acceptodds
Desk rejected at ICLR 2027

Revision, Not Restart: Revisable Visual Plans for Closed-Loop World-Action Models

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

World–action models use predicted visual futures to condition robot actions, yet execution feedback can invalidate parts of a prediction while leaving its task structure useful. We propose Revisable Temporal Planning (RTP), which maintains the visual future as a persistent action condition and revises it after feedback. Its central mechanism is a learned revision bridge: it resumes an intermediate state saved during visual generation and adapts its continuation to current observations. Visual and action supervision connect this revision to subsequent control. Time-aware history supplies observed evidence, and an adaptive policy selects retention, bridge revision, or fresh replanning from new noise before decoding the next action. On RoboMME and RMBench, RTP achieves task-averaged success rates of 48.6% and 84.8%, respectively. Matched comparisons support learned continuation; estimated checkpoint-source and action-prefix effects are positive but less precisely resolved. These results connect feedback-driven visual-plan revision to closed-loop task performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
world-action models, robotic manipulation, closed-loop control, visual plan revision, learned revision bridge, adaptive replanning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_23GgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Revisable Temporal Planning (RTP), a framework that enables world-action models to revise persistent visual plans based on execution feedback, rather than always restarting the generation process. The method introduces a learned revision bridge that resumes from intermediate diffusion states, a time-aware history representation to incorporate observed evidence, and an adaptive selector that chooses between retaining the current plan, revising it, and performing fresh replanning. Experiments on RoboMME and RMBench report task success rates of 48.6% and 84.8%, respectively. The results suggest that feedback-conditioned plan revision can improve task performance, although the contributions of individual components and the advantages of some design choices remain insufficiently established.

Critical strengths
  1. The paper addresses a practical problem in world-action models: how to incorporate execution feedback without unnecessarily discarding an existing predictive plan. The distinction between retaining, revising, and restarting a plan is meaningful for closed-loop planning.
  2. The paper includes a range of experiments and ablations covering revision strategies, checkpoint sources, history configurations, and training losses. The results provide encouraging evidence for the learned revision mechanism, particularly compared with zero-residual correction.
Critical weaknesses
  1. The ablation studies lack a systematic, incremental analysis from a common baseline, leaving the marginal contributions and interactions of key components unclear.
  2. The advantage of adaptive update selection remains statistically uncertain. RTP achieves 48.63% success, compared with 47.38% for fixed Bridge-10 and 46.38% for binary retain/fresh, but paired confidence intervals are not reported.
  3. The independent benefit of saved checkpoints is not fully established. Although native checkpoints outperform root-noise reconstruction (47.4% vs. 44.6%), the reported 95% confidence interval for the difference includes zero.
  4. The reliability of discrepancy metrics as indicators of task performance remains unclear. In Tables 13–14, recurrent fitting substantially reduces discrepancy tolerance exceedances but improves task success by only 0.125 percentage points, suggesting that improvements in these metrics do not necessarily translate into better task performance.
Questions for authors
  1. Could the authors provide a more systematic component-wise ablation, including controlled comparisons that quantify the marginal contributions of the major components and examine potential interactions between checkpoint reuse, learned revision, and adaptive update selection?
  2. Could the authors report paired confidence intervals for the success-rate differences between RTP and the fixed Bridge-10 and binary retain/fresh baselines, to clarify whether the observed improvements are statistically reliable?
  3. Given the uncertainty in the checkpoint-source comparison, how essential is the use of saved native checkpoints to the overall performance of RTP? Could the authors further isolate its contribution from that of the learned revision bridge?
1 comment

Thank you for the detailed response and additional experiments. The new ablations and statistical analyses address several of my concerns and strengthen the empirical evaluation. I appreciate the authors' efforts in providing these additional results and clarifications. After considering the response, I will maintain my current rating.

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

This paper proposes Revisable Temporal Planning (RTP) for closed-loop world–action models. RTP maintains a persistent visual plan and updates it using execution feedback by retaining the remaining prediction, resuming saved intermediate generation states with a learned correction, or replanning from fresh noise. The method combines time-aware observation history, a feedback-conditioned revision bridge, and an adaptive update selector. Implemented on LingBot-VA, RTP achieves success rates of 48.6% on RoboMME and 84.8% on RMBench. The paper provides ablations on learned correction, checkpoint sources, action conditioning, and computational cost.

Critical strengths
  1. The paper addresses a meaningful closed-loop control problem: revising partially useful visual plans after execution feedback. The explicit separation of observed history and predicted futures provides a coherent formulation.
  2. The matched ablations are informative. In particular, learned correction improves success by 5.75 percentage points over zero residual, with a paired 95% interval of [3.00, 8.50], providing clear evidence for this component.
Critical weaknesses
  1. No real-robot evaluation. The paper provides no real-world experiments, leaving the practical effectiveness of feedback-driven plan revision unverified.
  2. Limited simulation coverage. Evaluation is restricted to RoboMME and RMBench. Additional benchmarks, such as LIBERO-Pro and RoboTwin 2.0, would help establish broader generalization.
  3. Inconclusive evidence for saved checkpoints. The advantage over original-root-noise reconstruction has a 95% confidence interval crossing zero, leaving the distinctive contribution of saved intermediate states insufficiently established.
  4. Limited backbone generality. RTP is evaluated only on LingBot-VA, so its applicability to other world–action models remains unclear.
  5. Uncertain real-time applicability. Primary evaluation pauses environment dynamics during inference, while each controller call takes approximately 1.05 seconds on an H200. The controlled-delay tests do not fully establish performance under continuous execution.
  6. Limited gains over simpler update policies. RTP improves success over fixed Bridge-10 and binary retain/fresh by only 1.25 and 2.25 percentage points, respectively, without paired confidence intervals for these comparisons.
Questions for authors
  1. Can you provide real-robot experiments to validate whether feedback-driven plan revision improves performance under realistic sensing and execution errors?
  2. Can you provide results on additional simulation benchmarks, such as LIBERO-Pro and RoboTwin 2.0, to assess generalization beyond RoboMME and RMBench?
  3. What architectural requirements does RTP impose on the underlying WAM? Can it be applied to models such as Motus or UWM without substantial redesign?
  4. Given that the saved-checkpoint advantage over original-root-noise reconstruction has a confidence interval crossing zero, what evidence establishes the benefit of preserving intermediate generation states?
  5. How does RTP perform when environment dynamics continue during inference and execution delays reflect the measured controller latency?
  6. Can you report paired confidence intervals and results across independent training seeds for RTP versus fixed Bridge-10 and binary retain/fresh, to establish whether the additional complexity yields reliable gains?
Reviewer_wJRhOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Revisable Temporal Planning (RTP) for closed-loop world–action models. Instead of reusing a predicted visual plan unchanged or regenerating it from noise after execution feedback, RTP resumes intermediate denoising states saved during plan generation. It then continues them under current observations, using a learned velocity correction trained with visual and action supervision. A time-aware pyramid history supplies observed evidence. A calibrated selector chooses among retention, Bridge-5, Bridge-10 and fresh replanning before the next action is decoded. On RoboMME and RMBench, RTP reports task-averaged success of 48.6% and 84.8%, with matched internal controls for learned correction, checkpoint source and action conditioning.

Critical strengths
  1. A real problem with real difficulty. The paper asks a clear, practical question: how to revise an already generated visual plan after execution feedback, rather than reusing it unchanged or regenerating it from noise. The difficulties it identifies are real. Saved denoising states were formed under earlier observations, and better visual prediction does not guarantee better control.
  2. Broad set of controls. The ablations cover residual removal, checkpoint source, a 2×2 action-input factorial, loss terms, history factors, Fresh-k+C compute ladders, recurrent fitting, and delay, spatial-shift and horizon sensitivity tests.
Critical weaknesses
  1. Relevant SOTA is missing, and the gap over prior work is narrow.
    • Not cited: Real-Time Chunking (Black et al., NeurIPS 2025) continues plans under execution delay without retraining. Bidirectional Decoding (Liu et al., ICLR 2025) balances keeping the current plan against reacting to new observations. UVA (Li et al., RSS 2025) is a core joint video–action model.
    • Cited but not compared: FFDC decides whether to continue or replan. Feedback World Model corrects predictions from feedback without retraining. Diffusion ReRoll revises video–action plans by re-noising. Streaming Diffusion Policy already updates partially denoised buffers under new observations.
    • The external baselines are all memory-focused models, and none of them revises plans under feedback. It therefore remains unclear whether RTP's learned checkpoint revision outperforms simpler, training-free alternatives.
  2. The contribution is mainly engineering. The method is a three-stage pipeline: a pyramid history sampler, a calibrated mode selector, and two fixed checkpoint depths. The paper itself calls the depths an interface-defined choice rather than a principled one. No analysis explains when resuming a saved state should outperform restarting.
  3. The comparative claim is not substantiated.
    • On RMBench, RTP exceeds MemoryWAM by only +1.8 pp (84.8 vs. 83.0), with no interval.
    • The baseline entries appear to be taken from Yang et al. (2026). They match the published values exactly and are described as retaining their reported configurations. The paper reports no paired comparison against them.
    • A rough independent-binomial estimate (about ±3.4 pp) suggests this margin may be within evaluation noise.
    • Part of the overall gain comes from base adaptation: the authors' fine-tuned LingBot-VA reaches 79.9, compared with 78.2 published.
    • The paper does not explain why RTP outperforms memory-based methods such as MemoryWAM
Questions for authors

Refer Weaknesses Section

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

Related papers

Discussion (0)

Sign in to comment.