acceptodds
Withdrawn from ICLR 2027

WAVE: Learning to Verify Actions in 3D World Space for Generative Robot Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative robot policies, including vision-language-action models (VLAs) and world-action models (WAMs), produce temporally extended action chunks from observations and instructions. At deployment, native sampling exposes multiple candidate chunks, yet standard control commits to one without comparing how the candidates would change the scene. We introduce WAVE (World-space Action VErification), an overall policy enhancement for generative robot policies. WAVE reconstructs the current scene once as a dense Gaussian field, compresses it into a compact representation, and predicts a separate compact future rollout for every candidate chunk. A unified score fuses the policy's native consistency score with learned verifier scores and structured verifier adjustments in the shared world-space frame. Across three simulation benchmarks and nine matched frozen-policy comparisons, WAVE reaches 99.0% on LIBERO, 79.5% on RoboCasa, and 95.6% average success on RoboTwin 2.0. These results show that consequence comparison converts native candidate diversity into reliable manipulation decisions.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generative robot policies, World models, Vision-language-action models, Action verification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_TFGuOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes WAVE, a framework that enhances frozen generative robot policies through action candidate verification. It reconstructs a shared 3D Gaussian scene, predicts action-conditioned future states, and selects actions by combining native policy consistency with learned and structured verification scores. Experiments across LIBERO, RoboCasa, and RoboTwin 2.0 demonstrate consistent performance improvements.

Critical strengths
  1. The proposed framework provides an intuitive approach to improving generative robot policies through 3D consequence-aware action selection.
Critical weaknesses
  1. The paper provides very limited qualitative results. The only relevant visualizations are the illustrative figures in Figures 1 and 2, with the images appearing to be AI-generated, yet their provenance is not clarified in the AI Use Statement. No actual simulation rollouts, success/failure cases, or visualizations of the predicted 3D Gaussian states before and after actions are provided. Together with the absence of real-robot experiments, this makes it difficult to assess the practical reliability of the proposed method.

  2. There is a potential distribution mismatch between training and inference. The verifier is trained to distinguish expert actions from artificially perturbed negative samples, whereas inference requires ranking multiple plausible actions generated by the pretrained policy itself. Therefore, the preferences learned from synthetic perturbations may not generalize well to policy-generated action candidates, raising concerns about the reliability of the verifier during inference.

Questions for authors
  1. Could the authors provide actual simulation rollouts, including representative success and failure cases, as well as visualizations of the reconstructed 3D Gaussian states and their predicted changes under different candidate actions?

  2. How do the authors ensure that the verifier's ranking preferences learned from artificially perturbed expert actions generalize to plausible action candidates generated by the pretrained policy itself? Is there any evidence that the verifier can reliably distinguish between these policy-generated candidates rather than merely identifying synthetic perturbations?

  3. How accurately do the predicted 3D futures reflect actual execution outcomes? Could the authors evaluate whether the predicted action rankings agree with the true outcomes in simulation?

  4. How much of WAVE's improvement comes from consequence-based verification rather than increased candidate diversity? Could the authors include a random-selection baseline with the same K=4 candidates?

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

This paper introduces WAVE (World-space Action VErification), an inference-time enhancement for generative robot policies. Rather than executing a single sampled action chunk, WAVE samples multiple candidates from a frozen policy and ranks them according to their predicted consequences. It reconstructs a shared current scene as a dense 3D Gaussian field, compresses it into compact Gaussian slots, and predicts an action-conditioned future trajectory for each candidate. The final selection score combines native sampler consistency, learned verification scores, and structured adjustments covering goal progress, spatial consistency, contact, and plausibility. The method is evaluated on LIBERO, RoboCasa, and RoboTwin 2.0 across nine policy–benchmark pairings. All pairings improve over single-sample native deployment, with the best configurations reaching 99.0% success on LIBERO, 79.5% on RoboCasa, and 95.6% average success on RoboTwin 2.0. The paper also studies selection rules, candidate budgets, future representations, compact-state capacity, and training objectives. The central contribution is the integration of shared-state 3D consequence prediction with candidate selection, without directly optimizing the base policy.

Critical strengths
  1. Reconstructing the current scene once and branching only compact action-conditioned futures is a well-motivated architectural choice. It gives all candidates a common coordinate frame and avoids repeating dense reconstruction for each branch. This makes the proposed system more compelling than simply attaching an independent future predictor to every sampled action.

  2. The evaluation spans three benchmarks, multiple policy families, and both single-arm and bimanual manipulation settings. Improvements across all nine pairings support the usefulness of the overall enhancement. Importantly, Table 2 compares several scoring rules at the same candidate count, while Tables 7 and 8 examine compact capacity and training supervision. These controls provide evidence beyond the main comparison between one-sample deployment and four-candidate selection.

Critical weaknesses
  1. Does Table 6 evaluate expert-conditioned futures only, or also futures conditioned on native policy samples? Can you report prediction error and selection quality on held-out candidate banks whose outcomes are measured through simulator execution? In particular, how often does WAVE select a successful candidate when one exists in the bank?

  2. Does the GaussianDream world-query segment described in Appendix B.3 enter the action expert’s prefix attention, or is it used exclusively in a separate scene-context computation? With identical observations and initial noise, are the candidate actions identical between native deployment and WAVE before selection?

  3. How is the straight-line midpoint evaluation in Equation 16 obtained from the recorded sampler trajectory without another vector-field call? If an additional call is required, is its cost included in every runtime measurement?

  4. What action-to-workspace conversion is used for each embodiment? How much performance is retained when the centroid-based action-path terms are removed, or when they are replaced with forward-kinematic end-effector trajectories? This would help distinguish the contribution of calibrated geometry from that of action-coordinate heuristics.

  5. Can you provide uncertainty estimates for the main success rates, variability across independently trained WAVE models, and absolute decision latency on specified hardware? Please also clarify the execution-prefix lengths and whether all backbone passes and native-score evaluations are included.

Questions for authors

See the weakness

Reviewer_yyQ4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WAVE uses best-of-N action selection to improve frozen generative robot policies. It reconstructs the current scene as a Gaussian field and predicts a compact 3D future for each candidate action chunk. It then ranks candidates using policy consistency, learned verifier scores, and geometric criteria. The method improves performance across nine policy–benchmark pairings on LIBERO, RoboCasa, and RoboTwin 2.0.

Critical strengths
  1. WAVE improves existing policies while keeping their action generators frozen. Reusing one scene reconstruction across candidates also helps reduce the cost of comparing predicted outcomes.

  2. The method consistently improves performance across several policies and benchmarks, with substantial gains in some settings, including +12.4 percentage points on RoboCasa. The ablations help establish the value of future prediction and the different scoring components.

Critical weaknesses
  1. All experiments are in simulation. Although the benchmarks include cluttered and randomized scenes, it remains unclear how well the compact Gaussian representation handles real-world clutter, occlusion, and reconstruction errors. This leaves an important gap in the evidence for practical use.

  2. The verifier itself is not evaluated directly. The results show that selection helps overall, but do not explain when it chooses the wrong candidate or whether failures come from limited action diversity, inaccurate future predictions, or incorrect rankings. This is particularly relevant because future prediction is supervised on demonstrated actions, while inference involves policy-generated alternatives.

Questions for authors
  1. How accurately does the verifier rank held-out, policy-generated candidates based on their actual outcomes? How much improvement would an oracle selector achieve with the same candidate bank?

  2. Since only demonstrated-action futures receive reconstruction supervision, how well does the model predict outcomes for alternative actions, especially unsuccessful ones? Could you show predicted versus actual futures for both successful selections and failures?

  3. Table 7 studies slot capacity, but how does performance change as distractor count and occlusion increase? Is the default of 64 slots sufficient for more complex scenes?

  4. How are the prediction offsets aligned with action duration and the executed chunk prefix across benchmarks? How sensitive is performance to these choices?

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

Related papers

Discussion (0)

Sign in to comment.