipeline automatically generates action-aligned reasoning annotations from existing demonstrations, offering a practical alternative to collecting additional robot interaction data.

Critical weaknesses

1.ARC relies on a strong external VLM to generate action-level reasoning during inference. Although the paper includes an explicit-subtask baseline, it does not fully isolate the benefits of ARC fine-tuning from those of online planning and guidance provided by the external VLM.
2.The ablations demonstrate the importance of different trace components, but do not fully distinguish the benefits of causal structure from those of richer and more explicit action supervision. Stronger information-matched baselines would help establish whether causal reasoning is the key factor behind the improvements.
3.ARC-Trace uses full demonstration videos to generate action-aligned reasoning labels, which may contain information unavailable to the policy at the corresponding decision time. The paper does not sufficiently analyze how this train–test information mismatch affects performance.
4.Increased attention to reasoning tokens and the counterfactual training objective suggest stronger language grounding, but do not directly establish that the policy follows the intended causal relationships. More controlled behavioral interventions would strengthen this claim.
5.The reported reductions in training updates and denoising steps do not account for the full cost of offline annotation, fine-tuning, and external VLM inference. A more comprehensive comparison of runtime, GPU memory, and inference overhead would better establish the practical efficiency of ARC.


After rebuttal: The rebuttal is helpful and I choose to keep the score of 3, and encourage the authors to incorporate the discussion and experiments in the final revision.

Questions for authors

1.How much of ARC's improvement comes from the fine-tuned policy versus the online guidance of the external VLM? A comparison using the same VLM and refresh schedule but simpler action-level or subtask-level guidance would help clarify their respective contributions.
2.The trace ablations show clear benefits from the full representation, but how much of the gain is attributable to causal relationships rather than more detailed action descriptions?
3.Since ARC-Trace is generated with access to full demonstration videos, can the authors analyze whether the annotations contain information unavailable at the corresponding decision time? A comparison with traces generated from only current and past observations would be informative.
4.It would be helpful to report the total cost of offline trace generation, fine-tuning, and external VLM inference, including GPU memory and reasoning latency. How do these costs compare with the reported savings in training updates and denoising steps?
5.How robust is ARC to incorrect or stale reasoning traces, particularly in dynamic or partially occluded scenes? Additional experiments on different robot embodiments would also clarify the scope of its generalization.

4 comments

Comment (1/3)

We thank the reviewer for the thoughtful assessment and for recognizing ARC's improvements in simulation and on real robots. Several concerns are addressed experimentally in the submitted paper. We clarify those results and outline the additional analyses below.

Q1 and Weakness 1. ARC fine-tuning versus external VLM guidance

We appreciate this question. Table 3 already evaluates external VLM planning without ARC finetuning. As detailed in Appendix C.3, the baseline uses a strong VLM (GPT 5.5) as an active, closed-loop planner for the original π₀.₅-DROID policy. The VLM decomposes the task into subtasks and continuously monitors execution through visual feedback, determining when the policy should advance to the next subtask.

On RoboLab-Reasoning-50, the base π₀.₅ policy achieves 10.2% success. Explicit-subtask guidance raises this to 21.0%, whereas π₀.₅ + ARC achieves 57.0%. The VLM-in-the-loop baseline therefore improves success by 10.8 percentage points over the original policy, while ARC achieves an additional 36.0-point improvement over this strong planning baseline. This provides direct experimental evidence that the gains from ARC extend substantially beyond those obtained through active external VLM planning alone.

Additionally, fig. 9(a) independently evaluates the same frozen ARC-fine-tuned π₀.₅ checkpoint with different external reasoners and fig. 9(c) additionally establishes the importance of adapting the language components alongside action generation.

Together, these experiments provide complementary evidence for the contribution of ARC fine-tuning. The explicit-subtask baseline evaluates active external planning without policy adaptation, the reasoner ablation demonstrates stability across external VLMs, and the fine-tuning ablation establishes the importance of jointly adapting the language and action components. We have further highlighted these comparisons and their implications in the updated manuscript (Page 7, paragraph: Reasoning-focused evaluation).

Q2 and Weakness 2. Causal reasoning versus richer action descriptions

We appreciate this question. fig. 9(b) already compares separately fine-tuned π₀.₅ policies with different trace representations on RoboLab-120 (default), directly testing the effect of reasoning supervision on learned control.

Considering the subset necessary for this specific rebuttal, full ARC achieves 45.30% success. Removing Cause & Effect reduces success to 39.25%, a drop of 6.05 percentage points, while retaining the remaining components, including explicit action guidance. Action-only supervision achieves 27.42%, essentially the 28.00% base-policy result. Adding State to Action raises success to 32.25%, still 13.05 points below full ARC. In the other direction, removing Action from the full trace reduces success to 30.58%, a 14.72-point decrease. Removing State, Avoid, or Completion yields 37.17%, 35.83%, and 42.33%, respectively.

The implication we are trying to make is that these comparisons are directly relevant to the concern about richer action descriptions. Removing Cause & Effect hurts performance despite retaining explicit action guidance, while Action-only and State + Action supervision fall well short of full ARC. Causal fields identify which state change matters and why, while the action field connects that effect to an executable step (fig. 9(b) therefore provides experimental evidence for the additional contribution of causal content). Adding another descriptive or constraint field does not automatically improve performance. The benefit depends on how the components support the full causal account.

The related concern about causal faithfulness also deserves clarification from our end. Section 4.1 explicitly describes attention as a diagnostic, and we do not treat attention weights as proof of causal behavior. For π₀.₅, Equation (2) adds counterfactual supervision to approximately 5% of training examples, supplying traces that contradict the demonstrated action and penalizing the policy when its resulting prediction stays too close to that action. This discourages the controller from ignoring the reasoning signal.

We have expanded the discussion in the updated manuscript (Page 9, paragraph: Key ablations) to emphasize these findings and clarify how the ablations distinguish the benefits of causal supervision from those of explicit action descriptions.

Comment (2/3)

Q3 and Weakness 3. Full-video supervision and possible future-information leakage

We appreciate this concern and would like to clarify the distinction between offline annotation and online policy execution. ARC-Trace uses complete demonstrations (including the current and future states) to construct action-aligned reasoning supervision, while the deployed system generates its reasoning solely from observations available at inference time. The information provided to the inference-time reasoner and policy is fully specified in Appendices B.3 and G.3. All reported simulation and real-robot evaluations follow this setting, without access to future observations.

Access to the complete demonstration during offline annotation serves an important purpose. It allows the annotation pipeline to identify action boundaries, reconstruct task progress, and associate executed actions with their intended effects. This is particularly useful for generating consistent, action-aligned reasoning labels from existing demonstrations. The use of complete trajectories during supervision construction does not introduce privileged future information into the deployed policy's inference inputs.

Furthermore, the annotation procedure includes explicit safeguards against retrospective reasoning. As detailed in Appendix G.2, keyframes are processed chronologically, earlier annotations are preserved, and the annotator is explicitly instructed not to infer from future keyframes. These constraints encourage reasoning that is grounded in the state and intended action at the corresponding decision point.

We also wish to distinguish between predicting an intended effect and relying on future observations. ARC's causal traces naturally describe how an action is expected to change the environment. Such forward-looking reasoning is central to the proposed representation and can be generated from the current state without observing the future outcome.

While complete-video context supports the construction of action-aligned offline annotations, the deployed reasoner operates exclusively on currently available observations, without access to future frames or demonstration outcomes. The reported performance gains are therefore obtained under the intended online inference setting. We have clarified this distinction in the updated manuscript (Page 6, paragraph: Inference and Page 18, subsection: ARC-TRACE: Data Labeling Details).

Q4 and Weakness 5. Computational costs and inference overhead

We thank the reviewer for this suggestion and apologize for the lack of clarity regarding the scope of our efficiency results. Figures 7 and 8 specifically quantify improvements in action-generation efficiency and training-update efficiency, respectively. We have clarified this distinction and added a detailed end-to-end latency analysis to the main paper (Page 10, paragraph: End-to-end latency analysis). The analysis profiles a complete hardware rollout, showing that external reasoning accounts for approximately 37% of execution time, compared with 9% for policy inference and 50% for physical motion. This provides a more complete understanding of the latency trade-off associated with reasoning-conditioned control.

Regarding computational resources, Appendix D already reports the training configurations and hardware requirements. Specifically, π₀.₅ is fine-tuned for 30,000 steps using 8 A100 GPUs, while Cosmos3-Nano-Policy uses 128 GB200 GPUs, with a measured training runtime of 154.29 hours. Appendix B.3.1 additionally details our optimized external reasoning setup, including local VLM serving using vLLM, FP8 inference, and prefix caching. For offline ARC-Trace generation, we use external VLMs accessed through proprietary APIs. Since these services do not expose their underlying GPU allocation, utilization, or peak memory consumption, accurately estimating their provider-side computational costs is difficult. We have therefore focused our additional analysis on directly measurable training resources and end-to-end inference latency, while making the scope of the reported efficiency improvements explicit.

Comment (3/3)

Q5. Incorrect or stale traces and other robot embodiments

We appreciate the reviewer for highlighting this important limitation. Appendix F already explicitly discusses ARC's sensitivity to reasoning errors, incomplete observations, occlusions, and stale traces. ARC strengthens the dependency between intermediate reasoning and action generation. Consequently, incorrect or outdated reasoning can influence the resulting actions, potentially leading to execution failures. This is a current limitation of the current framework and a natural consequence of making reasoning an influential conditioning signal for robot control. Improving reasoning reliability is therefore an important direction for future work. In particular, verifying generated traces against current visual observations and triggering reasoning updates when the scene changes significantly could help mitigate incorrect or stale guidance.

Regarding embodiment generalization, ARC is evaluated on two distinct robot foundation model families, π₀.₅ and Cosmos3-Nano-Policy, using the DROID embodiment. The consistent improvements across these architectures demonstrate that the proposed training recipe is applicable to different policy formulations, including VLA and world-action model. Extending ARC to additional robot embodiments should be straightforward, as the same ARC-Trace pipeline can be applied to demonstrations from different robot platforms to generate action-grounded reasoning annotations. The underlying reasoning formulation and training recipe remain unchanged, making the approach readily adaptable to other embodiments given suitable demonstration data.

I thank the authors for providing the new results and clarifications. Fine-tuning ARC and implementing causal reasoning indeed yield more substantial performance improvements. As the authors' responses have addressed most of my concerns, I have decided to maintain my positive recommendation.

Comments on the paper

1 comment

We sincerely thank all reviewers for their thoughtful evaluations, constructive feedback, and recognition of ARC's contributions. We are encouraged by their positive assessment of our results and the practical potential of action-grounded causal reasoning for robot control.

We have carefully addressed the questions, concerns, and limitations raised in the reviews, supported by additional experiments, ablations, and clarifications of our methodology and evaluation protocols. These include new comparisons with ECoT-style supervision, an ablation of the counterfactual objective, and an end-to-end latency analysis.

We have updated both the main manuscript and the appendix accordingly, with all revisions and additions highlighted in red to make them easy to identify. We hope these additional results and clarifications address the reviewers' concerns and further strengthen the paper. We would be happy to address any further questions or clarifications the reviewers may have and look forward to continuing the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.