acceptodds
Under review as a conference paper at ICLR 2027

GainPilot: Steering Vision-Language-Action Policies with Gain-Triggered Replanning and Bounded Action Correction

Abstract

Action-chunked vision-language-action (VLA) policies can accumulate execution errors, but detecting deviations alone does not determine whether replanning helps. We introduce GainPilot, which combines gain-triggered replanning with bounded post-sampling correction of a base policy. A dynamics-informed trigger predicts the terminal-success gain of replacing the active queue with a subsequently sampled and corrected plan. An anchored critic jointly learns absolute success, anchor-relative gains, and outcome differences between paired local perturbations. One action-gradient computation then supplies a coordinate-aware correction direction, bounded separately for translation and gripper commands. Across 13 backbone–horizon configurations on Meta-World MT50, spanning , SmolVLA, and X-VLA, GainPilot improves success by 10.6–26.0 percentage points over the base policies. For at , success rises from 61.6% to 74.2%, compared with 69.4% for gain-triggered replanning alone. On LIBERO with 10-shot fine-tuning, mean success rises from 79.00% to 90.75%; across nine real-robot tasks, it increases from 53.3% to 73.9%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-language-action models, Gain-triggered replanning, Post-sampling action correction, Anchored critics, Action-chunked control, Robotic manipulation.

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

This paper proposes GainPilot, a framework for improving the execution reliability of VLA policies through gain-triggered replanning and bounded action correction. It uses a gain predictor to estimate whether replanning is likely to improve task success and triggers intervention only when the expected benefit exceeds a threshold. A critic trained with gain-based and pairwise supervision then provides bounded gradient-based corrections to newly generated action chunks. Experiments on Meta-World MT50, LIBERO, and real-world robotic manipulation tasks show improvements over the original VLA policies and several correction baselines, suggesting that GainPilot can improve execution reliability by combining selective replanning with local action refinement.

Critical strengths
  1. The paper focuses on the accumulation of errors and unnecessary re-planning during the execution process of VLA, attempting to determine when intervention would be more helpful rather than constantly re-planning frequently. This helps to enhance the reliability and efficiency of robot execution.
  2. GainPilot not only predicts whether re-planning can bring benefits, but also uses the critic to perform bounded gradient correction on the newly generated actions, making the intervention decision and action optimization complementary to each other.
Critical weaknesses
  1. Prior work such as AAC (https://arxiv.org/pdf/2604.04161) adapts action execution horizons based on action entropy to balance execution efficiency and responsiveness. Although GainPilot additionally uses gain prediction and critic-guided action correction, the paper could more clearly establish the unique contribution of its gain-triggered replanning mechanism over existing adaptive horizon methods.
  2. The comparison on the real robot is not sufficient. The results of the nine tasks are commendable, but each task was only tested 20 times, and it was mainly compared with the original π₀.₅. Therefore, it cannot fully prove that it is superior to other error correction methods.
Questions for authors
  1. The paper uses a fixed threshold Gt>0.05G_t > 0.05 to trigger re-planning. Is this threshold applicable to different tasks and VLA backbones? How significant are the effects of threshold changes on success rate and re-planning frequency?
  2. GainPilot modifies actions based on the gradient of the critic, but the prediction of the critic may not be accurate. How can we ensure that the gradient direction truly improves the success rate of the task, rather than generating unreasonable actions by exploiting the prediction error of the critic?
Reviewer_tQm7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GainPilot is a test time method for frozen action chunking VLA policies that decides when to abandon the current action chunk sequence and then refines the re-queried action chunk.

A gain model using action conditioned visual dynamics predictions estimates the terminal success of switching to a freshly sampled plus critic-corrected plan versus continuing, and then triggers replanning above some threshold. The replacement’s first 10 actions are then shifted by one bounded gradient step from a critic ensemble trained on terminal outcomes with absolute-success, anchor-relative, and paired-perturbation losses.

Training labels for both models come from paired simulator branches that restart from the same point and roll out to task completion. On Meta-World MT50 with pi0.5, SmolVLA, and X-VLA, the paper reports gains of 10.6–26.0% over native execution and 4.8–8.4% over uncorrected replanning; for π0.5 at H=10, success rises from 61.6% to 74.2%, versus 69.8% for VLA-Corrector. It also reports results on 10-shot LIBERO (79.0 to 90.75%) and nine real-robot tasks on an AgileX PiPER (53.3 to 73.9%).

Critical strengths

There is very limited strength to this paper. The only one I could think of would be:

  1. Controlled simulation comparisons: Gains over uncorrected replanning are consistent across three backbones and various action chunk horizon settings. The Replan and Matched Replan variants also separate the contribution of correction from the trigger’s training target.
Critical weaknesses
  1. One core requirement of this method depends on is impractical. Both the gain and critic labels come from paired branches that restart from the same execution point, replay the prefix, and roll out to terminal success. Appendix A.2 implies roughly 1,800 critic rollouts per MT50 task (50 sources × 9 candidates × 4 continuations), and the total collection protocol, method, and cost is never reported for real robot. Even in sim, this method requires additional data collection and training on these additional data, making their gains over baselines and comparable methods questionable.

For the real world PiPER arm experiments, the paper does not describe how real robot data were collected, pi0.5 was fine-tuned, how the auxiliary models were trained, or how branch labels were obtained on hardware. It also omits a hardware comparison to VLA-Corrector, which uses the same platform and a near-identical task suite with demonstration-only supervision and reports a comparable gain.

  1. The real-robot results are unsubstantiated. This paper provides no videos or supplementary material demonstrating tasks specifications, data collection, eval rollouts. So it's impossible to judge its merits. The real world tasks are also trivial in difficulty, does not justify why this method is even needed to solve these tasks in the first place.
Questions for authors

How is real robot data collected for your method?

Reviewer_QEiqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GainPilot, a test-time method for improving frozen action-chunking VLAs. It uses lightweight dynamics and gain models to decide when replanning would be beneficial, then adjusts the newly generated action chunk using a critic ensemble. The auxiliary models are trained offline without modifying the base VLA. Experiments on Meta-World MT50 show improvements of 10.6–26.0 points over native execution and 4.8–8.4 points over replanning without correction. Success also improves from 79.0% to 90.75% on LIBERO and from 53.3% to 73.9% across nine real-robot tasks.

Critical strengths
  1. Deciding when to replan based on expected improvement is an interesting idea, and evaluating the trigger before generating a new chunk avoids unnecessary inference.
  2. The gains are consistent across different policies and settings, with clear comparisons separating the benefits of replanning and correction.
  3. The real-robot results are encouraging, improving from 96 to 133 successes over 180 trials, with gains across all nine tasks.
  4. The paper includes relevant recent baselines, outperforming both VLA-Corrector (69.8%) and GAF (68.0%) with 74.2% success.
  5. The auxiliary models are relatively lightweight (3–10M parameters), and inference latency is reported.
Critical weaknesses
  1. It's unclear how much the gain-based trigger actually contributes, since different trigger targets perform similarly (69.4% vs. 68.8%), and comparisons with random triggers or shorter fixed horizons are missing.
  2. The LIBERO and real-robot experiments lack important setup details and a Replan baseline, making it difficult to separate the gains from replanning and correction.
  3. The method seems to require substantial interaction data, but the total cost is not reported, and possible train/evaluation overlap and hyperparameter tuning on evaluation conditions raise concerns.
  4. The correction always takes a fixed-size step in the critic's gradient direction, but there's no random-direction or adaptive-step baseline to show this is necessary.
  5. Most MT50 results are based on single 500-episode evaluations without confidence intervals, so some of the smaller improvements are hard to judge.
  6. The base policies are relatively weak, and inference latency increases noticeably, particularly for X-VLA (86.9 to 148.6 ms).
  7. Several implementation details are missing, including auxiliary-model training settings and baseline tuning, making the method difficult to reproduce.
Questions for authors

1a. How were the dynamics, gain and critic models trained for LIBERO and for PiPER? Appendix A describes only the π0.5–MT50 configuration. What data were collected on the robot, and how were branches (restart, prefix replay, forking) created on hardware?

1b. For PiPER, which checkpoint does "the same π0.5 checkpoint" (Sec. 4.4) refer to, and with what fine-tuning data and H was it used? How were initial conditions and disturbances randomized, and how were trials ordered? Who judged success, and was the judging blind? Undisclosed robot-side training or train/eval overlap would lower my score.

  1. Can you add the raw-gain trigger with the same correction, a random trigger at the matched intervention rate, native execution at H = 5, and a Replan arm on LIBERO and the robot? Please also report Matched Replan's intervention count, which Table 2 leaves unreported. Results showing that the gain target matters would raise my score.
  2. How many environment steps did data collection take per backbone? Appendix A.3 splits the data by initial scene. Are the 500 evaluation conditions drawn from the test scenes? Fig. 5 compares dynamics widths on the evaluation conditions, and the default (512) is the best one there. How were τ and the correction radii chosen? Can you report results with hyperparameters selected on a held-out split?
  3. How does a random correction direction of the same amplitude perform? How many evaluation episodes per fit does Table 3 use, and is Table 1's 74.2% one of these three fits? [Optional, verify first: the reported means and SDs, e.g. 73.33 ± 1.15, are consistent with per-fit success rates in steps of 2%. That would suggest 50 episodes per fit.]
  4. Can you report held-out predictive metrics for the gain models and the critic on the test scenes of Appendix A.3? For example, the correlation between predicted and realized gain, or pairwise ranking accuracy.
  5. Can you report GainPilot on the full-fine-tune LIBERO checkpoint (97.0)? This would show whether it preserves or lowers success when the base policy leaves little headroom. A converged π0.5 MT50 checkpoint would also help, if feasible.
  6. Can you give paired tests (e.g., McNemar over the shared initial conditions) for GainPilot vs. Replan and vs. the Table 2 baselines? How were VLA-Corrector and GAF implemented (official code or re-implementation), and how were they tuned? What learning rates, update counts and training time were used for the auxiliary models?

Comments on the paper

1 comment

Dear Reviewers,

I would like to kindly remind everyone to review the authors' responses and provide your feedback.

If you have any further questions or concerns regarding the authors' responses, please feel free to raise them early, so that the authors have sufficient time to address them before the rebuttal period ends.

Thank you all very much for your time and effort! And, have a nice weekend. :)

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.