acceptodds
Under review as a conference paper at ICLR 2027

CoRe-WAM: Condition-Aligned Distillation and Recurrent Action Refinement for Efficient World-Action Models

Abstract

World-Action Models (WAMs) integrate world prediction with action generation, but iterative sampling increases inference latency during closed-loop control. Reducing this latency is therefore important for timely action updates in response to new observations. In this paper, we focus on the sequential formulation, which first predicts future videos and then generates actions conditioned on them, rather than jointly sampling both modalities. Distribution Matching Distillation (DMD) offers a way to accelerate both stages by training few-step generators with distribution-level supervision. However, a direct branch-wise DMD adaptation that retains ground-truth video conditioning for action matching introduces a mismatch with the student-generated futures used at inference. Moreover, video distribution matching alone provides no explicit action-level feedback for updating the predicted futures. To address these issues, we propose CoRe-WAM, a framework that combines Condition-Aligned Distribution Matching Distillation (CA-DMD) with Recurrent Action Refinement (RAR). CA-DMD shares the same student-generated future across the action sampler and both the teacher and fake action score models. A vector-Jacobian product (VJP) further propagates the action distribution-matching gradient through the action sampler to the video generator, incorporating downstream action feedback alongside video distribution matching. During inference, RAR constructs a proposal from the preceding chunk's terminal action and refines the noise-perturbed proposal under the current video prediction. Experiments on RoboTwin 2.0, LIBERO, and LIBERO-Plus demonstrate competitive manipulation performance with significantly reduced inference latency. On RoboTwin 2.0, our method reduces the mean inference latency per recurrent chunk from ms for the original LingBot-VA teacher to ms on an NVIDIA PRO 6000 GPU, yielding a speedup.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot manipulation, world action models, efficient distillation

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

This paper presents CoRe-WAM which focusing on reducing the latency of World-Action Models (WAMs). CoRe-WAM has 2 main components:

  1. CA-DMD performs action distribution matching under teh same student-generated future video used by the action sampler instead of using ground-truth future videos during distillation.

  2. RAR: At each subsequent action chunk, noise is added to the most recent (previous) action latent. Such a perturbed latent will then be refined using the current context and video prediction.

Extensive results were presented, with the proposed method showing a good balance between manipulation performance and inference latency by achieving comparable performance with much lower latency.

Critical strengths
  • The tradeoff is a sensible approach. Given that latency is the bottleneck, the proposed method presented impressive results by having 19.35x speedup over the original teacher LingBot-VA.

  • The idea of passing action-distribution feedback back to the video generator is a sensible idea to support the sequential distillation approach in CA-DMD.

  • RAR is training-free, making it a very smart way to refine teh action latent.

  • Extensive evaluations were presented with a wide range of tasks and scenarios.

Critical weaknesses
  • The evaluation of teacher action prediction and score reliability across tasks could have been stronger.

  • The cumulative design in the ablation does not fully isolate component interactions which weakens the evidence to understand the performance gain associated with each individual component.

  • The recurrent refinement seems to be sensitive to the noise level with the large performance difference demonstrated. It is unclear how generalizable the settings are when transferring to other tasks.

Questions for authors

I would like the authors to clarify the 3 weaknesses stated above.

1 comment

Thanks for your positive assessment and valuable feedback. Our responses to the main concerns are given as follows.

(The evaluation of teacher action prediction and score reliability across tasks could have been stronger.)

Thanks for the suggestion. We extend the teacher action diagnostic from RGB block ranking to eight RoboTwin Clean tasks and add a directional test of the DMD feedback. Table R1 compares teacher XYZ RMSE under ground truth and student-generated futures using one action snippet from a single arm per task. We also compare Joint DMD and CA-only on the same tasks and scene seeds, with 50 closed-loop trials per task. CA-only improves task success even in cases where teacher RMSE is higher under student-generated futures. RMSE measures agreement with demonstration actions, whereas condition alignment adapts action supervision to the video conditions encountered at inference. These results support the benefit of addressing the conditioning mismatch.

Table R1. Teacher action RMSE on selected snippets and closed-loop success on the corresponding tasks.

TaskGT-future RMSE (mm)Student-future RMSE (mm)Joint DMD (%)CA-only (%)
place_fan1.2751.91832.0066.00
lift_pot2.3943.26676.0096.00
place_cans_plasticbox1.1800.83972.0092.00
place_dual_shoes1.1480.94242.0060.00
stack_blocks_two1.7351.67082.00100.00
stack_blocks_three1.4031.22426.0084.00
blocks_ranking_rgb1.7572.05512.0058.00
blocks_ranking_size1.5002.27020.0054.00

To assess whether the DMD feedback provides useful control guidance, we conduct a paired directional test across these eight tasks, with 25 trials per task (200 paired evaluations per setting). The DMD update direction is −ga-g_a, where gag_a is the action DM gradient in Eq. (5). Starting from the same student actions, we compare the original actions with corrections along this direction, its opposite direction, and a random direction of equal magnitude. Following the DMD update direction achieves the highest success rate in Table R2, supporting its control utility under the evaluated conditions.

Table R2. Directional test of DMD feedback.

Action settingAverage success (%)
Original student actions73.00
Random correction of equal magnitude71.50
Correction opposite to the DMD update68.50
Correction along the DMD update80.00

(The cumulative design in the ablation does not fully isolate component interactions, which weakens the evidence to understand the performance gain associated with each individual component.)

Thanks for the comment. We include additional ablations in Table R3 that show that each component improves success individually. We evaluate CA, VJP, and RAR separately alongside their combined variants on a subset of 25 RoboTwin 2.0 tasks, with 50 trials per task in each of the Clean and Randomized splits. Success rates are averaged over the two splits.

Table R3. Component ablation on a subset of 25 RoboTwin 2.0 tasks. “Student future” denotes condition alignment (CA).

VariantStudent futureVJPRARSuccess (%)
Joint DMD———66.20
+ CA only✓——70.12
+ VJP only—✓—68.04
+ RAR only——✓70.32
+ CA + VJP (CA-DMD)✓✓—78.60
+ CA + VJP + RAR (Ours)✓✓✓83.08

In the variant with VJP alone, the action sampler and both score models use ground-truth futures. The VJP is evaluated under this condition and incorporated into the video update as in Eq. (7).

(The recurrent refinement seems to be sensitive to the noise level with the large performance difference demonstrated. It is unclear how generalizable the settings are when transferring to other tasks.)

Thank you for the comment. We include an additional validation in Table R4 that supports the shared setting σr=0.50\sigma_r = 0.50. We compare the three noise levels on a held-out validation split, which covers 10 RoboTwin 2.0 Clean tasks, with 50 trials per task. This split is independent of the main evaluation set and used only for this additional noise comparison, with the checkpoint and all other hyperparameters fixed.

Table R4. Noise-level validation on the held-out split.

σr\sigma_rSuccess (%)
0.2534.60
0.5090.40
0.7588.20

As shown in Table R4, σr=0.50\sigma_r = 0.50 performs best, consistent with the shared setting already used in our main evaluations. Both 0.500.50 and 0.750.75 achieve high success. The lower success at 0.250.25 may arise from retaining too much of the historical proposal, limiting adaptation to the current observation and predicted future. These results support using the shared setting across tasks while confirming sensitivity to lower refinement noise.

We will include these additional experiments in the revised PDF.

Reviewer_yEJwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoRe-WAM addresses the cost of generating future video and robot actions in world action models, and the mismatch between training and testing conditions. CA-DMD uses future video generated by the student as the shared condition for action generation and distribution matching. This reduces the shift caused by training with real future video but testing with predicted video. A vector-Jacobian product passes gradients from the downstream action objective back to the video generator, so video generation is also guided by action quality. RAR uses the final action of the previous action chunk to construct a candidate for the current chunk, then adds noise and refines it with a single evaluation.

Critical strengths
  1. The paper addresses a relevant problem. Both inference cost and the mismatch between training on real future video and testing on predicted video can limit the usefulness of world action models for robot control.

  2. The method consistently improves over Joint DMD on all three benchmarks, and on LIBERO-Plus it slightly exceeds the original teacher.

Critical weaknesses
  1. The core novelty seems limited. The method builds on LingBot-VA and improves an existing distillation approach, DMD. Although the experiments report gains, the paper does not make a strong enough case that CA-DMD and RAR amount to a substantial new contribution rather than an incremental modification.

  2. Reducing latency is a central motivation, but the reported improvement appears to be only a few milliseconds over Joint DMD and Flash-WAM. The paper does not establish how meaningful this gain is in practice. It also lacks comparisons with MECo-WAM [1], Fast-WAM [2], and PFD [3], which appear to offer better performance and lower latency. Without these comparisons, the claimed efficiency advantage is hard to judge.

[1] Jianjun Zhang, Jian Zhu, Taiyi Su, Chong Ma, Zitai Huang, Yi Xu, and Hanli Wang. Learning 4d geometric priors for inference-efficient world action models. arXiv preprint arXiv:2607.05468, 2026.

[2] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026.

[3] Pengcheng Fang, Hongli Chen, and Xiaohao Cai. Privileged foresight distillation: Zero-cost future correction for world action models. arXiv preprint arXiv:2604.25859, 2026.

Questions for authors
  1. Latency matters for robot control, but the method is only a few milliseconds faster than Joint DMD and Flash-WAM in Table 4, a difference that may fall within normal timing variation. Since VJP is used only during training and RAR keeps one action evaluation per chunk, neither module should change inference cost relative to Joint DMD. What accounts for this difference? How large is the gain across repeated runs, and does it produce a measurable improvement in the actual control loop?

  2. Table 1 appears to show better results in the Clean setting but weaker results in the Randomized setting. MECo-WAM [1], Fast-WAM [2], PFD [3], LaWAM [4], and Motus [5] seem to perform substantially better there. What explains this gap, and how should the method’s overall performance be judged against these methods?

  3. Why does the Table 5 ablation use LIBERO-Plus Long rather than RoboTwin 2.0? Please report the effect of CA-DMD and RAR on the RoboTwin 2.0 success rate.

[1] Jianjun Zhang, Jian Zhu, Taiyi Su, Chong Ma, Zitai Huang, Yi Xu, and Hanli Wang. Learning 4d geometric priors for inference-efficient world action models. arXiv preprint arXiv:2607.05468, 2026.

[2] Tianyuan Yuan, Zibin Dong, Yicheng Liu, and Hang Zhao. Fast-wam: Do world action models need test-time future imagination? arXiv preprint arXiv:2603.16666, 2026.

[3] Pengcheng Fang, Hongli Chen, and Xiaohao Cai. Privileged foresight distillation: Zero-cost future correction for world action models. arXiv preprint arXiv:2604.25859, 2026.

[4] Jialei Chen, Kai Wang, Kang Chen, Shuaihang Chen, Feng Gao, Wenhao Tang, Zhiyuan Li, Weilin Liu, Zhuyu Yao, Boxun Li, et al. Lawam: Latent world action models for efficient dynamics-aware robot policies. arXiv preprint arXiv:2606.15768, 2026.

[5] Hongzhe Bi, Hengkai Tan, Shenghao Xie, Zeyuan Wang, Shuhe Huang, Haitian Liu, Ruowen Zhao, Yao Feng, Chendong Xiang, Yinze Rong, et al. Motus: A unified latent action world model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 35101–35113, 2026.

3 comments

(1/3 of the responses)

Thanks for your valuable feedback and questions. Our responses to the main concerns are as follows. All references to the methods discussed below were provided by the reviewer.

(The core novelty seems limited. The paper does not make a strong enough case that CA-DMD and RAR amount to a substantial new contribution rather than an incremental modification.)

Thanks for the comment. Our contribution is to identify and address two issues that arise when DMD is directly applied to sequential WAMs: the mismatch between training and inference conditions, and the absence of downstream action feedback in video distillation. CA-DMD evaluates action matching under shared futures generated by the student and propagates the action distribution matching gradient through the action sampler to the video generator. This aligns action distillation with inference conditions and makes video updates responsive to downstream action requirements. RAR further refines a proposal constructed from the preceding chunk’s terminal action segment under the current observation and predicted future.

To further demonstrate that these contributions extend beyond LingBot-VA, we conduct additional experiments on FastWAM-IDM and Motus in its sequential VGM→IDM mode (which explicitly generate future video). We compare CA-DMD with Joint DMD using the same backbone and sampling budget, and include teacher references. Our distilled models are evaluated on 50 RoboTwin 2.0 tasks with 50 trials per task in each split. Table R1 reports the results across these backbones.

Table R1. Distillation results across backbones on RoboTwin 2.0 (%). Bold marks the best distilled result within each backbone.

Backbone/MethodVideo/Action NFECleanRandomizedAverage
FastWAM-IDM Teacher10/1091.1691.3491.25
FastWAM-IDM+Joint DMD1/177.5275.9276.72
FastWAM-IDM+Ours1/190.6490.3290.48
Motus Teacher10/1088.6687.0287.84
Motus (VGM→IDM)+Joint DMD1/180.4876.0478.26
Motus (VGM→IDM)+Ours1/182.4081.4481.92

These results support the broader applicability of CA-DMD to DMD of sequential WAMs beyond LingBot-VA.

We will revise the PDF to clarify these contributions and include the additional results.

(2/3 of the responses)

(The reported improvement appears to be only a few milliseconds over Joint DMD and Flash-WAM. What accounts for this difference? How large is the gain across repeated runs, and does it produce a measurable improvement in the actual control loop? It also lacks comparisons with MECo-WAM, Fast-WAM, and PFD, which appear to offer better performance and lower latency.)

Thanks for the comment. We agree that Joint DMD, Flash-WAM, and CoRe-WAM have comparable per-chunk generation costs at the same sampling budget. The small gaps in Table 4 of the main text come from comparing mean latencies over 30 timed runs per method after 10 warm-up runs, which may reflect implementation differences and timing variability. The gap can also vary with the baseline. For example, on FastWAM-IDM (Table R3), CoRe-WAM has slightly higher per-chunk generation latency than directly reducing the NFE (202.92 vs. 181.49 ms), while remaining significantly faster than the original teacher (789.19 ms). We do not claim the few-millisecond gaps as an advantage of our method. As we state in the paper, our goal is to accelerate the original multi-step teacher while improving manipulation success over Joint DMD at the same low sampling budget, rather than to achieve a large additional speedup over other distillation methods already operating at 1v1a.

More importantly, CoRe-WAM achieves shorter mean completion times in the actual control-loop on the evaluated successful rollouts in practice. Following Fast-WAM [2], we report task completion time to assess execution efficiency. To quantify completion times across repeated runs, we evaluate each method on the same three RGB block ranking scenarios using an NVIDIA PRO 6000 GPU, with 10 runs per scenario. We measure elapsed time from the first control loop iteration until the evaluator confirms task success, including model generation, action execution, environment stepping, and rendering. Tables R2 and R3 report mean completion times over successful rollouts, where our method produces a measurable improvement in the actual control loop by finishing the tasks earlier.

Table R2. Mean full simulation control loop completion time on successful rollouts.

MethodCompletion time (s)
LingBot-VA teacher231.62
Joint DMD128.50
Flash-WAM119.39
CoRe-WAM (Ours)110.64

Regarding comparisons with MECo-WAM, Fast-WAM, and PFD, these methods differ from our LingBot-VA baseline in their base policies and training and deployment designs. MECo-WAM builds on the Fast-WAM framework and introduces geometric supervision during training. Fast-WAM uses Wan2.2-5B as its video backbone and removes explicit future generation from its main inference path. PFD uses Fast-WAM as its base policy and transfers action corrections conditioned on future information into a policy that uses only current observations. These choices may partly account for their better performance and lower latency.

Our approach retains the base architecture of LingBot-VA and improves DMD for distillation to fewer sampling steps. Our focus on step distillation is therefore complementary to these designs. Since their main deployment paths do not explicitly generate future video, using them as baselines for step distillation with our fixed teacher would require changes to the training or inference formulation. Thus, these methods serve as indirect baselines for assessing the effect of step distillation.

Despite that, to address the reviewer’s concern about the efficiency advantage more directly, we additionally apply our method to FastWAM-IDM, which generates future video before predicting actions. Its success rates are included in Table R1. Furthermore, Table R3 reports generation time and control-loop completion time under the same setting described above. Our method reduces generation time relative to the original teacher and achieves shorter mean control-loop completion times than both the teacher and directly reduced NFE on the evaluated successful rollouts.

Table R3. Timing results on FastWAM-IDM. Control-loop completion times are means over successful rollouts.

MethodGeneration (ms/chunk)Control-loop completion (s)
Original teacher789.1958.28
Directly reduced NFE181.4946.76
Ours202.9239.77

These additional experiments support the efficiency advantage of our method beyond LingBot-VA, with faster generation than the original teacher and shorter task completion times in the measured successful rollouts.

We will add these results and a comparison of the modeling paradigms to the revised PDF.

(3/3 of the responses)

(Table 1 appears to show better results in the Clean setting but weaker results in the Randomized setting. MECo-WAM, Fast-WAM, PFD, LaWAM, and Motus seem to perform substantially better there. What explains this gap, and how should the method’s overall performance be judged against these methods?)

Thanks for your questions. The low sampling budget is one factor contributing to the Randomized gap. We additionally evaluate CoRe-WAM with one video and two action evaluations per recurrent chunk (1v2a). Increasing the action budget from one to two evaluations raises Randomized success from 78.98% to 83.08%, while retaining one video evaluation. Table R4 compares CoRe-WAM with Flash-WAM at the same 1v2a budget.

Table R4. RoboTwin 2.0 success rates at 1v2a (%).

MethodVideo/Action NFECleanRandomizedAverage
Flash-WAM1/288.4282.6685.54
Ours1/288.8883.0885.98

Regarding overall performance, the cited methods [1]~[5] differ in their base policies and training/inference designs, while reported latencies also depend on hardware and timing protocols. Our focus on few-step DMD for sequential WAMs is complementary to these policy designs. We therefore assess our contribution under matched backbones and sampling budgets. Our formulation directly applies to FastWAM-IDM and Motus in VGM→IDM mode. The main deployment paths of MECo-WAM, Fast-WAM, and PFD omit future generation, while LaWAM predicts latent subgoals, whose designs require further adaptation of our current formulation.

Within the limited rebuttal period, we prioritize the adaptation of our method on two cited sequential WAMs: FastWAM-IDM and Motus in VGM→IDM mode. Table R1 shows improvements over Joint DMD on both Clean and Randomized across these backbones. In particular, our distilled policy retains explicit future generation at 1v1a while achieving success rates close to the cited current-observation policies. Tables R2–R3 further report generation and control-loop timings under our evaluation setup.

These results across three backbones and both splits support the applicability of our method beyond LingBot-VA.

(Why does the Table 5 ablation use LIBERO-Plus Long rather than RoboTwin 2.0? Please report the effect of CA-DMD and RAR on the RoboTwin 2.0 success rate.)

Thanks for the comment. We used LIBERO-Plus Long in Table 5 in the main paper to examine the components during challenging long-horizon execution. Following the reviewer’s suggestion, we additionally conduct component ablations on RoboTwin 2.0, evaluating a subset of 25 tasks with 50 trials per task in each split. We report success rates averaged over Clean and Randomized.

Table R5. Component ablation on a subset of 25 RoboTwin 2.0 tasks. Success rates are averaged over Clean and Randomized. “Student future” denotes condition alignment (CA).

VariantStudent futureVJPRARSuccess (%)
Joint DMD———66.20
+CA only✓——70.12
+VJP only—✓—68.04
+RAR only——✓70.32
+CA+VJP (CA-DMD)✓✓—78.60
+CA+VJP+RAR (Ours)✓✓✓83.08

We will include these additional experiments and discussions in the revised PDF.

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

This paper proposes CoRe-WAM, a framework for reducing inference latency in sequential world-action models (WAMs), which first predict a future video and then generate actions conditioned on it. CoRe-WAM distills both stages of LingBot-VA into one-step generators using Distribution Matching Distillation (DMD). A direct branch-wise DMD setup matches actions under ground-truth futures even though actions are conditioned on student-generated futures at inference. To fix this, CA-DMD conditions the action student, teacher and fake score models on the same student-generated video, and it backpropagates the action distribution-matching gradient to the generated video through a vector–Jacobian product, so video distillation also receives action-level feedback. At inference, the training-free Recurrent Action Refinement (RAR) starts each action chunk from the previous chunk’s noised terminal action and refines it in a single step under the current video prediction. With one video and one action evaluation per chunk, CoRe-WAM reaches 83.4% average success on RoboTwin 2.0, 97.75% on LIBERO and 74.35% on LIBERO-Plus Long, close to the full teacher. It outperforms vanilla joint DMD and reduced-step teacher sampling while cutting per-chunk latency from 5.4 s to 281 ms (19.35×). A cumulative ablation on LIBERO-Plus Long shows that condition alignment, VJP feedback and RAR each improve success.

Critical strengths
  • The paper identifies a concrete problem in naive DMD for sequential WAMs of the actions being distilled under ground-truth futures but executed under generated ones. The fix follows directly from the KL chain-rule decomposition (Eq. 4). It adds no inference cost, and gives large, consistent gains over Joint DMD at the same budget (+17.0 points on RoboTwin, +1.9 on LIBERO, +10.1 on LIBERO-Plus Long). The ablation attributes most of this gain to condition alignment and VJP feedback.
  • At one function evaluation per stage, the student stays close to the multi-step teacher on LIBERO (within 0.85 points) and matches it on LIBERO-Plus Long, with a 19× latency reduction. This can be very useful for closed-loop deployment of video-based policies.
Critical weaknesses
  • The approach seems to be only marginally better than the strongest distillation baseline. At the matched budgets, CoRe-WAM leads Flash-WAM by only about 2 points on RoboTwin (83.39 vs 81.41), and Flash-WAM is better on the Randomized split (80.26 vs 78.98). Flash-WAM’s paper also reports stronger settings that are not shown. The so the results show CA-DMD is comparable to consistency-based distillation, not clearly better.
  • RAR closely resembles the training-free warm-start of One-Step Flow Policy (Li et al., 2026c), which the paper cites only for a generic denoising principle. Its noise level is also chosen on the RoboTwin Clean evaluation itself, where the selected value’s score is the headline result. Performance is sensitive to that choice.
Questions for authors

How does RAR differ from the warm-start in One-Step Flow Policy (Li et al., 2026c)? Can the refinement noise-level be chosen on a held-out split?

1 comment

Thanks for your positive assessment and valuable feedback. Our responses to the main concerns are given as follows.

(The approach seems to be only marginally better than the strongest distillation baseline. Flash-WAM’s paper also reports stronger settings that are not shown. The results show CA-DMD is comparable to consistency-based distillation, not clearly better.)

Thanks for the comment. We agree that the average gain over Flash-WAM on RoboTwin 2.0 is modest. Our contribution addresses sequential distillation under student-generated future conditions, which is a specific challenge within DMD. Consistency distillation and DMD provide parallel approaches to step distillation, and Flash-WAM is a strong representative of the former.

To complete the comparison, we additionally evaluate CoRe-WAM following Flash-WAM’s stronger 1v2a setting. As shown in Table R1, CoRe-WAM achieves 88.88% on Clean and 83.08% on Randomized, compared with 88.42% and 82.66% for Flash-WAM, respectively, yielding comparable performance with higher success on both splits.

Table R1. RoboTwin 2.0 success rates (%). † denotes results reported in Flash-WAM. Bold marks the best result.

MethodNFECleanRand.Avg.
LingBot-VA + Flash-WAM†1/288.4282.6685.54
Ours1/288.8883.0885.98

Across benchmarks, the improvement is more pronounced on LIBERO. At the same 1v1a budget, CoRe-WAM achieves 97.75% average success compared with 95.10% for Flash-WAM (Table 2 in the main paper).

We will include the expanded comparison and clarify the relationship between the two distillation approaches in the revised manuscript.

(RAR closely resembles the training-free warm-start of One-Step Flow Policy (Li et al., 2026c). How does RAR differ from this warm-start? Performance is sensitive to the choice of noise level. Can the refinement noise-level be chosen on a held-out split?)

Thanks for the comment. RAR differs from OFP in how the historical action proposal is constructed. Warm starting is an established strategy in generative planning and control, as already discussed in Diffuser [1] and Diffusion Policy [2]. OFP uses it to complement its self-distillation framework. Let aa denote the preceding action chunk, with prediction horizon HH and execution horizon hh. OFP shifts the unexecuted action suffix and pads it with the terminal action, retaining the remaining portion of the previous plan: $ \bar a_{\mathrm{OFP}} = [a_{h+1},\ldots,a_H,\underbrace{a_H,\ldots,a_H}_{h\text{ times}}]. $ Our RAR instead repeats the unexecuted action suffix twice to construct the next proposal: $ \bar a_{\mathrm{RAR}} = [a_{h+1},\ldots,a_H,;a_{h+1},\ldots,a_H]. $ Here, H=32H=32 and h=16h=16, measured in low-level actions. The terminal action latent in the paper corresponds to the 16-action suffix used here. Repeating this segment preserves its internal temporal variation throughout the proposal, whereas OFP uses a constant terminal action to pad the unexecuted suffix. This proposal is then perturbed with noise and refined under the current observation and newly predicted future.

To examine how these two strategies affect performance, we compare them using the same CoRe-WAM student and inference settings on a 12-task RoboTwin 2.0 subset, with 50 trials per task in each split. As shown in Table R2, RAR achieves higher success than OFP-style initialization on both splits, indicating that its proposal construction is better suited to our setting.

Table R2. Warm start comparison on RoboTwin 2.0 (%).

InitializationCleanRand.
OFP-style warm start80.3367.83
RAR84.8378.00

Regarding noise selection, we additionally compare the three noise levels on a held-out validation split covering 10 RoboTwin 2.0 Clean tasks, with 50 trials per task. This split is independent of the main evaluation set and used only for this additional noise comparison, with the checkpoint and all other hyperparameters fixed. As shown in Table R3, these results support σr=0.50\sigma_r = 0.50, the shared setting already used in our main experiments. Both σr=0.50\sigma_r = 0.50 and 0.750.75 achieve high validation success. The lower success at σr=0.25\sigma_r = 0.25 may arise from retaining too much of the historical action proposal.

Table R3. Noise-level validation on the held-out split.

σr\sigma_rSuccesses / TrialsSuccess (%)
0.25173/50034.60
0.50452/50090.40
0.75441/50088.20

We will revise the PDF to include these discussions and experiments.

[1] Janner et al. “Planning with Diffusion for Flexible Behavior Synthesis.” ICML, 2022.

[2] Chi et al. “Diffusion Policy: Visuomotor Policy Learning via Action Diffusion.” IJRR, 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.