acceptodds
Under review as a conference paper at ICLR 2027

REVIVE: Inference-Time Return with Phase-Aware Controller for VLA Models

Abstract

Vision-Language-Action (VLA) models have shown strong general-purpose robot control by leveraging large vision-language models. However, imitation-learned VLA policies remain brittle in multi-step manipulation: small execution errors can compound off the demonstrated trajectory, and the policy may silently advance to a later phase without completing the current subgoal, which we formalize as . Existing recovery approaches often rely on additional human interventions, reward supervision, or offline retraining. We propose REVIVE, an execution-time recovery framework that improves VLA success without retraining the underlying policy. Instead of learning explicit recovery from every failure, REVIVE intervenes before local deviations escalate. An external Return Controller monitors phase-local progress, boundary evidence, and action-effect signals to detect phase-level stalls; upon detection, it performs phase-local rollback to the latest valid phase anchor, clears the current action chunk, and triggers replanning. Failure-directed decoder adaptation generates alternative actions, while failure memory ranks candidates to avoid repeating past mistakes. Experiments on simulated manipulation benchmarks and two real-robot long-horizon tasks show that REVIVE improves task success, including under out-of-distribution task configurations and object changes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generalist Robot Policy, Vision-Language-Action Models, Close-loop Recovery

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_U8vVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes REVIVE, which monitors execution progress and phase boundaries, performs limited local action rollback upon detecting anomalies, and regenerates actions. To reduce repeated failures, the method incorporates Decoder Adaptation, which temporarily adjusts some action decoder parameters, and Failure-Guided Selection, which ranks candidate actions using the similarity and density of recorded failed actions. The authors evaluate overall performance on LIBERO, LIBERO-plus, and two real robot tasks with their OOD settings, and analyze the roles of different recovery components through ablations.

Critical strengths

The issue of stage failure propagation addressed by the paper is of practical significance. Local rollback can attempt to preserve previously completed task progress, while failure-guided action generation and selection target the specific problem of repeatedly executing similar failed actions. Experiments cover multiple VLA backbones, simulation perturbations, and real robots, and the reported results show that the overall approach has the potential to improve task completion rates. The authors also explicitly discuss limitations such as action replay being unable to restore complete physical states, and failure density not being equivalent to success probability.

Critical weaknesses

The critical ablation results in the main text and appendix are inconsistent. For the π0 model on LIBERO with configurations of the same name, Table 4 reports average success rates of 97.45% and 97.75% for Return + FGS and Return + DA respectively, while Table 8 reports corresponding values of 91.10% and 92.00%. However, the average results for the base model, Return, and complete REVIVE are consistent across both tables.

Questions for authors

Please explain why the results for π0's Return + FGS and Return + DA in Table 4 and Table 8 are different. Do they come from different implementations, checkpoints, evaluation samples, or experimental settings? If so, please clearly label the configurations and differences for each row, and indicate which set of results should serve as the primary basis for ablation; if this is a reporting error, please unify the tables, related main text conclusions, and cost statistics.

1 comment
AuthorsOpenReview ↗

Response to Reviewer U8vV

We thank the reviewer for checking the tables closely; this was the most consequential issue in the review.

Q1: Why do Table 4 and Table 8 differ for π₀ Return + FGS and Return + DA?

"For the π0 model on LIBERO with configurations of the same name, Table 4 reports average success rates of 97.45% and 97.75% for Return + FGS and Return + DA respectively, while Table 8 reports corresponding values of 91.10% and 92.00%."

This is a reporting error on our part, not a difference in implementation, checkpoint, evaluation samples, or experimental settings. The two rows in the main table were copied from an inconsistent intermediate result, while the appendix contains the final configuration. The reviewer's observation about which rows agree is exactly right: Baseline, Return, and complete REVIVE were transcribed correctly, and only these two were not.

Corrected π₀ ablation, now identical in both tables:

MethodAvg. success rate
Baseline90.80%
Return91.70%
Return + FGS91.10%
Return + DA92.00%
REVIVE97.90%

Following the reviewer's request to indicate which results should serve as the primary basis: the appendix configuration is the primary one, and we have unified the main table, the appendix table, the accompanying text, and the cost statistics around it. We have also re-derived the bold and underline markings from the corrected values, labeled the checkpoint, per-suite episode count, and controller protocol in both captions, and noted in the main caption that the Return-to-Start row uses a different evaluation budget (400 versus 2,000 episodes) — a detail that was previously only in a source comment.

As we noted in our response to Reviewer sirW, who raised the same discrepancy, the corrected numbers do not weaken the ablation: the complete method exceeds every single-component variant on every suite, with +5.9 pp average and +10.4 pp on the long-horizon suite over the strongest single component.

Reviewer_sirWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper characterizes “silent phase failures” as one of the common failure modes of VLA models for robot manipulation, especially in tasks involving multiple steps. It proposes a method to: detect such failures at execution time, do partial roll backs to a state where the model can re-attempt the subtask or "phase", and steer the newly sampled trajectories away from previously attempted failing trajectories. The proposed approach is evaluated on the LIBERO and LIBERO+ simulation benchmarks with three open weights VLA models as the base, demonstrating performance gains. The method is also evaluated in real-world robot experiments on two multi-step pick-and-place tasks.

Critical strengths
  • The work addresses a relevant problem in action chunking VLA models
  • The presentation is clear and fairly easy to understand
  • The evaluation spans multiple backbones and simulation and real-world settings, and is generally sound
Critical weaknesses
  • Relevance of the work. My first and main concern is related to the relevance of the work given the state of the field. I think the need for the specialized machinery is not sufficiently justified in light of recent developments. First, the progress monitoring and recovery decisions implemented by the Return controller could plausibly be handled by a frontier VLM, as it has been shown by recent orchestration frameworks using VLMs for subgoal planning and task monitoring. The performance gains introduced by the DA and FGS components appear to diminish when switching to a more capable model, pi0.5. This raises the question of whether these mechanisms address a persistent limitation or weakness that stronger base policies increasingly overcome.
  • Inconsistent backbone choice across failure analysis, main experiments, and ablation study. Table 1 motivates the method by looking at the failures of GR00T N1.7 on the LIBERO benchmark, the component ablation study in the main paper uses pi0, and the robot experiments use pi0.5. While including multiple backbones in the main experiments to demonstrate transfer is a good idea, switching backbone for each of the supporting experiments does not establish a consistent connection between the diagnosed failure mode in real-world scenarios.
  • Reporting issues on the ablation study. First, the numbers on the ablation study in the main paper do not match those in the appendix. In the main paper, the avg scores for are 90.80 - 91.70 - 97.45 - 97.75 - 97.90, in the appendix B.2 the avg scores for pi0 in the complete ablation results are 90.80 - 91.70 - 91.10 - 92.00 - 97.90. Secondly, the main paper reports the ablation study numbers for pi0 instead of pi0.5, where according to the appendix, the gains are marginal. Which brings me to the next point.
  • Benchmark choice for failure analysis and ablation study. The LIBERO benchmark is almost saturated by some of the selected baselines - mainly pi0.5 - so the most interesting results are those on LIBERO+. However, both the failure analysis and the ablation study are done LIBERO, where the gains are marginal.
Questions for authors
  • Could a frontier VLM replace the learned progress monitor and Return decision rules? What advantages does the proposed controller offer?
  • How frequently does π₀.₅ exhibit the silent phase failures identified for GR00T? Does the failure analysis support the same motivation across backbones?
  • What explains the discrepancies between the main-paper and appendix ablation results? Is it just a reporting error?
  • Do the ablations result hold on LIBERO+?
4 comments
AuthorsOpenReview ↗

Response to Reviewer sirW

We thank the reviewer for the detailed assessment. We address the four questions in order.

Q1: Could a frontier VLM replace the learned monitor and Return decision rules?

"The progress monitoring and recovery decisions implemented by the Return controller could plausibly be handled by a frontier VLM... What advantages does the proposed controller offer?"

We tested this on the same 2,000-episode LIBERO protocol as the ablations, with π₀.₅ as the base policy. We replaced the learned return decision with a locally deployed 27B open-source VLM (Qwen/Qwen3.8-27B) and kept the rest of the pipeline identical — same rollback, same DA, same FGS — giving the VLM a simple and favourable context:

You are a conservative robot manipulation recovery monitor. Task: {description}. Approximate forward control steps: {forward_hint}. Inspect the current robot image. Decide whether the policy has silently stalled or deviated enough that a local rollback should be triggered. Use RETURN only when there is visible evidence of a stalled/deviated phase; otherwise use CONTINUE. Reply with exactly one word: RETURN or CONTINUE.

(given a single frame plus the task text and a step counter)

The VLM never produced a single RETURN. Across all 1,028 invocations it answered CONTINUE, so it triggered no recovery at all, even though the prompt explicitly invited it to return on visible evidence of a stall. Its agreement with the reference decision was 94.4% over 742 calls on LIBERO-10, 98.4% over 108 on Goal, 98.6% over 147 on Object, and 98.4% over 31 on Spatial — the four groups partition the 1,028 calls exactly.

That agreement is the degenerate part of the result, not a virtue. Because the reference label is RETURN only in a minority of calls, a monitor that always says CONTINUE inherits the majority class for free. Reading the counts backwards, the reference was RETURN in roughly 46 of the 1,028 calls (≈5.7% on LIBERO-10, ≈1.9% on Goal, ≈1.4% on Object, 0% on Spatial), so the VLM agreed on nearly every CONTINUE and missed essentially every reference return — the opposite of what a replacement monitor must do.

The cost side rules it out even if it had worked: the 27B model needs 55.23 GB (four-GPU sum, excluding the base VLA) and 531.9 ms per decision, against 2.39 MB and negligible time for the auxiliary head.

On the same episodes the unmodified π₀.₅ baseline reaches 97.10% and full REVIVE 99.50%, while the VLM-judge configuration reaches 97.55%. Since no RETURN was ever issued, that configuration cannot recover anything; the 0.45 pp over baseline is not distinguishable from noise (95% CI ≈ [−0.55, +1.45] pp), and the meaningful gap is the 1.95 pp the learned head contributes.

We read this as a context problem rather than a model-size problem. Deciding when to return requires a phase-local progress trajectory — how predicted subgoal progress changed over the last window, whether boundary evidence has decayed, what recent arm actions look like. A single image plus a step counter exposes none of it, which is why the VLM defaults to the safe answer. Supplying that context means rebuilding the interface the auxiliary head already consumes, now at ~500 ms and 55 GB per query. The learned head is not a cheaper approximation of a VLM judge; it operates on the signal the decision actually requires.

AuthorsOpenReview ↗

Response to Reviewer sirW

Q2: How frequently does π₀.₅ exhibit silent phase failures? Does the analysis support the same motivation across backbones?

"The performance gains introduced by the DA and FGS components appear to diminish when switching to a more capable model, pi0.5. This raises the question of whether these mechanisms address a persistent limitation..."

We repeated the diagnostic on two further frozen backbones under the same annotation protocol; the full table is now in Appendix B.3.

BackboneVariantsGrasp FailR_silentR_stallTask SR
GR00T N1.7654610.2460.7540.682
π₀.₅654160.3120.6880.745
π₀6431240.5810.4190.379

Three things follow. First, the failure mode is not specific to GR00T — it appears on both π-family backbones under identical rules. Second, and more substantively, the silent share increases monotonically as the policy weakens: 0.246, 0.312, 0.581. Silent transitions are a minority of grasp failures for the stronger policies and a majority only for π₀. The weaker the backbone, the more often a grasp failure becomes a silent phase transition rather than a local stall — which is precisely why the mode deserves a dedicated mechanism rather than being treated as ordinary task failure.

Third, a stronger backbone does not remove the problem. π₀.₅ has far fewer grasp failures (16 versus 124) and much higher success (74.5% versus 37.9%), yet about 31% of its grasp failures remain silent. On LIBERO the gains do narrow for π₀.₅ (+2.40 pp versus +7.10 pp for π₀), and we agree with the reviewer's reading that this reflects the baseline nearing saturation at 97.10% rather than the mechanism failing. On LIBERO-plus, where headroom is larger, the gain is substantial for both models (π₀ 28.86 → 49.85; π₀.₅ 85.10 → 91.05). The recovery problem does not disappear with a stronger backbone — the measurable headroom does.

Q3: What explains the discrepancies between the main-paper and appendix ablation results?

"In the main paper, the avg scores for are 90.80 - 91.70 - 97.45 - 97.75 - 97.90, in the appendix B.2 the avg scores for pi0 in the complete ablation results are 90.80 - 91.70 - 91.10 - 92.00 - 97.90."

This is a reporting error, and we thank the reviewer for catching it. There were no differences in implementation, checkpoint, evaluation samples, or settings. The main table's Return + FGS and Return + DA success rates were copied from an inconsistent intermediate result; the appendix values come from the final π₀ configuration. Baseline, Return, and full REVIVE were transcribed correctly.

We determined which run is authoritative from the cost columns, which must come from the same run as the success rates. The main table's Steps / Ret. / Share entries reproduce the appendix's per-suite interaction records exactly (Return: 220.92 steps, 0.75 returns, 6.07%; Return + FGS: 223.23, 0.77, 6.20%; Return + DA: 219.80, 0.75, 6.07%). The appendix values are therefore correct.

The corrected π₀ ablation, now identical in both tables:

MethodSpatialObjectGoalLongAvg
Baseline93.4096.4092.2081.2090.80
Return95.6098.4091.0081.8091.70
Return + FGS95.2097.8089.6081.8091.10
Return + DA95.0098.6091.2083.2092.00
REVIVE99.40100.0098.6093.6097.90

We have unified the main table, appendix table, accompanying text, and cost statistics, and re-derived the best/second-best markings. The corrected values strengthen the ablation narrative: the complete method exceeds every single-component variant on every suite, +5.9 pp over the strongest single component in average success, concentrated in the long-horizon suite (93.60 versus 83.20, +10.4 pp).

On the reviewer's related point that the main paper ablates π₀ rather than π₀.₅: we agree the π₀.₅ margins are smaller, and Appendix B.2 gives that decomposition alongside GR00T N1.7 so the reader can see both. We retain π₀ in the main text because the separation between single components and the complete method is widest there, which is what makes the component attribution legible; the π₀.₅ numbers are reported in full rather than omitted.

On ablating pi0 rather than pi0.5

"The main paper reports the ablation study numbers for pi0 instead of pi0.5, where according to the appendix, the gains are marginal."

We agree the pi0.5 margins are smaller, and Appendix B.2 reports that decomposition in full alongside GR00T N1.7. We keep pi0 in the main text because it is the backbone where the separation between single components and the complete method is widest, which is what makes the component attribution legible; this was a presentational choice and we should have said so.

Q4: Do the ablation results hold on LIBERO-plus?

We are not able to run the full component decomposition on LIBERO-plus within the current compute budget. Adding three further ablation configurations on top would multiply that budget beyond what we can currently afford. We prefer to state this plainly rather than report a partial or unreproducible grid, and we have not added such a table.

On the reviewer's broader observation that LIBERO is near-saturated while LIBERO-plus is the more informative benchmark, we agree. That is why our robustness evidence and real-robot evaluation carry the main weight, and why we do not lean on the +2.40 pp on π₀.₅ as support for the mechanism.

Reviewer_B4kaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

REVIVE is an inference-time recovery framework that does not require re-training of the basic VLA. It mainly addresses the phased failures and error accumulation in robot operations by monitoring the execution progress, locally reverting the most recent actions, temporarily adjusting the decoder, and filtering out alternative actions. The paper has verified this method on multiple LIBERO benchmarks, three VLA models, and real robot tasks. The results show that this approach can increase the task success rate and improve the OOD execution performance to a certain extent.

Critical strengths
  1. REVIVE enables online recovery without changing the VLA architecture or requiring offline retraining. It combines phase-aware monitoring, local rollback, Decoder Adaptation, and Failure-Guided Selection. This design can be deployed as an external recovery module for existing VLA policies.

  2. The experiments cover multiple VLA backbones, simulation benchmarks, and real-robot tasks. REVIVE is evaluated with π0, π0.5, and GR00T N1.7 on LIBERO and LIBERO-plus, as well as on long-horizon and OOD real-robot tasks. The results show consistent improvements across models and environments.

  3. The paper compared the effects of direct rollback, Failure-Guided Selection, Decoder Adaptation, and different rollback ranges, and reported the number of execution steps, rollback times, and interaction costs. The results showed that the combination of DA and FGS was superior to using any single component alone, and local rollback could better maintain the existing progress compared to returning to the task start point.

Critical weaknesses
  1. The stage is mainly based on the changes in the gripper and the peak of the action changes in the successful trajectory, and the training target of the Return Controller also relies on these offline boundaries. The paper has not fully explained whether the stage segmentation and the monitor remain reliable when the task stage is unclear, the action changes are not obvious, or the task structure differs significantly from the training distribution.

  2. In real deployments, rollback operations are not allowed in some cases, such as when the object moves or there are physical irreversible situations.

  3. The comparison or contrast description with existing recovery and re-planning methods is still insufficient. The paper mainly compares the basic strategy and its own component variants, lacking a direct comparison with other re-planning methods, or other inference-time intervention methods (World-In-the-Loop, Critic in the Loop, etc.).

  4. It is an additional suggestion whether the method in this paper can be tested in a dedicated LIBERO-RECOVER to more directly demonstrate its performance.

[1] World-in-Loop: Online Correction via Event-Triggered World Models for Robust VLA Policies

[2] Critic in the Loop: A Tri-System VLA Framework for Robust Long-Horizon Manipulation

[3] LIBERO-RECOVER: Beyond Task Success Towards Failure Recovery in Robotic Manipulation Models

Questions for authors

Please refer to Weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer B4ka

We appreciate the positive evaluation of REVIVE's practicality and of our component and cost analysis. Below we address the three weaknesses in the order given.

On stage segmentation and monitor reliability

"The stage is mainly based on the changes in the gripper and the peak of the action changes in the successful trajectory... The paper has not fully explained whether the stage segmentation and the monitor remain reliable when the task stage is unclear, the action changes are not obvious, or the task structure differs significantly from the training distribution."

We agree this was under-justified, and we have added an external validation in Appendix A.2. The key point is that the rule is not fitted to LIBERO: we applied it with default parameters, no benchmark-specific tuning, to two benchmarks that carry phase annotations produced by their own authors.

BenchmarkSegmentation ruleRecall @±20 framesWithin-phase (%)
RoboCasa365gripper only422/838 (50.4%)23.4
RoboCasa365hybrid (ours)728/838 (86.9%)85.0
CALVINgripper only16/17—
CALVINhybrid (ours)17/17—

On RoboCasa365 — a kitchen benchmark with a different action space and stage structure from LIBERO — the hybrid rule locates 86.9% of 838 annotated transitions within ±20 frames, and 85.0% of predicted segments contain no ground-truth boundary inside them. That last figure is the one that matters for rollback: returning to the previous anchor replays actions from the same phase rather than undoing a completed one. The comparison against the gripper-only rule also shows why the action-change criterion is not decorative: many composite phases involve moving or operating a mechanism rather than grasping, and gripper transitions alone recover only half.

One design point may deserve emphasis. Our phases are operational, defined by locally coherent action patterns, not semantic subgoals. This is deliberate, for exactly the reason the reviewer raises: semantic stage labels are often ambiguous in expert demonstrations. Using an operational definition lets the controller be trained without manual annotation, at the cost that its phases need not map one-to-one onto human-named subgoals — acceptable, because the controller only needs to detect that a local attempt has stopped progressing and that a rollback target exists.

The residual scope is real and now stated in the Conclusion: smooth continuous execution offers no boundary, and a stage structure far from the training distribution remains a risk.

On rollback when the scene is physically irreversible

"In real deployments, rollback operations are not allowed in some cases, such as when the object moves or there are physical irreversible situations."

We agree, and we do not claim otherwise. Rollback reverses recent joint commands; it does not reconstruct contacts, object poses, or grasp state. Once an object has moved irreversibly, a contact has broken, or the scene has degraded, returning to the previous anchor cannot restore the original state.

This is precisely why REVIVE intervenes before the failure becomes irreversible. The controller tracks phase-local progress over a sliding window and triggers Return while the deviation is still local, so the previous anchor remains approximately recoverable. We have made this scope explicit in the Conclusion, and we now motivate the preventive framing in the Introduction with LIBERO-Recover's own measurements: with episodes initialized at recorded failure states, GR00T N1.5 recovers only 5.7%–26.3% and OpenVLA-OFT 15.6%–32.6%, far below their ordinary LIBERO success. Preventing a failure from arising is more tractable than repairing one afterwards.

AuthorsOpenReview ↗

Response to Reviewer B4ka

On comparison with existing recovery and re-planning methods

"The paper mainly compares the basic strategy and its own component variants, lacking a direct comparison with other re-planning methods, or other inference-time intervention methods (World-In-the-Loop, Critic in the Loop, etc.)."

We agree, and we now discuss World-in-Loop in Section 2.3 of the revision. The two methods share the central idea: when an event-triggered monitor flags an unsuitable grasp, the agent returns to a previously healthy state. The difference is in the machinery. World-in-Loop obtains the post-rollback behaviour from a video-generation model that synthesizes a corrected future, later converted to actions by an inverse-dynamics decoder. REVIVE needs neither a VLM judge nor a video-generation model: its deviation signal comes from a lightweight auxiliary head over the frozen VLA's own representation (2.39 MB), and its alternatives come from temporary decoder adaptation that repels recorded failed actions, filtered by an episode-local failure-density criterion.

Critic in the Loop addresses a different cost — the latency of placing VLM-level reasoning in the control loop — which it solves by dynamic scheduling so the expensive system is queried sparingly. On trigger, the arm discards planned future actions and re-plans from its current state; it never returns to an earlier execution point, so the question of where the robot should go back to does not arise. A re-plan in place inherits whatever local deviation has accumulated, whereas returning to the last confirmed anchor gives the policy a chance to act differently from a state that is still recoverable.

On testing on LIBERO-Recover

"It is an additional suggestion whether the method in this paper can be tested in a dedicated LIBERO-RECOVER to more directly demonstrate its performance."

We respectfully note a regime mismatch rather than a gap we can close by running the benchmark. LIBERO-Recover initializes every episode directly at a recorded failure state and provides no pre-failure history. It therefore measures recovery from an already-degraded scene. Such a setting supplies neither a phase anchor to return to nor a pre-failure progress trajectory to compare against — the two signals the Return Controller is built on. REVIVE is designed and evaluated as a preventive mechanism that keeps the VLA out of the failure states LIBERO-Recover starts from.

That said, we take the suggestion seriously: a method whose recovery ability is never exercised is less useful than one that has been tested. We agree LIBERO-Recover has reference value for asking whether recovery capability degrades gracefully once the scene is damaged, and we will cite it in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.