acceptodds
Under review as a conference paper at ICLR 2027

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

Abstract

Vision-Language-Action (VLA) models have made substantial progress in robotic manipulation, but their generalization remains constrained by the diversity and coverage of action-labeled robot demonstrations. Human videos provide much broader behavioral and environmental experience, while learning from them faces a fundamental trade-off between scalable data and explicit action grounding. We introduce ThinkWorld, a VLA learning framework that grounds action reasoning in world dynamics and strengthens it through reinforcement learning with world imagination. ThinkWorld represents each action thought as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that captures the intermediate dynamics connecting the current world to a future state and serves as a common action representation across embodiments. We first initialize CoD with action-labeled demonstrations, then predict the future induced by each CoD and reinforce those whose predictions better match the observed outcome. This enables videos without action annotations to provide direct supervision from observed state transitions, allowing human and robot experience to refine the same reasoning space. Experiments in simulation and on real robots demonstrate effective human-to-robot transfer and strong generalization to unseen tasks and environments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, World Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JpiNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ThinkWorld learns a shared motion-token representation from human and robot data, then uses agreement between predicted and observed future features as a GRPO reward to refine the VLM. The trained VLM supports continuous action generation, while motion-token generation and future prediction are not required at deployment.

Critical strengths
  1. The method offers an interesting way to use videos without action annotations, combining motion supervision with feedback from observed future states.

  2. The real-world generalization results are encouraging. The future-feature visualizations and CoD perturbation experiments provide evidence that the motion tokens help predict relevant scene changes.

  3. The additional prediction branches are used only during training, avoiding their generation overhead during robot execution.

Critical weaknesses
  1. Stronger direct comparisons are needed. I would like to see comparisons with π₀.₅ and Fast-WAM under matched downstream settings, particularly on real robots. Please also report training GPU-hours, including the additional GRPO cost, and inference latency and memory on the same hardware.

  2. The cross-embodiment analysis needs more evidence. The t-SNE plot appears to contain relatively few samples, and the sample count and selection procedure are unclear. Quantitative cross-embodiment retrieval or motion-classification results would make the alignment claim more convincing.

  3. The contribution of CoD and GRPO needs clearer isolation. Please clarify how CoD differs from existing shared or latent action representations, and separately evaluate the benefit of human videos. Since GRPO adds only 1.2 percentage points in the reported ablation, a comparison with continued future-prediction training using the same data and a comparable compute budget would be useful.

Questions for authors

Please see the above weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JpiN

We thank the reviewer for the constructive feedback. We summarize the additional experimental results in the general response and address each point below. Changes addressing your comments are marked in purple in the revised manuscript.

[W1] Comparison with π0.5 and Fast-WAM; training and inference cost → Figs. 1, 3; Tabs. 5, 6, 9, 10

We have added real-robot comparisons with π0.5 and Fast-WAM. Since their released checkpoints were not trained on YAM data, we first adapt both baselines on MolmoAct2-BimanualYAM, then fine-tune them on the same task-specific robot demonstrations as ThinkWorld, with the same amount of task data. All methods use the same evaluation hardware and control settings. ThinkWorld achieves the highest OOD success and outperforms π0.5 overall, with fewer parameters and no proprietary robot pretraining data.

Inference latency and peak memory are measured on the same RTX 3090 GPU at batch size 1.

MethodOOD success, avg. over tasks (%)Latency (ms)Memory (GiB)
ThinkWorld88.94107.6
π0.582.22408.9
MolmoAct246.744711.6
Fast-WAM13.325712.8

In addition, RL accounts for only 4.2% of the total pretraining compute. Since CoD is used only during training, ThinkWorld incurs no inference-time reasoning overhead and runs faster than MolmoAct2.

[W2] Cross-embodiment analysis on CoD space beyond t-SNE → Sec. 4.5 (Tab. 13); App. C.6 (Tab. 14, Fig. 15)

In our submission, Fig. 6 visualizes randomly selected post-CoD </think> representations from human and robot datasets, colored by motion direction.

We have added motion-classification and retrieval evaluations in Sec. 4.5 and App. C.6 (Tabs. 13–14), using 2,398 windows from EgoDex, YAM, and RoboCOIN with four motion directions. With frozen representations, a linear probe trains on two sources and tests on the third; retrieval searches only the other sources.

RepresentationAvg. balanced accuracy (%)Retrieval P@1 (%)Retrieval P@5 (%)
Action-token histogram28.224.526.8
<think> before CoD37.131.028.5
</think> after empty CoD45.532.330.8
</think> after CoD53.944.341.9

Post-CoD representations perform best on every held-out source and in retrieval, supporting transferable motion information across embodiments. Fig. 15 further shows shared motion-direction structure in a subspace learned from human videos, with some dataset-specific structure remaining.

[W3] Contributions of CoD, human videos, and GRPO → Fig. 14; Tabs. 11-12

(1) CoD vs. latent actions. As described in lines 196–202, CoD represents human and robot trajectories as FAST-style motion tokens generated autoregressively by the VLM. Unlike pixel-derived latent actions such as LAPA [1] and UniVLA [2], CoD serves as a shared reasoning sequence whose generation is refined by comparing imagined consequences with observed futures.

(2) Benefit of human videos. With robot data fixed, we vary the human-video share in SFT and apply the same RL stage (Fig. 14). Results below report LIBERO-PRO average success.

Human-video share in SFTBefore RL (%)After RL (%)RL gain
0% (robot only)45.3747.50+2.13
10%46.0447.75+1.71
25%47.4448.84+1.40
50%47.7749.05+1.28
Robot only, same total data scale as 50%46.6——

Human videos improve performance beyond the robot-only control with the same amount of data, and RL improves every variant. On real-robot skill composition, adding human videos raises success from 3/20 to 13/20 with identical robot demonstrations and instructions (Tab. 12), providing full-sequence supervision absent from the atomic robot data.

(3) GRPO vs. continued training. Starting from the same SFT checkpoint and using the same data, we compare 2k additional iterations of continued SFT with future prediction against GRPO (Tab. 11). Continued SFT lowers performance, while GRPO consistently improves it across seeds.

TrainingLIBERO-PRO Avg. (%)
SFT checkpoint54.4
+ 2k continued SFT53.9
+ 2k GRPO, mean ± std over three seeds55.4 ± 0.3

We are happy to discuss any remaining questions.

[1] Latent Action Pretraining from Videos, ICLR 2025

[2] UniVLA: Learning to Act Anywhere with Task-centric Latent Actions, RSS 2025

Thanks for the authors' rebuttal and my questions have been solved. I will retain my positive score.

We are glad that our response has addressed your concerns, and we sincerely thank you for your thoughtful review and engagement.

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

This paper proposes ThinkWorld, a VLA training framework. It unifies robot end-effector actions and human wrist trajectories into a delta-EEF representation, which a FAST-style tokenizer discretizes into motion tokens. These tokens, called Chain-of-Dynamics (CoD), are generated autoregressively by Qwen3-VL-2B. Learnable future queries follow the CoD and predict frozen DINOv3 features of the short-horizon future. The model is trained in three stages, which lets videos without action annotations also provide supervision.

Critical strengths

The experiments cover a broad range of settings, and the method adds no inference overhead. The model is post-trained only on standard LIBERO, yet it achieves the best average performance among the methods in the tables on distribution-shift benchmarks such as LIBERO-PRO and LIBERO-Plus. Its advantage is especially clear on dimensions such as task-goal changes and perturbations to the robot’s initial state. On real-robot tasks, it also clearly outperforms the compared methods.

Critical weaknesses

1.The central claim of the title and abstract lacks experimental support. That claim is that world-imagination RL strengthens CoD and thereby enables learning from unlabeled videos and human videos. In Table 3, the gains from RL are small, and some dimensions even decline. Appendix D also states that all results come from a single training seed. 2.The only evidence for human-to-robot transfer is a visualization of a single oven task. There is no quantitative evaluation, such as success rates, and no controlled comparison. The remaining real-world tasks are fine-tuned only on robot demonstrations, and the paper includes no ablation that removes human data. The test instructions explicitly state the order of skills, and every atomic skill appears in the robot data. It therefore cannot be ruled out that these tasks could be completed without relying on human videos. 3.The real-robot experiments include only system-level comparisons. There are no ablations that remove CoD or RL. 4.The RL reward itself does not require labels. However, according to Tables 4 and 5, the RL stage still uses data from the same sources as SFT, and this data carries motion annotations. The paper does not validate RL on unlabeled videos. 5.The experimental setup is not clearly described.

Questions for authors

1.Does the RL stage use any videos without motion annotations? If not, could you provide results of RL on unlabeled videos, with varying data scale? 2.Could you report results for “+CoD” and “+RL” in Table 3 across multiple training seeds, and state the number of evaluation episodes for LIBERO-PRO? This determines whether the small improvement from RL is credible. 3.How were the baselines initialized in the real-robot experiments? For example, does MolmoAct2 start from the BimanualYAM checkpoint? Could you also provide real-robot results for ThinkWorld without RL and without human data? 4.Could you provide quantitative evaluations and results on more tasks or task combinations, compared against the same model trained without human videos? 5.The results in Figure 1 appear to come from Table 6. Could you explain why Figure 1 is inconsistent with Table 6?

4 comments
AuthorsOpenReview ↗

Response to Reviewer fAkT (1/2)

We thank the reviewer for the detailed review and constructive feedback. We have conducted additional experiments to address the concerns raised and summarized the new results in the general response. Changes addressing your comments are marked in orange in the revised manuscript.

[W4, Q1] Does RL use videos without motion annotations? → Sec. 4.5, Fig. 14

In our original RL stage, the videos come from the same sources as SFT, while the world-imagination reward itself uses no motion annotations. To test whether RL can benefit from new human videos, we train a robot-only SFT model without any human video and then apply RL with human videos, which the model first encounters during RL (Sec. 4.5).

TrainingHuman video in SFTHuman video in RLLIBERO-PRO Avg. (%)
SFT onlyNone—45.37
SFT + RLNone50%47.50 (+2.13)
SFT only (reference)25%, motion-labeled—47.44
SFT only (reference)50%, motion-labeled—47.77

RL raises the robot-only model by 2.13 points, comparable to adding 25% human video to SFT. Fig. 14 shows that the same RL stage improves every human-data scale, with the largest gain when SFT sees the least human data.

[W1, Q2] Is the RL gain on LIBERO-PRO credible? → App. C.5 (Tab. 11)

We added two RL seeds, each evaluated on 10,000 LIBERO-PRO episodes, and a control that continues SFT for the same 2k iterations instead of RL. Across three RL seeds initialized from the same +CoD SFT checkpoint, RL achieves 55.4 ± 0.3%, with every seed outperforming the starting checkpoint (54.4%). In contrast, continuing SFT for the same 2k iterations lowers the average to 53.9%.

The +CoD SFT baseline remains a single training run, as repeating its full 150k-step training across seeds was infeasible within the rebuttal period.

LIBERO-PRO (%)Avg.
SFT checkpoint (start)54.4
+ 2k continued SFT53.9
+ 2k RL, 3 seeds55.4 ± 0.3

[W2, Q4] Do human videos improve skill composition beyond robot demonstrations and explicit instructions? → Tab. 12

We compare the same model trained with vs. without human videos on two oven-task combinations: open → close and open → put-left → close (10 trials each). Both variants use identical robot atomic-skill data and test instructions specifying the skill order. Human videos provide composite demonstrations with language descriptions but no motion labels.

As shown in Tab. 12, human videos improve success from 3/10 to 8/10 and from 0/10 to 5/10, respectively (3/20 → 13/20 overall). Here, human videos are the only source of demonstrations showing how atomic skills connect into complete sequences. They provide essential compositional supervision beyond knowing the individual skills and their execution order.

Composition (oven, 10 trials each)Without human videoWith human video
open → close3/108/10
open → put-left → close0/105/10
Overall (20 trials)3/2013/20

[W3, Q3] Real-robot ablations and baseline initialization → Sec. 4.3 (Tab. 12); App. B.3-B.4 (Tabs. 10, 16)

We have added an RL ablation on the real-robot sorting task (Tab. 16). Removing RL lowers success from 15/15 to 8/15 in-distribution and from 13/15 to 6/15 out-of-distribution. In the setting requiring a new trajectory, ThinkWorld succeeds in 5/5 trials, compared with 0/5 without RL.

Sorting taskIDOOD
ThinkWorld15/1513/15
w/o RL8/156/15

For the human-video ablation, we report the real-robot skill-composition comparison from W2 (Tab. 12). Both variants use identical robot atomic-skill demonstrations and test instructions; adding human videos of the complete sequences raises overall success from 3/20 to 13/20.

In addition, all methods in Fig. 3 use the same robot-only demonstrations and the same amount of task data for each task, and follow their official LeRobot training recipes. MolmoAct2 starts from its released BimanualYAM checkpoint. Other pretrained baselines start from official released checkpoints, with π0.5 and Fast-WAM first adapted on the same BimanualYAM data, while Diffusion Policy is trained from scratch. We have added these details to Appendix B.4.

AuthorsOpenReview ↗

Response to Reviewer fAkT (2/2)

[W5] Experimental setup and implementation details → Sec. 4.1; App. A, B.2-B.4 (Tab. 10)

We have described the model configuration, training data, and hyperparameters in Sec. 4.1 and App. A, and the real-robot hardware, control settings, task data, and evaluation protocols in App. B.2-B.3 in our submission draft. We have further added App. B.4 to clarify each method’s training recipe, checkpoint initialization, and BimanualYAM adaptation, with batch sizes and task-specific fine-tuning costs reported in Tab. 10 in the revised version.

[Q5] Inconsistency between Fig. 1 and Tab. 6 → Fig. 1

We thank the reviewer for pointing this out. Fig. 1 retained incorrectly calculated scores for MolmoAct2 and LaWAM, while Tab. 6 and Fig. 3 reported the corrected results. We have updated Fig. 1 to match Tab. 6.

We hope these results address the reviewer's main concerns, and we would be grateful if the reviewer would consider revisiting the score.

Thank you again for your time and consideration.

As the discussion period draws to a close, we would like to follow up on our response to your review. We would be grateful if you could let us know whether our response addresses your concerns. We would also be happy to provide further clarification or discuss any remaining questions.

Thank you once again for your time and consideration.

With the discussion period nearing its end, we wanted to kindly follow up on our previous response. We would greatly appreciate it if you could share whether our clarifications have adequately addressed your concerns. Please feel free to let us know if there are any remaining questions or points that would benefit from further discussion.

Reviewer_qJ3uOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ThinkWorld, a vision-language-action (VLA) learning framework that represents action reasoning as a Chain-of-Dynamics (CoD): a sequence of discrete motion tokens describing intermediate world dynamics from the current observation to a future state. ThinkWorld first supervises CoD and CoD-conditioned future visual prediction using action-labeled human/robot demonstrations, then applies reinforcement learning with a world-imagination reward: sampled CoDs are rewarded by the similarity between their imagined future representation and the observed future representation. This allows action-unlabeled videos to supervise the same CoD reasoning space. Experiments include real-robot OOD generalization, human-to-robot skill composition, LIBERO/LIBERO-PRO, and analyses of CoD attention, cross-embodiment clustering, and future-prediction quality. The main claimed contributions are a shared VLM-native CoD interface, consequence-grounded RL from observed futures, and human-video scaling for robot skill transfer.

Critical strengths

Consequence-grounded RL for action-unlabeled video is a clean and consequential idea. Rewarding sampled CoDs by alignment between imagined and observed futures turns ordinary human/robot videos into supervision for action reasoning without requiring executable actions or hand-pose retargeting. This directly addresses a central bottleneck in learning from human video and is the paper’s strongest conceptual contribution.

The empirical scope is strong and includes human-to-robot skill composition. The real-robot OOD evaluation, LIBERO-PRO results, and the demonstration that a human video can specify a compositional sequence over atomic robot skills provide meaningful evidence beyond standard in-domain imitation. The CoD attention and t-SNE analyses also help support the claimed motion-centric representation.

Critical weaknesses

The causal role of CoD in continuous control is not fully isolated. The action expert consumes VLM representations rather than CoD tokens directly, and during RL only the VLM is updated while the action expert is frozen. The ablation in Table 3 is useful, but it cumulatively adds history, future prediction, CoD, and RL. It is therefore difficult to tell whether the gain comes from CoD as a discrete action-reasoning interface, from the auxiliary future-prediction objective, or from extra VLM representation shaping. A more direct ablation—e.g., matching future-prediction training without CoD, or probing whether CoD-conditioned representations are necessary for the action expert—would materially strengthen the claim.

Experimental fairness and statistical evidence need clarification. The paper reports a single training seed and real-robot results over limited tasks/trials. Baselines such as MolmoAct2 and LaWAM should be shown to use comparable data, tuning, and inference budgets. Figure 3 is qualitative/plot-based without exact success-rate values in the main text, and Table 6 is referenced for real-robot success rates but is not visible in the provided manuscript. These issues do not invalidate the approach but limit confidence in the magnitude and robustness of the reported OOD gains.

Questions for authors

Can you provide a more direct ablation isolating CoD from future-prediction training—e.g., same future-prediction loss and RL reward but without CoD tokens—to show that CoD itself, rather than only richer VLM features, causes the control gains?

Why is the future prediction horizon fixed at 1 second? How does performance vary with horizon, especially for long-horizon or slow/fast tasks?

Were all real-robot baselines trained with the same robot data, auxiliary human video, tuning budget, and action-chunk/replanning setup? Please report exact success rates and trial counts for Figure 3 and Table 6.

The RL reward uses DINO feature similarity with motion weighting. Could the policy exploit static or easy-to-predict regions? How stable is GRPO training across seeds and reward normalizations?

4 comments
AuthorsOpenReview ↗

Response to Reviewer qJ3u (1/2)

We thank the reviewer for the positive assessment. We summarize the additional experimental results in the general response and address each point below, with references to the corresponding revised sections. Changes addressing your comments are marked in blue in the revised manuscript.

[W1, Q1] Where does the gain come from? → Tabs. 3, 15; App. C.7

We thank the reviewer for suggesting more direct ablations. We isolate CoD from future prediction and compare RL with continued SFT using the same data and number of training iterations.

CoD and future prediction. With the same data, training steps, and future-prediction loss, adding CoD improves success from 50.4% to 54.4% (Tab. 3). We additionally include a CoD-only variant to separate CoD supervision from future prediction: CoD alone reaches 51.4%, above future prediction alone (50.4%), and combining the two gives the best result. Perturbing CoD also degrades future prediction (Tab. 2), showing that predicted futures depend on CoD content.

RL. Our GRPO samples CoD sequences and rewards their imagined consequences. Without CoD, there is no sampled sequence to produce candidate-dependent rewards, so the same GRPO objective cannot be applied unchanged. We therefore add a control that continues SFT for the same 2k iterations as RL. Continued SFT lowers success to 53.9%, while RL reaches 55.6%.

Variant (LIBERO-PRO)Future lossCoDRLAvg.
+Future prediction✓✗✗50.4
+CoD only (new)✗✓✗51.4
+Future prediction + CoD✓✓✗54.4
SFT continued, step-matched (new)✓✓✗53.9
+RL✓✓✓55.6

The action expert does not need CoD at inference. As shown in App. C.7 (Tab. 15), feeding the generated CoD to the action expert does not improve success (53.6% vs. 55.6% without CoD) and increases latency from 600 ms to 1,029 ms per call (1.7×; A100, batch size 1). This indicates that CoD's benefit is already carried by the VLM representations it shapes during training, consistent with prior work on internalized reasoning, e.g., ECoT-Lite [1], so our final design does not feed it to the action expert.

AuthorsOpenReview ↗

Response to Reviewer qJ3u (2/2)

[W2, Q3] Are the real-robot comparisons fair and the gains reliable? → App. B.3-B.4 (Tabs. 6, 10)

All methods compared in Fig. 3 are fine-tuned on the same task-specific robot demonstrations with the same amount of task data, following their official LeRobot [2] training recipes. Pretrained baselines start from official released checkpoints, with π0.5 and Fast-WAM first adapted on MolmoAct2-BimanualYAM, while Diffusion Policy is trained from scratch. All models share the same RTX 3090, 15 Hz control, and half-chunk replanning.

We have further detailed the training settings and BimanualYAM adaptation in App. B.4, with batch sizes and fine-tuning GPU-hours reported in Tab. 10. Exact success rates and trial counts for Fig. 3 are provided in App. B.3 (Tab. 6), now supplemented with 95% Wilson confidence intervals. On pooled OOD trials, ThinkWorld achieves 52/60 (86.7%) versus 25/60 (41.7%) for MolmoAct2, with non-overlapping intervals. These intervals quantify uncertainty over physical trials; the real-robot results use a single training seed.

[Q2] Why a 1-second horizon?

We use a 1-second horizon to match our motion-tokenization window, consistent with the action-tokenization practice of FAST [3] and MolmoAct2 [4]. This aligns the predicted future with the motion encoded by CoD and provides a common physical time span across human and robot sources recorded at 3–50 Hz. The prediction horizon is therefore coupled to the CoD window. We have not evaluated alternative horizons and leave their effects on slow, fast, and long-horizon tasks to future work.

[Q4] Can RL exploit static regions? Is GRPO stable? → App. C.5 (Tab. 11)

To reduce reliance on static regions, we use motion weighting to emphasize motion-related changes. The additional experiments in App. C.5 (Tab. 11) show that GRPO consistently improves over SFT across three seeds, with the strongest results using both group and batch normalization.

We further conducted ablations of reward normalization and motion weighting in App. C.5 (Tab. 11). Our full method achieves 55.4 ± 0.3%, compared with 54.3% for group-only normalization and 54.9% for batch-only normalization, supporting the combined design. It also achieves a slightly higher average than the global reward without motion weighting (55.1%), suggesting a modest benefit from emphasizing dynamic regions.

SettingLIBERO-PRO Avg.
SFT, before RL54.4
RL, full method (3 seeds)55.4 ± 0.3
Group-only normalization54.3
Batch-only normalization54.9
Global reward, no motion weighting55.1

We are happy to discuss any remaining questions.

[1] Training Strategies for Efficient Embodied Reasoning, CoRL 2025

[2] LeRobot: State-of-the-art Machine Learning for Real-World Robotics in Pytorch, Github 2024

[3] FAST: Efficient Action Tokenization for Vision-Language-Action Models, RSS 2025

[4] MolmoAct2: Action Reasoning Models for Real-World Deployment, CoRL 2026

Thank you again for your time and consideration.

As the discussion period draws to a close, we would like to follow up on our response to your review. We would be grateful if you could let us know whether our response addresses your concerns. We would also be happy to provide further clarification or discuss any remaining questions.

Thank you once again for your time and consideration.

With the discussion period nearing its end, we wanted to kindly follow up on our previous response. We would greatly appreciate it if you could share whether our clarifications have adequately addressed your concerns. Please feel free to let us know if there are any remaining questions or points that would benefit from further discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Revisions

We thank all reviewers for their careful and constructive feedback. Reviewer qJ3u described consequence-grounded RL as “a clean and consequential idea” that “directly addresses a central bottleneck in learning from human video.” Reviewers fAkT and JpiN noted the broad experiments, the best average performance on LIBERO-PRO, and clear real-robot gains. They also valued that CoD and future prediction add no inference overhead.

The shared concerns were:

  1. the statistical robustness of the RL gain,
  2. isolating the effectiveness of CoD and RL,
  3. direct evidence for learning from action-unlabeled human videos, and
  4. real-robot setup and baseline details.

We address these points with the additional results below and refer reviewers to the updated sections for details. Changes in the revised manuscript are marked in blue, orange, and purple for reviewers qJ3u, fAkT, and JpiN, respectively. Figure regions addressing two reviewers are outlined in both corresponding colors. Table and figure numbers refer to the revised manuscript; additional figures and tables are numbered after those in the submission, preserving existing numbers.

#Additional results requested by reviewersReviewersReferred Section
1Effectiveness of RL on unseen human video across human-data scalesfAkT, JpiNSec. 4.5 (Fig. 14)
2RL robustness across RL seeds and reward designs, and a step-matched comparison with continued SFTqJ3u, fAkT, JpiNApp. C.5 (Tab. 11)
3CoD isolation with a CoD-only variant, and an action expert that directly consumes CoDqJ3uTabs. 3, 15; App. C.7
4Real-robot ablation of ThinkWorld without RL on the sorting taskfAkTApp. B.3 (Tab. 16)
5Skill composition success rates with and without human videosfAkT, JpiNSec. 4.3 (Tab. 12)
6Additional real-robot baselines π0.5 and Fast-WAM, added to the real-robot evaluationJpiNFig. 1, 3; Sec. 4.2; App. B.3 (Tab. 6)
7Training recipes and costs for each real-robot policy, including fine-tuning and pretraining GPU-hoursJpiN, fAkT, qJ3uApp. A.4, B.4 (Tab. 5, 10)
8Inference latency and memory for all six policies on the deployment RTX 3090JpiNApp. C.4 (Tab. 9)
9Quantitative cross-embodiment analysis with motion classification and retrieval across human and robot datasetsJpiNSec. 4.5 (Tab. 13); App. C.6 (Tab. 14, Fig. 15)

Clarifications. We have updated the MolmoAct2 and LaWAM success rates in Figure 1 to match Table 6 (Appendix B.3) and Figure 3. We have also added the confidence intervals mentioned in Appendix D to the corresponding tables. We thank reviewer fAkT for pointing this out.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.