acceptodds
Withdrawn from ICLR 2027

ActionCritic: Action-Aware Self-Criticism for Reliable Vision-Language-Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-Language-Action (VLA) models provide a scalable interface for mapping multimodal observations and language instructions to robot actions. However, most VLA policies ex- ecute predicted actions without explicitly assessing whether those actions remain appropriate under the current scene con- figuration, where a locally incorrect command can cause an irreversible task failure. In this paper, we introduce Ac- tionCritic, a unified VLA framework that internalizes ac- tion verification, corrective instruction generation, and dy- namic re-planning within a single shared multimodal repre- sentation. Rather than relying on computationally expensive external simulators or disjointed auxiliary correction mod- ules, ActionCritic proactively evaluates the physical suitabil- ity of proposed event-triggered key actions against the ob- servation history and task semantics prior to execution. Un- suitable action proposals are preempted and translated into concise atomic corrections, which directly condition subse- quent action generation to enable bounded, local recovery. Training follows a two-stage schedule that first establishes critic capability with QA supervision and then mixes plan- ning and correction-conditioned re-planning while retaining a low-weight QA objective. Furthermore, ActionCritic learns a robust, deployment-aware decision boundary by compre- hensively leveraging both successful and failed trajectories while rigorously excluding failed motor commands from im- itation targets. Extensive evaluations across the SIMPLER and RoboTwin benchmarks demonstrate that ActionCritic sig- nificantly improves closed-loop task success, critic discrim- ination, and recovery effectiveness while maintaining strict inference efficiency.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotics

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_h96BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ActionCritic, which adds a pre-execution self-check mechanism to StarVLA. Before grasping or releasing an object, the model re-observes the scene and judges whether the action is appropriate; if not, it generates a language correction and replans. Training uses four types of QA tasks to learn action understanding and error correction, with failed trajectories serving as negative samples. Experiments on SIMPLER, LIBERO, LIBERO-Plus, and real-world ALOHA tasks show gains of roughly 1–8 percentage points.

Critical strengths
  1. Checking before critical actions such as grasping and releasing prevents hard-to-recover errors early. Triggering only at key moments also avoids unnecessary computation.
  2. With the same backbone, data, and action representation as StarVLA, the method improves performance across model scales, simulation benchmarks, and real-robot tasks.
  3. Failed trajectories are used only to learn error recognition, not to imitate erroneous actions. Using a single model for action generation, verification, and correction also simplifies the architecture.
Critical weaknesses
  1. VLA-in-the-Loop already studies verification and correction before critical actions. Action Draft and Verify uses the VLA's internal capability to verify actions. VLAC unifies policy and critic in a single model. OneTwoVLA, RACER, and others also address self-correction and recovery. The paper should clearly explain its differences from these methods and compare against them empirically.
  2. There are missing key controls; the revierer is unclear why the method works. The results only show that adding ActionCritic improves performance. They do not answer three questions. Would simply letting the VLA re-observe and replan before critical actions work equally well? Is a shared critic better than a separate one? Is language correction actually necessary? Table 5 also omits the most important setting: zero corrections. These controls bear directly on the core contribution. Moreover, unlimited corrections yield the highest success rate in Table 5, suggesting that simply retrying more contributes substantially.
  3. The method uses extra QA supervision, failed trajectories, and VLM-generated correction labels. The current experiments cannot separate gains due to the self-check mechanism from gains due to the additional supervision.
  4. The paper claims that sharing one model reduces overhead, but it does not clearly report critic inference time or compare against a separate critic. The current latency results are insufficient to support the efficiency claims.
Questions for authors

Please address my comments above, especially:

  • Would simply letting the VLA re-observe and replan before critical actions work equally well?
  • Is a shared critic better than a separate one?
  • Is language correction actually necessary?
Reviewer_UxrDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

ActionCritic is a VLA policy that verifies its own key actions before execution. A hand-written trigger detects the first gripper-close in a predicted chunk, and the same backbone judges whether it is suitable. If rejected, it generates a short language correction and replans. Training is two-stage: QA-only critic training, then joint planning, replanning, and QA. The policy is building on top of StarVLA, and evaluated on SIMPLER, LIBERO, LIBERO-Plus, and ALOHA.

Critical strengths
  1. Verifying irreversible actions like grasps before execution is well motivated.
  2. A single shared model with no external corrector or video prediction is simple and low-overhead.
  3. Atomic language corrections are interpretable, and using failures only as critic negatives is sensible.
  4. Gains are consistent across scales, simulators, and a real robot, and the limitations section is honest.
Critical weaknesses
  1. Critic is a proxy, not a physical check. It outputs a text yes/no learned from demo states vs. earlier frames. The abstract claims improved “critic discrimination,” but no critic accuracy or precision/recall is reported.
  2. Correlated failures. Policy and critic share a backbone, and LIBERO-Plus camera (61.0 to 61.6) and robot (40.1 to 41.0) barely improve. The OOD scenarios for VLA, would be OOD for the verifier too. No separate-critic comparison.
  3. Failure data source unspecified, especially for SIMPLER/Bridge, where demos are all successes. Base-policy rollouts in the eval simulator would raise leakage concerns and give the baseline less data.
  4. Missing ablations. No separation of QA co-training without the loop, failure data vs. none, trigger vs. always-verify, or two-stage vs. single-stage, so gains may come from extra supervision rather than verify-and-replan.
  5. Unclear details. Only the grasp trigger is defined. The action head and key-action serialization are unstated, so compatibility with different action representations is not shown. It is also unclear how the 0.1 QA weight balances against action losses.
  6. Inconsistent baseline coverage. The LIBERO and LIBERO-Plus tables compare different sets of models. OpenVLA-OFT, π0.5, UniVLA, NORA, RIPT-VLA, and WorldVLA appear only on LIBERO-Plus, while CogACT, CoT-VLA, CronusVLA, TriVLA, 4D-VLA, and SpatialVLA appear only on LIBERO. Strong baselines like OpenVLA-OFT and π0.5 are therefore absent from the main LIBERO comparison, and CogACT is absent from LIBERO-Plus. LIBERO-Plus also reports only the 2B model. Beyond that, all baselines are vanilla VLAs, with no reflection/recovery methods (e.g., Phoenix, RACER, AHA) or external verifier.
  7. Weak statistics. No seeds or confidence intervals, a small LIBERO gain (+1.4) on a near-saturated benchmark, and SIMPLER-Bridge +12.5 gains are about 3 episodes.
  8. Numerical and citation Errors. The LIBERO 0.8B suites average 96.25, not 96.4. SpatialVLA’s LIBERO average does not match its suites. “Figure 6” should be Table 6, and \citet is used where \citep is needed.
Questions for authors
  1. Where do the failure trajectories come from, and how are they labeled?
  2. What is the critic’s accuracy on held-out data, including under shift?
  3. How does a separate critic compare with the shared one?
  4. Can you ablate QA co-training without the loop, and trigger vs. always-verify?
Reviewer_8ZQxOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces ActionCritic, a method in which a VLA policy evaluates the suitability of an irreversible key action, such as grasp or release, by itself before execution. If the proposed action is unsuitable, the policy generates a concise language correction and performs replanning conditioned on it.

Critical strengths
  • Simple architecture. The method performs verification, correction, and replanning within a single model, without relying on external models or a recovery policy.
  • The use of failure data is clear. Failure cases are used only as negatives for the critic, and their motor commands are not imitated. Correction labels are grounded in later successful interaction states, and only those that agree with the recorded end-effector motion are retained.
  • Clear writing. The paper is written intuitively and is easy to understand.
Critical weaknesses
  • It is difficult to identify precisely where the performance gains come from. The paper claims that the performance benefit comes from event-triggered verification and correction, but only the maximum number of correction attempts is ablated. The method introduces additional training, QA co-training, the replanning target, and the critic and correction loop all at once, so a rigorous ablation of each component is needed.
  • Missing comparison with baselines. Various works address correction and verification for VLAs [1, 2, 3, 4]. The benefit of the shared parameter design can only be verified through a comparison with a separately trained critic, with all other variables controlled.
  • Missing quantitative results on critic discrimination and correction behavior. The paper claims to evaluate critic discrimination and correction behavior. However, it does not report the precision of the critic, the trigger and reject frequency per episode, or the success rate after correction.
  • Missing Appendix and real-world robot videos. The main text defers several explanations to the Appendix, but the Appendix is not provided. In addition, the authors must include videos of the real-robot experiments.

[1] FPC-VLA: A Vision-Language-Action Framework with a Supervisor for Failure Prediction and Correction (Yang et al., arXiv 2025)

[2] OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning (Lin et al., ICLR 2026)

[3] Hume: Introducing System-2 Thinking in Visual-Language-Action Model (Song et al., arXiv 2025)

[4] ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning (Huang et al., NeurIPS 2025)

Questions for authors

See above. The authors must add the Appendix and provide real-robot videos in a manner permitted by the ICLR guidelines.

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

Related papers

Discussion (0)

Sign in to comment.