acceptodds
Under review as a conference paper at ICLR 2027

Vision-Grounded Energy Critics for Physically Feasible Vision-and-Language Action Policies

Abstract

We introduce Vision-Dominated Energy Learning (VEL), a lightweight, backbone-agnostic framework that augments any frozen or fine-tuned vision-and-language action (VLA) policy with a compact energy head. The energy head maps the hidden representations of the VLA and a candidate action to a scalar energy value, where lower energy indicates higher compatibility between the action and the current scene. Training uses a contrastive objective on demonstration data, and inference applies a single gradient-based residual correction that shifts the proposed action toward a lower-energy alternative. This procedure requires no environment rollouts, no additional data collection, and no modification of the VLA backbone. At the population optimum, the contrastive objective recovers the log density ratio between the conditional expert action distribution and the action marginal, and a single gradient step is guaranteed to reduce the learned energy under local smoothness conditions. Over five training seeds of the energy head, VEL raises the LIBERO-Long success rate of OpenVLA-OFT from 94.4% to 95.8% and that of from 92.4% to 93.6%, which removes 24% and 16% of the residual failures, while the shorter suites show smaller changes of mixed sign. Controlled comparisons show that search over the same energy, a direct correction head of matched capacity, and a score-matching energy head do not reproduce the long-horizon gain. On a physical screw-insertion task with a backbone, VEL raises full-task success from 35% to 50%, and the gain concentrates in the precision-demanding insertion stage. The energy head introduces fewer than 4.0M additional parameters and 4.61 GFLOPs per correction step, an increase of 0.4% over OpenVLA-OFT and 5.2% over .

Primary area:
applications to robotics, autonomy, planning
Keywords:
Energy-based Model, Robotic Policy, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.80 over 5 reviews, above 27% of the others.
Reviewer_9wEJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VEL. Keeping the VLA backbone frozen, VEL trains a small energy head using existing demonstration data, and corrects actions using a single gradient descent step. The authors evaluate VEL on LIBERO using OpenVLA-OFT and π0.5\pi_{0.5} as backbones. They also report improved performance with VEL on a real-world screw grasping and insertion task.

Critical strengths
  1. The method provides a lightweight approach to adapting existing VLA policies. VEL keeps the VLA backbone frozen and trains the energy head using existing demonstration data, without additional data collection.

  2. The method is evaluated on two VLA backbones in simulation and is also tested on a real-world manipulation task.

Critical weaknesses

My main concern is the consistency of the method's performance gains across task suites. As shown in Table 1, VEL does not consistently improve on the corresponding base policies across all task suites. Specifically, on LIBERO-Object, both VLA backbones have slightly lower reported mean success rates with VEL than without it. It therefore remains unclear how to determine in advance whether applying VEL will improve performance in a given deployment setting. A clearer characterization of when VEL is likely to help or hurt would make it easier to decide whether to use it in deployment.

Questions for authors
  1. Table 1 reports results over five energy-head training seeds for VEL, whereas the other rows are described as single-run results. How sensitive are the baseline success rates to evaluation randomness?

  2. In Figure 5, why does the energy appear to vary mainly with the noise standard deviation and remain nearly constant along the noise-mean axis? Does the model still assign low energy to perturbed actions when the noise standard deviation is small but the absolute noise mean is large?

  3. In Appendix E.5, the results suggest that fewer correction steps and smaller step sizes yield higher average success rates over the tested settings. What happens if α\alpha is reduced below 0.1?

Reviewer_GpNWOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Vision-Dominated Energy Learning (VEL), which uses a contrastively trained energy head to refine VLA actions through a single gradient step. Experiments on LIBERO and a real-world screw-insertion task report improvements over the baseline policies.

Critical strengths
  1. The method is lightweight and can be integrated into existing VLA backbones without additional environment interactions.
  2. The paper evaluates two VLA architectures and includes ablation studies and a real-robot experiment.
Critical weaknesses
  1. The method combines established techniques, including energy-based modeling, contrastive learning, and gradient-based action refinement. The theoretical analysis largely follows existing results, and the technical contribution beyond prior work remains limited.
  2. The simulation and real-world experiments use substantially different robotic platforms (a parallel gripper versus a dexterous hand), making the results difficult to compare directly. The real-world evaluation is limited to a single screw-insertion task, with only a few images and success-rate statistics presented. Without complete demonstration videos, detailed failure analysis, or quantitative measurements of action corrections, the reported improvements are difficult to assess. Moreover, the modest gains on LIBERO do not sufficiently support the broader claims about physical feasibility.
Questions for authors
  1. Can the authors provide complete real-robot demonstration videos, including both successful and failed trials, to substantiate the reported results?
  2. Can the authors quantify how the energy-based correction improves physical feasibility, rather than simply increasing the likelihood of expert-like actions?
Reviewer_JsMLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Vision-Dominated Energy Learning (VEL), a method for refining the actions of a frozen VLA policy. The main component is a small energy-based critic network that takes the VLA’s last hidden layer features and a candidate action and outputs a scalar energy. The critic is trained on the demonstrations with a contrastive objective (InfoNCE), so that the expert action receives lower energy than expert actions from other states in the batch. At inference, the VLA’s action is refined by a single gradient step on the learned energy function. The method is applied on top of OpenVLA-OFT and π0.5\pi_{0.5} on the LIBERO simulation benchmark. It improves LIBERO-Long success from 94.4% to 95.8% for OpenVLA-OFT and from 92.4% to 93.6% for π0.5\pi_{0.5}, with small changes of mixed sign on the other three suites. On a real-robot screw-insertion task with π0.5\pi_{0.5} full-task success rises from 35% to 50%. The authors also show that the contrastive optimum recovers a log density ratio, and they give a one-step descent bound.

Critical strengths

The introduced method is a simple, lightweight extension that refines the actions of a VLA at inference time, without the need to retrain the backbone. The real-robot results are promising, as VEL substantially improves π0.5\pi_{0.5} on the precision-demanding insertion task. The paper is clearly written and easy to follow. The controls in Table 2, such as search over the learned energy, are informative, even though they expose limitations of the method.

Critical weaknesses
  • The main weakness of the method is the limited size of its gains. The main evaluation is on LIBERO, which consists of four task suites, and the only improvements reported as significant are on LIBERO-Long: +1.4 pp on top of OpenVLA-OFT and +1.2 pp on top of π0.5\pi_{0.5}. These gains are modest, and the changes on the remaining three suites are small and of mixed sign. Moreover, LIBERO is largely saturated (the base policies already reach 92–99%), so it remains unclear how the method performs on more challenging benchmarks. The real-robot results are promising but not sufficient on their own, as they cover a single task with a single energy head and no other baselines. Finally, the related work (Appendix A) claims that VEL is complementary to RL fine-tuning and can be applied on top of an RL-fine-tuned policy, but this claim is not tested, so the benefit of VEL for such policies remains unclear.

  • The learned energy does not seem to reflect action quality reliably. Table 2 shows that selecting lower-energy actions by search reduces the success rate below that of the base policy, even for the anchored CEM, which only searches a small neighborhood of the base action. The same pattern appears in Table 8, where additional gradient steps and larger step sizes steadily reduce the success rate on LIBERO-Long. These results raise concerns about the robustness of the method, as the step size must be carefully tuned to stay within the region where the learned energy is reliable.

  • The motivating claim in the introduction that the standard VLA pipeline has a "structural gap", namely "the lack of a mechanism that evaluates a candidate action against the current scene configuration and revises it before execution", seems unjustified. The VEL critic is trained on the same expert demonstrations as the behavior-cloned VLA and is conditioned on the VLA's own hidden states, so it has access to no information that the policy lacks. Any benefit must therefore come from a difference in estimation, which the paper does not identify.

Questions for authors

Could you evaluate VEL on a less saturated benchmark where the base policies have more room for improvement?

Reviewer_KyLvOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Vision-Dominated Energy Learning (VEL), a lightweight action-refinement framework for VLA policies. VEL attaches an energy critic to a pretrained VLA backbone and trains the critic using an in batch contrastive objective over expert state action pairs. The method performs a single gradient-based update to refine actions proposed by the base policy, without retraining the backbone or requiring additional environment rollouts. Experiments on LIBERO with OpenVLA-OFT and π0.5 show improvements primarily on LIBERO-Long.

Critical strengths
  1. This paper focus on an limitation of VLA that absence of an explicit mechanism to evaluate and refine proposed actions before execution. The proposed energy critic can be integrated into existing policies without expensive online interaction.

  2. The experiment analysised alternative refinement mechanisms. The density-ratio interpretation and local descent analysis clarify what the learned energy represents and the limited guarantees associated with its optimization.

Critical weaknesses
  1. The contrastive objective learns to discriminate matched expert state action pairs from mismatched pairs. But it does not directly establish physical feasibility. The method does not explicitly supervise collision avoidance, contact stability, dynamical constraints, or task-precondition satisfaction. A low energy action is therefore not necessarily physically feasible.

  2. The main simulated improvements are concentrated on LIBERO-Long, with gains of approximately 1.2–1.4 percentage points. Results on the shorter suites are small and occasionally negative. It should be evaluated through comparative experiments across a broader range of benchmarks rather than being limited to a single benchmarkand hardware configuration.

  3. I inspected the supplementary code and found a discrepancy that require clarification. The energy head training code uses ground-truth actions from a normalized training pipeline. However, the inference implementation appears to pass actions that have already been unnormalized by the VLA into the energy correction function. I could not identify the corresponding normalization step before the critic evaluation. If these paths were used for the reported experiments, the critic would receive actions in different coordinate scales during training and inference.

Questions for authors
  1. Can the authors provide evidence, preferably using existing evaluation trajectories, that the learned energy distinguishes physically feasible and infeasible actions rather than merely ranking actions by similarity to expert demonstrations?

  2. For the real-robot experiment, could author clarify the insertion-stage success denominator and whether the baseline and VEL trials use matched initial conditions?

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

This paper proposes Vision-Dominated Energy Learning (VEL), a lightweight energy-based action refinement framework for vision-language-action (VLA) policies. The method introduces a compact energy critic trained with an InfoNCE objective on expert demonstrations. During inference, the critic evaluates the compatibility between visual-language representations and candidate actions, and a single gradient-based residual update is applied to refine the actions generated by the original VLA policy. The authors provide theoretical analyses of the contrastive energy objective and local energy descent, and evaluate the proposed approach using OpenVLA-OFT and π0.5 on the LIBERO benchmark. Real-robot experiments on a screw-insertion task are also presented. The idea of improving pretrained VLA policies without modifying their backbones is practically appealing. However, I have several concerns regarding the novelty of the approach, the relationship between the learned energy and physical feasibility, the robustness of the experimental evidence, and the consistency between the described method and the supplementary implementation.

Critical strengths
  1. Practical motivation and lightweight integration. The paper addresses a relevant problem in robotic manipulation: correcting potentially suboptimal actions generated by pretrained VLA policies. The proposed framework is relatively simple and introduces limited additional model parameters and computation.
  2. Evaluation across different VLA architectures. The experiments cover both OpenVLA-OFT and π0.5, providing initial evidence that the proposed energy-based correction mechanism is compatible with different action-generation architectures.
  3. Real-robot experiments and refinement baselines. The real-robot screw-insertion experiment provides encouraging results, with full-task success increasing from 35% to 50%. The comparisons with direct correction, score matching, and energy-based search are also useful for investigating the proposed refinement mechanism.
Critical weaknesses
  1. Limited methodological novelty and theoretical justification. The proposed framework mainly combines existing techniques, including contrastive energy learning, cross-attention, and gradient-based action refinement. Although the integration is practical, the theoretical propositions largely follow established results and do not demonstrate that minimizing the learned energy necessarily improves action quality or physical feasibility.
  2. Limited experimental improvements and insufficient validation. The improvements on LIBERO-Long are relatively modest, with gains of only 1.4 and 1.2 percentage points for OpenVLA-OFT and π0.5, respectively, while other suites show mixed results. Furthermore, the real-robot evaluation is limited to a single insertion task. Additional evidence, particularly regarding collision avoidance, action correction effectiveness, and statistical reliability, is needed to support the generality of the claimed benefits.
  3. Insufficient ablation analysis and reproducibility concerns. The ablation studies do not fully establish why energy-based correction outperforms alternative approaches. In particular, the substantial performance degradation of the direct correction baseline raises concerns about comparison fairness. Additionally, several discrepancies between the supplementary code and the described methodology, including correction step sizes and gradient clipping strategies, require clarification to ensure reproducibility.
Questions for authors
  1. Can the authors provide quantitative evidence that the learned energy gradient corresponds to physically meaningful action correction, rather than merely increasing compatibility with the demonstration distribution? In particular, how often does the gradient correction reduce action error or physical constraint violations?
  2. Can the authors clarify the statistical evaluation protocol, including whether baseline and corrected policies use identical initial states? Could they report confidence intervals and paired rescue/failure statistics for the main LIBERO-Long results?
  3. How was the direct residual-correction baseline optimized, and were its correction magnitudes constrained to match those of VEL? Could the authors explain its substantial degradation relative to the original policy?
  4. Can the authors reconcile the supplementary implementation with the method described in Section 3.5, particularly regarding gradient clipping, action projection, step-size defaults, and energy-output ranges?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.