acceptodds
Under review as a conference paper at ICLR 2027

Future Anchored Verification and Online Recovery for World Action Models

Abstract

World action models (WAMs) have emerged as a promising paradigm for robotic manipulation. They act by first predicting how a task should be performed and then decoding the actions from that future. However, the remaining actions are invalid once execution drifts from the prediction. Simply replanning from the already out of distribution state rarely restores what the task still requires; existing execution monitors decide when to stop, but not what to restore. We observe that the answer is already in hand: the future the WAM predicted before acting depicts exactly the states it intended to pass through. We introduce FAVOR (Future Anchored Verification and Online Recovery), a lightweight framework that keeps these predicted frames as anchors and uses them for verification and recovery. An Anchor Verifier compares each observation with its anchor, together with the executed actions, to flag deviations that break the task. Anchor-Guided Recovery then hands the alarmed anchor to a vision-language model, which turns it into a short corrective instruction that WAM could execute under strengthened instruction guidance to get back to the intended future before resuming the task. FAVOR raises the task success of the base WAM from 97.85% to 98.10% on LIBERO and from 72.60% to 72.98% on LIBERO-Plus without modifying the policy.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
robot manipulation, world models, failure detection and recovery, CFG

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_purgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FAVOR, an execution verification and recovery framework for world action models. It retains the visual futures predicted before action execution and uses them as anchors for both detecting deviations and specifying recovery targets. A learned Anchor Verifier combines observation–prediction features, executed actions, and temporal information to identify deviations. Upon an alarm, a VLM translates the current observation and the retained anchor into a corrective instruction, which the frozen WAM executes with strengthened classifier-free guidance. Experiments on LIBERO and LIBERO-Plus show improved failure detection and modest increases in task success, from 97.85% to 98.10% and from 72.60% to 72.98%, respectively.

Critical strengths
  1. Reusing predicted futures as both verification references and recovery targets is intuitive and provides a coherent connection between failure detection and correction.
  2. The Anchor Verifier achieves strong detection performance on LIBERO, outperforming the evaluated distance-based and learned detectors.
Critical weaknesses
  1. The success-rate gains are marginal: only 0.25 percentage points on LIBERO and approximately 0.38 on LIBERO-Plus. No uncertainty estimates are reported to establish their reliability.
  2. End-to-end comparisons mainly involve internal variants of Fast-WAM-IDM. Comparisons with strong WAMs such as LingBot-VA and Cosmos Policy, as well as competing execution-monitoring methods, are missing. Cross-backbone generality is also untested.
  3. Simulation evaluation is limited to two closely related benchmarks, LIBERO and LIBERO-Plus. Additional benchmarks such as RoboTwin 2.0 would provide stronger evidence of generalization.
  4. No real-robot experiments are provided, leaving the effectiveness of verification and recovery under physical execution errors insufficiently validated.
  5. Episode-level detection metrics do not establish timely or effective recovery. Detection delay, successful rescue rates, and harmful intervention rates are not reported.
  6. The additional inference latency from verification, VLM calls, and strengthened CFG is not quantified, making the practical trade-off against the small success gains unclear.
Questions for authors
  1. Are the reported gains consistent across random seeds? Please provide confidence intervals or paired statistics on successful rescues and harmful interventions.
  2. How does FAVOR compare with LingBot-VA and Cosmos Policy under matched evaluation settings, and with execution-monitoring methods such as FFDC using the same backbone? Is there evidence that FAVOR also benefits another compatible WAM backbone?
  3. Do you have results on a more distinct simulation benchmark, such as RoboTwin 2.0, to assess generalization beyond the LIBERO task family?
  4. Do you have real-robot results demonstrating effective verification and recovery under physical execution errors and perception noise?
  5. How early are task-breaking deviations detected, and what fraction of triggered interventions ultimately rescues the task?
  6. What additional inference latency does FAVOR introduce, and how does it affect overall task completion time?
Reviewer_wsGoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FAVOR, an inference-time framework for world action models that uses the model's own predicted future frames as anchors for both execution verification and recovery. An Anchor Verifier monitors whether execution remains aligned with the predicted future, while Anchor-Guided Recovery uses a VLM-generated corrective instruction and strengthened classifier-free guidance to steer the policy back toward the intended trajectory without modifying the underlying WAM.

Critical strengths

A key strength of the paper is the central insight that although execution deviations invalidate the remaining actions in a chunk, the predicted future states themselves remain useful recovery targets. This leads to a simple and well-motivated framework that unifies verification and recovery around the same future predictions.

The authors also provide convincing evidence that the frozen WAM relies far more on visual context than language instructions. Their attention analysis shows almost identical attention patterns when using the original instruction and an empty prompt, which directly motivates the use of stronger classifier-free guidance during recovery.

The Anchor Verifier is particularly strong. By combining DINOv2-based observation-anchor comparisons with action and horizon information in a temporal model, it achieves an AUROC of 0.992, an AUPRC of 0.781, and 97.1% failure coverage at a 5% false alarm rate, substantially outperforming both learned detectors and distance-based baselines.

Another practical advantage is the plug-and-play nature of the method. FAVOR operates entirely at test time on top of a frozen Fast-WAM model, uses an off-the-shelf VLM, and requires only a lightweight verifier without any policy retraining.

Critical weaknesses

Despite the strong detection performance, the gains in end-to-end task success are relatively modest. FAVOR improves average success on LIBERO from 97.85% to 98.10% and on LIBERO-Plus from 72.60% to 72.98%, suggesting that many detected failures either occur infrequently or are difficult to recover from in practice.

The recovery benefits also appear limited on the more challenging LIBERO-Plus benchmark. Since verification and recovery depend on anchors generated by the underlying WAM, the approach is naturally constrained when the predicted future itself becomes less reliable under distribution shift.

Finally, the paper lacks some useful ablations of the Anchor Verifier. In particular, it would be valuable to quantify the contribution of action embeddings, horizon embeddings, and the two-stage training procedure, as the current results do not isolate which components are most responsible for the verifier's strong performance.

Questions for authors

What is the extent of performance degradation in the Anchor Verifier, measured by AUROC, AUPRC, and coverage, when trained only with local perturbations, only with episode-level labels, or without incorporating the executed action and horizon embeddings?

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

The paper proposes FAVOR, which uses the future frames predicted by the World Action Model (WAM) before execution as anchors, while simultaneously performing execution deviation detection and online recovery. The Anchor Verifier combines the current observations, the predicted anchors, and the task-related deviations to determine the task. Once an anomaly is triggered, the Anchor-Guided Recovery uses the visual language model to convert the anchor state into a brief corrective instruction, and through the enhanced classifier-free guidance, guides the WAM to return to the expected state. This method does not require modifying the basic strategy.

Critical strengths
  1. The core idea is concise and targeted: directly reuse the future predictions generated by WAM, simultaneously converting the "plan representation" into verification targets and recovery targets, avoiding the need to build an independent target state model.
  2. The method maintains the freezing of basic WAM parameters and incorporates detection and recovery capabilities in a modular way. The deployment interface is clear, and the intrusion into the existing WAM is relatively low.
  3. The design of Anchor Verifier fully utilizes the relationship among observation, anchor point, and action. On LIBERO, it achieves an AUROC of 0.992, an AUPRC of 0.781, and a failure coverage of 97.1%, significantly outperforming the baseline that only uses feature distances or internal representations of strategies.
  4. The experiments include direct re-planning, no anchor point correction, and complete recovery as controls, enabling the distinction of the contributions of anomaly detection, general correction, and anchor-guided recovery.
Critical weaknesses
  1. The success rate improvement is relatively small: The absolute gains of LIBERO and LIBERO-Plus are only 0.25 and 0.38 percentage points respectively. It is difficult to determine the stability of the gains.
  2. It has only been verified on Fast-WAM-IDM, which is not sufficient to support universal conclusions for different WAM architectures.
  3. The recovery process relies on additional VLM generating correction instructions and rejudging whether the recovery is completed, but there is a lack of system analysis of delay, computing cost, and error instruction ratio. What is the maximum number of recoveries and what will be done if the correction fails.
  4. The method defaults that the original future prediction contains correct and recoverable task states. When the future prediction itself is incorrect or the target is unreachable, the anchor may generate an incorrect recovery target, and the related failure modes lack quantitative evaluation.
  5. The training part of the validator depends on artificially constructed action perturbations and episode-level labels. It has not yet explained its generalization ability to more complex real disturbances, continuous contact errors, and real robot sensor noise.
  6. Can correction examples generated by VLM be provided? Are the correction examples generated by VLM too deviated from the normal task instructions in the training set? Can WAM understand these corrective instructions?
Questions for authors

Please refer to Weakness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.