acceptodds
Under review as a conference paper at ICLR 2027

When to Trust Imagination: Adaptive Action Execution for World Action Models

Abstract

World Action Models (WAMs) jointly predict future visual observations and actions, but typically execute a fixed number of actions before replanning, regardless of whether the imagined future remains consistent with reality. We formulate adaptive WAM execution as future–reality verification and propose Future Forward Dynamics Causal Attention (FFDC), a lightweight verifier that reasons over predicted actions, predicted visual dynamics, real observations, and instruction-conditioned semantics. FFDC allows continued execution when a plan remains reliable and triggers replanning when its validity deteriorates. We combine verification with Mixture-of-Horizon Training to improve long-horizon trajectory coverage. On RoboTwin, FFDC-WAM achieves the highest average success rate among the evaluated methods in both clean and randomized settings. Under randomization, it improves success by 2.54 percentage points over Base-Motus while reducing WAM planning calls by 69.10% and completion time by 33.88%. Zero-shot evaluation on DOMINO shows a favorable trade-off over fixed 16-action execution. Across six real-world tasks on a Unitree G1D robot, FFDC-WAM improves average success by 9.17 percentage points over Base-Motus-LC16 while reducing calls by 74.06% and completion time by 51.31%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Adaptive Action Chunking, Future–Reality Verification, Closed-Loop Control, Robotic Manipulation

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_Jhu1OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an adaptive execution framework for World Action Models (WAMs) that compares the WAM’s imagined future with the physical rollout. The goal is to avoid error accumulation during long open-loop chunk execution while reducing the number of WAM calls. The authors introduce Future Forward Dynamics Causal Attention (FFDC), a lightweight verifier that estimates whether the remaining WAM-predicted action segment is still reliable, based on the predicted actions, predicted visual dynamics, the current real observation, and instruction-conditioned semantics. The method is evaluated on 50 RoboTwin tasks in simulation and on six real-world manipulation tasks, and the results show the effectiveness of using the WAM’s predicted future for execution verification.

Critical strengths
  1. The paper addresses an important problem, and the proposed module is simple, lightweight, and intuitive.
  2. The paper includes extensive ablation studies on the verifier inputs, MoH training, and the decision threshold.
Critical weaknesses
  1. Mismatch between the framing and the training data. The paper frames FFDC as future–reality verification. However, a large part of the negative samples is constructed by corrupting demonstration actions (Section 3.3). In these samples, the real observation is consistent with the prediction and only the actions are perturbed. The verifier may therefore learn whether the actions are consistent with the predicted video, rather than whether reality has deviated from the prediction.

  2. Possible repeated short chunks in difficult phases. In difficult phases, the verifier may repeatedly reject the new plan, which could lead to repeated replanning with very short chunks and cause the robot to get stuck. The paper does not describe a safeguard for this case or report the distribution of executed chunk lengths.

Questions for authors
  1. In the synthetic negatives (Section 3.3), the real observation is consistent with the prediction and only the actions are perturbed. Could the authors evaluate the verifier separately on negatives from failed rollouts and from synthetic corruptions, to clarify whether it detects deviations between reality and the prediction? This is relevant because removing the real observation reduces the success rate only from 76.4% to 72.4% (Table 3).

  2. In difficult phases, the verifier may repeatedly reject similar plans, leading to very short chunks and preventing progress. Did the authors observe such cases, and is there a safeguard such as a minimum chunk length or a limit on consecutive replans? The distribution of executed chunk lengths on the hard tasks would help clarify this.

  3. Please specify how often the verifier is evaluated during execution and the value of k used for verification.

  4. The Abstract and Introduction report differences in percentage points as percentages (e.g., 9.17% and 13.33%). The name “Base-Motus-LC16” is also ambiguous, as it is unclear whether it refers to Base-Motus or to LC-16.

  5. Several tables and figures are placed in the middle of the text. Placing them at the top of the page would improve readability.

  6. The paper does not provide videos of the real-robot experiments, and Figure 4 shows only a single snapshot sequence.

I am willing to raise my score if those points are addressed.

9 comments
AuthorsOpenReview ↗

Response to W1 and Q1: Future–reality framing versus action-corruption supervision

Dear Reviewer Jhu1,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1 and Q1: Future–reality framing versus action-corruption supervision

Thank you for highlighting this concern. We have added the supervision analyses to Appendix A, “Analysis of verifier supervision” (Table 4), with a cross-reference in Section 3.3, “Training strategy and dataset construction.”

Our action-corruption strategy is motivated by a failure pattern observed in our WAM rollouts: predicted visual futures often depict successful execution even when the corresponding actions are incorrect. Corrupting demonstration actions while retaining the original visual context approximates this inconsistency. These synthetic examples are complemented by actual successful and failed rollouts, whose real observations, predicted visual tokens, and executed actions provide supervision for naturally occurring execution discrepancies.

1. To examine explicit visual-mismatch supervision, we additionally train a verifier variant whose synthetic negatives perturb or mismatch only the predicted visual tokens, while keeping real observations and remaining actions unchanged.

MethodRandomized SR (%)Completion time (s)WAM calls
FFDC (visual-negative variant)88.1615.761.70
FFDC-WAM88.2016.201.69

The visual-negative variant achieves 88.16% SR, compared with 88.20% for the original FFDC, maintaining comparable downstream performance with an alternative synthetic-negative construction.

2. To directly address Q1, we evaluate the FFDC checkpoint with the fixed threshold of 0.5, without retraining, on two independent test sets. The actual-rollout set contains segments from successful executions and critical failure segments from failed executions. The demo-derived set contains unmodified demonstration segments as positives and synthetically corrupted segments as negatives, generated from demonstration episodes excluded from training. Each set contains 10 positive and 10 negative examples per task under each of the clean and randomized settings, covering all 50 tasks.

Test setPositive samplesNegative samplesTotal samplesClassification accuracy (%)
Actual rollouts1,0001,0002,00091.65
Demo-derived1,0001,0002,00093.75

The verifier achieves 91.65% accuracy on actual rollout segments, compared with 93.75% on demo-derived samples, a difference of 2.10 percentage points. Its high accuracy on actual executions provides evidence that its classification ability extends beyond recognizing synthetic corruption patterns. We have included the separate classification results, clarified the motivation and composition of the negative training data, and reported the visual-negative experiment in Appendix A of the revised manuscript.

AuthorsOpenReview ↗

Response to W2 Q2 Q3: Repeated replanning, execution safeguards, and checking frequency, Q4 Q5: Percentage points, baseline naming, and layout, and Q6: Real-robot videos

W2 Q2 Q3: Repeated replanning, execution safeguards, and checking frequency

Thank you for highlighting this concern. We have clarified the verification schedule in Section 3.1, “Preliminary,” and added the detailed execution protocol and chunk-length distributions to Appendix B, “Execution protocol and chunk-length distribution” (Table 5).

The verifier is evaluated every 16 executed action steps: after steps 16, 32, and 48 within a 64-step plan. If the score falls below 0.5, the remaining actions are discarded and replanning is triggered; otherwise, execution continues. Once all 64 actions are exhausted, a new plan is generated if the episode is still ongoing.

The verification horizon is k=16k=16 in Eq. (3) and Section 3.2. Each check assesses the next 16 predicted actions.

The analysis of the executed chunk lengths: We additionally analyzed the execution from the clean and randomized evaluations, each covering 50 tasks with 100 episodes per task. The table below reports the executed chunk lengths between successive WAM calls. We exclude each episode’s final chunk because episode termination can truncate it independently of the verifier.

SettingTaskMean chunk length (steps)chunk 16 (%)chunk 32 (%)chunk 48 (%)chunk 64 (%)
CleanAll 50 tasks51.5125.240.860.6473.26
CleanBlocks Ranking Size53.1321.251.740.7076.31
CleanHanging Mug42.0444.540.841.9652.66
CleanPlace Mouse Pad48.6729.174.170.0066.67
CleanPut Object Cabinet55.8410.795.866.9276.43
CleanScan Object54.8715.012.027.9775.00
RandomizedAll 50 tasks53.6421.100.430.6177.86
RandomizedBlocks Ranking Size59.428.810.880.4489.87
RandomizedHanging Mug37.2455.070.680.6843.58
RandomizedPlace Mouse Pad44.4437.043.703.7055.56
RandomizedPut Object Cabinet55.3711.676.016.8975.43
RandomizedScan Object54.6315.223.885.1775.73

FFDC preserves long execution chunks overall: 73.26% of non-final chunks in the clean setting and 77.86% in the randomized setting reach the full 64 steps, with mean lengths of 51.51 and 53.64 steps, respectively. On Hanging Mug, 16-step chunks are more frequent (44.54% and 55.07%), but 52.66% and 43.58% of chunks still reach 64 steps, yielding mean lengths of 42.04 and 37.24 steps. Thus, FFDC retains substantial long-chunk execution even on this difficult task while allowing earlier replanning. We have clarified the execution protocol and included these distributions in Appendix B of the revised paper. We will investigate recovery mechanisms for persistent replanning in future work.

Q4 Q5: Percentage points, baseline naming, and layout

Thank you for pointing out these issues. We have addressed them in the revised manuscript:

  • In the Abstract and Section 1, “Introduction,” we have corrected “9.17%” and “13.33%” to “9.17 percentage points” and “13.33 percentage points,” respectively (lines 024 and 100 of the revised PDF).
  • We have consistently renamed “Base-Motus-LC16” to “Base-Motus” throughout the manuscript (see lines 387 and 410 of the revised PDF for the revised name in Section 4.2 and Table 2).
  • We have revised the float-placement settings for the main-text tables and figures to favor the top or bottom of pages.

Q6: Real-robot videos

Thank you for this question. We have added a description of the supplementary real-robot videos to Section 4.2, “Real-world experiments,” and provided the videos in the supplementary material. The videos are presented at 1× playback speed, with annotations identifying action execution, WAM inference, and verifier-triggered replanning.

Dear Reviewer Jhu1,

Thank you again for your careful review and constructive feedback. We have provided point-by-point responses to your comments in our replies above.

With two days remaining in the discussion phase, we would be grateful to know whether our responses and additional evidence have adequately addressed your concerns. If you find that our rebuttal and additional evidence have adequately addressed them, we would be deeply grateful for your consideration of revisiting the evaluation. If any questions remain or further clarification would be helpful, we would be happy to continue the discussion.

We look forward to hearing from you. Thank you again for your time and consideration.

Best regards,

The Authors

Reviewer_zfAJOpenReview ↗
Rating
3(+1)
Confidence
2(−2)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes FFDC, a lightweight verifier that decides whether to continue executing a world action model's action chunk or request replanning. It jointly uses real observations, predicted video features, action tokens and semantic features, with structured attention and caching. A variable-horizon Motus backbone enables longer execution. The method reaches 88.20% on RoboTwin random and 75.00% on the real-robot evaluation.

Critical strengths
  1. Determining whether the predicted future state remains sufficiently reliable to support continued execution is a valuable question. The FFDC’s approach, which uses predicted visual outcomes and planned actions to decide when to interrupt open-loop control, is also a reasonable method.
  2. The evaluation encompassed a wide range of simulation experiments, real-world robotic tasks, and component elimination experiments, providing evidence of the method’s practical value.
  3. The FFDC architecture is relatively simple and conceptually clear, making the method quite attractive from an implementation perspective.
Critical weaknesses
  1. The discrepancy with the original Motus results requires clarification. The paper reports that Base-Motus achieved success rates of 87.66% and 85.66% on the RoboTwin Clean and Random datasets, respectively, whereas the original Motus manuscript reported success rates of 88.66% and 87.02%. This affects the interpretation of the reported improvement: the FFDC results of 88.90% and 88.20% are only 0.24 and 1.18 percentage points higher than the original results. Although comparisons across different papers may involve different experimental protocols, the authors should specify relevant differences in training, checkpoints, action time limits, or evaluation settings.
  2. The contributions of longer execution times and validation times should be distinguished more clearly. On the RoboTwin Random dataset in Table 1, increasing the number of cycles from LC-16 to LC-64 raised the success rate from 83.76% to 87.40%. In contrast, the transition from LC-64 to FFDC resulted in only a 0.80 percentage point further increase in the success rate, from 87.40% to 88.20%, simultaneously increasing the average completion time from 13.7 seconds to 16.2 seconds. These results indicate that the improvement in success rate is primarily attributable to longer execution times, while validation, although yielding a small additional benefit, also significantly increases the time overhead
  3. The validity of the synthetic negative examples needs further examination. Section 3.3 describes additional verifier training data that include synthetically corrupted action segments. Such perturbations may introduce recognizable artifacts that differ from naturally occurring execution failures. Without further controls, it remains unclear whether the verifier learns generalizable failure cues or partly relies on the construction of these synthetic negatives.
Questions for authors
  1. Could you provide more details and describe the method used to generate synthetic negative data, as described in Section 3.3?
  2. As noted in Weakness 2, is Base-Motus’s lower performance due to the use of 16-action chunks? Could you provide a fair comparison using 64-action Motus and FFDC-WAM, or conversely, provide FFDC-WAM with 16-action chunks?
  3. In real-world experiments, FFDC-WAM achieved a success rate improvement of 5–15 percentage points per task compared to Base-Motus-LC16. Given that only 20 trials were evaluated per task, this corresponds to an increase of 1–3 successful trials per task; uncertainty estimates are still needed to determine whether the improvement is stable.
5 comments
AuthorsOpenReview ↗

Response to W1: Difference from the original Motus results, and W2 Q2: Fair horizon comparison and verification overhead

Dear Reviewer zfAJ,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1: Difference from the original Motus results

Thank you for pointing out this discrepancy. We have clarified the checkpoint source, training data and configuration, and evaluation protocol in Section 4, “Experiments,” under “Experimental setups.” Our evaluation also uses RoboTwin [1] under both clean and randomized settings, with 50 tasks and 100 evaluation episodes per task in each setting. For Base-Motus, we use the publicly released checkpoint from Motus [2]. To measure completion time alongside success rate under a consistent hardware setup, we rerun its evaluation. All evaluations are conducted on a single NVIDIA A100 GPU. The Base-Motus success rates reported in our paper therefore come from our own evaluation of the released checkpoint.

For our WAM training with MoH, we use the same publicly available RoboTwin [1] demonstration data as Motus [2]: 50 clean-setting demonstrations and 500 randomized-setting demonstrations per task across all 50 tasks. We train on four NVIDIA A100 GPUs for 320,000 steps, with a per-GPU batch size of 8, giving a total batch size of 32. This matches the cumulative number of training samples processed under the 40,000-step, batch-size-256 configuration reported in Motus.

[1] Chen, Tianxing, et al. "Robotwin 2.0: A scalable data generator and benchmark with strong domain randomization for robust bimanual robotic manipulation." arXiv preprint arXiv:2506.18088 (2025).

[2] Bi, Hongzhe, et al. "Motus: A unified latent action world model." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

W2 Q2: Fair horizon comparison and verification overhead

Thank you for raising this point. We have expanded the contribution analysis in Section 4.3, “Ablation study,” under “MoH training and decision threshold.” We have also added the 16-action backbone comparison under “Verification with a 16-action backbone,” with detailed results in Appendix E, “Verification with the original 16-action backbone” (Table 8). LC-64 already performs well on the 45 easy RoboTwin tasks, achieving 88.98% and 88.89% SR under clean and randomized conditions. However, its SR falls to 73.80% and 74.00% on the five hard tasks. FFDC improves these hard-task results to 76.00% and 76.40%, respectively. Since easy tasks constitute 90% of the benchmark, the overall average places substantially more weight on settings where fixed long-chunk execution already works well.

To isolate the verifier's contribution, LC-64 and FFDC use the same MoH-trained 64-action backbone. Their comparison therefore holds backbone training and the maximum execution horizon fixed while changing the execution strategy. Furthermore, Tables 1 and 3 provide the following comparison on the five hard tasks under randomization:

Backbone trainingFixed LC-64 SRFFDC SRGain from verification
Without MoH68.0%71.4%+3.4 pp
With MoH74.0%76.4%+2.4 pp

Verification improves SR both with and without MoH. Separately, MoH improves fixed execution by 6.0 pp and adaptive execution by 5.0 pp. These comparisons support complementary contributions from long-horizon training and online verification.

The benefit is more pronounced in real-world execution: FFDC increases SR from 61.67% to 75.00%, a 13.33 pp improvement over LC-64, while only needs 0.22 additional planning calls. Meanwhile, completion time remains 51.31% lower than the 84.42 s reported for Base-Motus-LC16, retaining much of the efficiency benefit of long-chunk execution.

To address Q2, Table 3(b) in the submitted paper already includes a controlled comparison using a 64-action backbone trained without MoH. With the backbone and maximum prediction horizon held fixed, adding FFDC increases success from 68.0% to 71.4% on the five hard RoboTwin tasks under randomization. This comparison isolates the benefit of adaptive verification without introducing MoH training.

We added short-horizon comparison by evaluating FFDC on the original Base-Motus with a 16 action chunk. We adapt FFDC to its action chunks while keeping the Base-Motus checkpoint and prediction horizon unchanged. The results on randomized RoboTwin are:

MethodRandomized SR (%)Time (s)WAM calls
Base-Motus85.6624.55.47
Base-Motus+FFDC87.7425.25.78

Adding FFDC improves success by 2.08 percentage points, from 85.66% to 87.74%, while increasing completion time by 0.7 s (2.9%) and planning calls by 0.31. These results show that FFDC also improves reliability with the original 16-action backbone, with a modest additional execution cost. Together, the controlled comparisons at 16-action and 64-action horizons demonstrate benefits from adaptive verification while keeping the underlying WAM fixed.

AuthorsOpenReview ↗

Response to W3 and Q1: Synthetic negative construction and possible artifacts

W3 Q1: Synthetic negative construction and possible artifacts

Thank you for raising this concern. We have documented the verifier training-data composition and clean-to-randomized evaluation in Appendix C, “Verifier generalization across environmental conditions” (Table 6), and the association between verifier scores and execution outcomes in Appendix D, “Verifier scores and execution risk” (Table 7). FFDC receives supervision from both naturally occurring execution failures and synthetic action corruptions. Specifically, we collect 10 successful and 10 failed rollouts per RoboTwin task, totaling 500 successful and 500 failed rollouts across the 50 tasks. Segments from actual failed rollouts provide examples of naturally occurring future–reality mismatches. Existing demonstrations provide positive segments and synthetic negatives generated through temporal swapping, gripper flipping, late-stage Gaussian noise, and tail scaling. Synthetic corruptions therefore supplement supervision from actual execution failures.

We also examine whether FFDC scores are associated with actual rollout outcomes. In the task-and-seed-matched FFDC and LC-64 analysis, using final LC-64 episode outcomes as labels, FFDC achieves AUROC = 0.786 and AUPRC = 0.761. At the operating threshold of 0.5, matched LC-64 episodes fail in 77.8% of cases associated with a replan prediction. These results provide evidence that FFDC scores capture execution-risk information beyond classification of synthetically corrupted training examples. The matched-episode analysis measures association with eventual outcomes, rather than a direct intervention at identical intermediate states.

To further assess transfer across environmental conditions, we evaluate a verifier trained exclusively on clean-setting demonstrations and successful/failed rollouts, with synthetic negatives generated from those clean-setting trajectories. It is evaluated on randomized RoboTwin without verifier fine-tuning. We keep the WAM backbone, evaluation protocol, and decision threshold unchanged.

Method / verifier training dataRandomized SR (%) ↑Completion time (s) ↓WAM calls ↓
FFDC / standard training88.2016.21.69
FFDC / clean-setting data only88.0016.41.72

The clean-only verifier achieves 88.00% SR, only 0.20 percentage points below standard FFDC, with increases of 0.2 s in completion time and 0.03 in planning calls. This supports transfer from clean to randomized environmental conditions within the same task families.

Together, these results support the usefulness of the learned verification signal for actual execution and environmental transfer. We have included these analyses in Appendices C and D of the revised manuscript.

Thank you for requesting details of the synthetic negative data generation. We have detailed the four corruption operators and their parameters in Appendix A, “Analysis of verifier supervision,” under “Synthetic negative generation,” with a cross-reference in Section 3.3, “Training strategy and dataset construction.” We randomly sample a action window from the demonstrations and apply one of four corruption operators, each selected with 25% probability. Under the default settings:

  1. Temporal swap: Perform two swap attempts. For each attempt, independently sample two time indices uniformly and exchange their complete action vectors if the indices differ.

  2. Gripper flip: Reverse the intended opening/closing direction of the gripper commands across the action window, keeping the resulting values within the valid gripper-opening range. Other action dimensions remain unchanged.

  3. Late-stage Gaussian noise: Preserve the first half of the window and add independent Gaussian noise, ϵ∼N(0,0.52)\epsilon \sim \mathcal{N}(0, 0.5^2), to all action dimensions in the second half, approximating reduced prediction accuracy toward the end of a predicted action chunk.

  4. Tail scaling: Uniformly sample a suffix length L∈{1,…,16}L \in \{1, \ldots, 16\} and a factor r∈[0.5,0.9]r \in [0.5, 0.9], then multiply all action dimensions in the last LL steps by the same factor.

Each corrupted sample is assigned the negative label y=0y=0 by construction, without additional rollout verification.

AuthorsOpenReview ↗

Response to Q3: Real-world experiments

Q3: Real-world experiments

Thank you for highlighting the limited number of real-world evaluation trials. We have summarized the expanded evaluation in Section 4.2, “Real-world experiments,” and provided the original, additional, and combined trial counts, together with task-wise 95% Wilson confidence intervals, in Appendix G, “Expanded real-world evaluation” (Table 10). Our initial evaluation used 20 trials per task and method, following the evaluation scale used for several real-world tasks in Motus [2]. We have now conducted 20 additional trials per task for Base-Motus and FFDC-WAM, increasing the evaluation to 40 trials per task. The updated results are:

TaskMethodOriginal 20 trialsAdditional 20 trialsCombined 40 trials
Bowl PouringBase-Motus-LC1613/20 (65%)14/20 (70%)27/40 (67.5%)
Bowl PouringFFDC-WAM15/20 (75%)15/20 (75%)30/40 (75.0%)
Clean TableBase-Motus-LC1612/20 (60%)11/20 (55%)23/40 (57.5%)
Clean TableFFDC-WAM15/20 (75%)16/20 (80%)31/40 (77.5%)
Put BlockBase-Motus-LC1614/20 (70%)15/20 (75%)29/40 (72.5%)
Put BlockFFDC-WAM16/20 (80%)16/20 (80%)32/40 (80.0%)
Put CupBase-Motus-LC1614/20 (70%)12/20 (60%)26/40 (65.0%)
Put CupFFDC-WAM16/20 (80%)17/20 (85%)33/40 (82.5%)
Stack BlocksBase-Motus-LC1611/20 (55%)12/20 (60%)23/40 (57.5%)
Stack BlocksFFDC-WAM12/20 (60%)13/20 (65%)25/40 (62.5%)
Stack BowlsBase-Motus-LC1615/20 (75%)14/20 (70%)29/40 (72.5%)
Stack BowlsFFDC-WAM16/20 (80%)15/20 (75%)31/40 (77.5%)
OverallBase-Motus-LC1679/120 (65.83%)78/120 (65.00%)157/240 (65.42%)
OverallFFDC-WAM90/120 (75.00%)92/120 (76.67%)182/240 (75.83%)

With equal trial counts across tasks, the combined mean success rate is 75.83% for FFDC-WAM, compared with 65.42% for Base-Motus-LC16, an improvement of 10.42 percentage points. FFDC-WAM maintains a higher observed success rate on all six tasks. These additional trials support the consistency of the observed performance advantage. We have included the expanded evaluation results and trial counts in Appendix G of the revised manuscript.

Reviewer_N1ztOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FFDC-WAM, an adaptive action execution framework for World Action Models (WAMs). Instead of executing fixed-length action chunks, the method introduces a lightweight verifier that compares predicted visual dynamics with real observations to decide whether to continue execution or replan. The authors also introduce Mixture-of-Horizon Training to support longer action predictions. Experiments on RoboTwin and six real-world manipulation tasks demonstrate improved success rates and reduced inference costs compared with several fixed-horizon and adaptive execution baselines.

Critical strengths
  1. The motivation is reasonable. Adaptive action execution can improve the trade-off between inference efficiency and closed-loop responsiveness, particularly for computationally expensive WAMs.
  2. The experimental evaluation is fairly comprehensive, including 50 RoboTwin tasks and six real-world manipulation tasks on a Unitree G1D. The reduction in WAM inference calls and task completion time is encouraging.
Critical weaknesses
  1. The core idea of using a learned confidence or consistency score to determine when to replan is closely related to existing adaptive action chunking and uncertainty-based execution methods. Although FFDC incorporates predicted visual dynamics, it is unclear whether the proposed attention architecture provides a sufficiently novel contribution beyond existing verification-based approaches.
  2. In Table 1, FFDC improves the average success rate over LC-64 by only 1.44 percentage points under clean conditions and 0.80 points under randomization. In real-world experiments, FFDC improves success over LC-64 but actually increases completion time and inference calls slightly. These results suggest that much of the efficiency improvement comes from longer action chunks rather than the proposed verifier.
  3. The verifier is trained using successful/failed rollouts and synthetic action corruptions. It is unclear whether these examples adequately represent real-world execution failures, especially under unseen object configurations, contact dynamics, or environmental disturbances. The evaluation does not sufficiently isolate how well the verifier generalizes beyond the training distribution.
Questions for authors
  1. What is the fundamental difference between FFDC and existing learned verification or adaptive action chunking methods? Is the proposed attention architecture necessary, or would a simpler learned classifier achieve comparable results?
  2. Given the relatively small improvements over LC-64 on RoboTwin, can the authors better isolate the contribution of the verifier from Mixture-of-Horizon Training and longer action execution?
  3. How well does the verifier generalize to unseen disturbances or failure modes that are not represented by the synthetic action corruptions used during training?
  4. How sensitive is the method to inaccurate future visual predictions from the WAM? Could prediction errors cause unnecessary replanning or, more concerningly, allow invalid actions to continue?
9 comments
AuthorsOpenReview ↗

Response to W1 and Q1: Comparison with Existing Methods

Dear Reviewer N1zt,

Thank you for your careful review and constructive feedback. Due to the 5,000-character limit per response, we have divided our replies into multiple comments.

W1 and Q1: Comparison with Existing Methods

Thank you for this question. We have expanded the comparison with related adaptive execution methods in Section 2. The comparison with a simpler learned classifier is provided in our next response to Q1.

FFDC is a lightweight inference-time decision module designed for WAMs: it turns the visual future already predicted jointly with an action plan into evidence for deciding whether the remaining actions should still be executed. After a WAM planning call, FFDC reuses the cached predicted visual, action, and semantic tokens and incorporates newly acquired observations. A verification check requires no additional WAM call, resampling of candidate action plans, or regeneration of predicted futures. The structured attention links actions to their temporally corresponding predicted dynamics, real observations, and task semantics to assess the validity of the remaining segment.

The original experiments support the effectiveness and computational practicality of this design. In our A100 evaluation, each FFDC check takes 0.015 s, compared with 1.9 s for the SGAC implementation that requires fresh WAM inference (Section 4.1). On randomized RoboTwin, FFDC achieves 88.20% SR, versus 84.40% for SGAC and 85.44% for latent cosine gating (Table 1). Thus, the reported results support using the WAM's existing predictions for efficient execution verification and show a benefit over a simple visual-similarity trigger.

The following table clarifies the relationship to adaptive execution methods. References [4-6] were already cited in the submitted manuscript; [1-3] are additional related works identified during rebuttal, which we have now cited and discussed in Section 2, “Related work,” of the revised manuscript.

MethodDecision basisDecision timingKey distinction from FFDC
PAV [1] (2026-10-15)A learned predictor forecasts consequences in policy representation space; validity scores select a prefix.Before prefix execution.Requires an additional future predictor; does not reassess the selected prefix using observations arriving during execution.
CA³C[2] (2026-10-06)Predicted disagreement sets the horizon; consensus selects an action candidate.After candidate generation, before execution.Requires multiple candidates and separate future rollouts; does not reassess the selected prefix during execution.
Knowing When to Stop [3] (2026-09-16)Action-to-observation cross-attention entropy determines the horizon.Before chunk execution.New observations do not update the selected horizon within the active chunk.
AutoHorizon: VLA Knows Its Limits [4] (2026-06-20)Action self-attention estimates the horizon.Once per policy call, before execution.Does not reassess the active chunk when unexpected deviations occur.
AAC: Adaptive Action Chunking at Inference-time [5] (2026-04-10)Entropy of parallel action samples determines the prefix.At the planning boundary.Requires multiple action samples; FFDC verifies one cached joint prediction.
SGAC: Self-Guidance and Adaptive Chunking [6] (2025-10-14)Similarity between queued and newly predicted actions determines queue updates.Fresh policy inference using updated observations.Requires fresh policy inference for consistency checks; FFDC reuses cached WAM predictions.
FFDC-WAM (ours)Structured attention over cached predictions, actions, semantics, and real observations.During execution of the current plan.No additional WAM inference or candidate sampling for verification; 0.015 s per check on an A100.

[1] Kim, Minkyu, et al. "Policy-Aligned Predictive Verification for Adaptive VLA Action-Chunk Execution." arXiv preprint arXiv:2610.21434 (2026).

[2] Li, Yuyan, et al. "Commit While Futures Agree: Consequence-Aware Adaptive Action Chunking for Robot Manipulation." arXiv preprint arXiv:2610.07949 (2026).

[3] Xu, Runze, et al. "Knowing When to Stop: Adaptive Action Chunking via Internal Cross-Attention Dynamics in VLAs." arXiv preprint arXiv:2609.00908 (2026).

[4] Wang, Haoxuan, et al. "Vla knows its limits: Adaptive execution horizons for robot policies." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.

[5] Liang, Yuanchang, et al. "Adaptive action chunking at inference-time for vision-language-action models." arXiv preprint arXiv:2604.04161 (2026).

[6] So, Junhyuk, et al. "Improving generative behavior cloning via self-guidance and adaptive chunking." Advances in Neural Information Processing Systems 38 (2026): 84432-84456.

AuthorsOpenReview ↗

Response to Q1: Necessity of the attention architecture

Q1: Necessity of the attention architecture

We have compared FFDC with existing methods and discussed their key differences in our preceding response to W1/Q1.

Thank you for suggesting a simpler learned classifier.

We have summarized the MLP comparison in Section 4.3, “Ablation study,” under “Verifier architecture,” and provided the experimental setup and complete results in Appendix F, “Comparison with a simpler learned verifier” (Table 9). The relevant question is whether the same evidence can be combined effectively without structured attention. To examine this, we replace the FFDC Transformer with an MLP classifier while retaining the same input modalities, WAM backbone, verifier training data, execution schedule, and evaluation protocol. The MLP uses LayerNorm and GELU in its hidden layers and outputs a scalar confidence score through a sigmoid head. Both variants are evaluated on the five hard RoboTwin tasks under randomization, following the ablation setting in Table 3.

Randomized hard taskMLP SR (%)MLP T (s)MLP CallsFFDC SR (%) — paper Table 3FFDC T (s) — paper Table 3FFDC Calls — paper Table 3
Blocks Ranking Size86.019.42.3093.025.33.27
Hanging Mug29.020.13.0344.027.53.96
Place Mouse Pad87.012.61.2988.012.41.27
Put Object in Cabinet72.022.02.0373.021.72.00
Scan Object85.015.91.2584.015.71.20
Average over five tasks71.818.01.9876.420.52.34

FFDC achieves an average success rate of 76.4%, compared with 71.8% for the MLP, an improvement of 4.6 percentage points. The largest gains occur on Hanging Mug (44% vs. 29%) and Blocks Ranking Size (93% vs. 86%). These results show that the simpler classifier does not match FFDC's execution success rate under this evaluation setting. This comparison supports the benefit of structured attention over the tested MLP for execution reliability, while making the associated time and replanning costs explicit.

AuthorsOpenReview ↗

Repsonse to W2 and Q2: Separating the verifier from MoH and long chunks,

W2 Q2: Separating the verifier from MoH and long chunks

Thank you for raising this point. We have clarified the separate contributions of MoH, long-chunk execution, and online verification in Section 4.3, “Ablation study,” under “MoH training and decision threshold.” LC-64 already performs well on the 45 easy RoboTwin tasks, achieving 88.98% and 88.89% SR under clean and randomized conditions. However, its SR falls to 73.80% and 74.00% on the five hard tasks. FFDC improves these hard-task results to 76.00% and 76.40%, respectively. Since easy tasks constitute 90% of the benchmark, the overall average places substantially more weight on settings where fixed long-chunk execution already works well.

To isolate the verifier's contribution, LC-64 and FFDC use the same MoH-trained 64-action backbone. Their comparison therefore holds backbone training and the maximum execution horizon fixed while changing the execution strategy. Furthermore, Tables 1 and 3 provide the following comparison on the five hard tasks under randomization:

Backbone trainingFixed LC-64 SRFFDC SRGain from verification
Without MoH68.0%71.4%+3.4 pp
With MoH74.0%76.4%+2.4 pp

Verification improves SR both with and without MoH. Separately, MoH improves fixed execution by 6.0 pp and adaptive execution by 5.0 pp. These comparisons support complementary contributions from long-horizon training and online verification.

The benefit is more pronounced in real-world execution: FFDC increases SR from 61.67% to 75.00%, a 13.33 pp improvement over LC-64, while only needs 0.22 additional planning calls. Meanwhile, completion time remains 51.31% lower than the 84.42 s reported for Base-Motus-LC16, retaining much of the efficiency benefit of long-chunk execution.

We agree that longer chunks provide a substantial part of the acceleration over short-chunk execution. FFDC adds selective correction to improve reliability, with an additional execution cost. We have made this attribution explicit in Section 4.3, distinguishing the efficiency gained from long chunks from the additional robustness provided by the verifier.

Comments on the paper

1 comment

Dear Reviewers and Area Chairs,

Thank you very much for your careful reviews and constructive suggestions. Your feedback has helped us strengthen the experimental evidence and improve the clarity of our paper. We have revised the manuscript and provided point-by-point responses to each reviewer. The main changes are marked in blue in the revised paper:

1. Isolated the verifier’s contribution (Section 4.3 and Appendix E).
Addresses: Reviewer N1zt (W2, Q2); Reviewer zfAJ (W2, Q2).
Verification improves hard-task SR by 2.4/3.4 percentage points with/without MoH. On the original 16-action backbone, it improves SR by 2.08 points.

2. Clarified novelty and compared verifier architectures (Sections 2 and 4.3, and Appendix F).
Addresses: Reviewer N1zt (W1, Q1).
We expanded related-work comparisons and added an MLP baseline. With the same inputs and evaluation protocol, FFDC achieves 76.4% SR, versus 71.8% for the MLP.

3. Expanded supervision analyses (Section 3.3 and Appendix A).
Addresses: Reviewer zfAJ (W3, Q1); Reviewer Jhu1 (W1, Q1).
We detailed synthetic-negative generation. Classification accuracy is 91.65% on actual-rollout samples and 93.75% on demonstration-derived samples.

4. Tested environmental generalization (Section 4.3 and Appendix C).
Addresses: Reviewer N1zt (W3, Q3); Reviewer zfAJ (W3).
A verifier trained only on clean-setting data achieves 88.00% randomized SR without fine-tuning, compared with 88.20% for standard training.

5. Analyzed execution risk and visual inputs (Section 4.3 and Appendix D).
Addresses: Reviewer N1zt (Q4); Reviewer zfAJ (W3).
Scores correlate with matched rollout outcomes (AUROC 0.786; AUPRC 0.761). Removing predicted visuals or real observations reduces SR by 4.8/4.0 points, respectively.

6. Clarified execution and reported chunk distributions (Section 3.1 and Appendix B).
Addresses: Reviewer Jhu1 (W2, Q2, Q3).
For 64-action plans, verification occurs every 16 actions, with k=16k=16. Full-length chunks account for 73.26%/77.86% under clean/randomized conditions.

7. Expanded real-world trials (Section 4.2 and Appendix G).
Addresses: Reviewer zfAJ (Q3).
We increased evaluation to 40 trials per task for Base-Motus and FFDC. Combined SR is 65.42%/75.83%, respectively, with task-wise 95% Wilson confidence intervals reported.

8. Clarified settings and improved presentation (Abstract and Sections 1 and 4).
Addresses: Reviewer zfAJ (W1); Reviewer Jhu1 (Q4, Q5, Q6).
We clarified checkpoint, training, and evaluation settings; corrected terminology and naming; adjusted layout; and provided annotated real-robot videos at 1× speed.

Thank you again for your time and constructive feedback.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.