acceptodds
Withdrawn from ICLR 2027

EDGE: RESIDUAL-CONDITIONED CORRECTION ACQUISITION FOR VLA POST-TRAINING

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A failed robot trajectory is not automatically a useful training example. Under a fixed correction budget, the relevant question is which logged states deserve new expert actions. We study this offline acquisition problem for vision-language-action policies and introduce EDGE (Experience Discovery via Gap-aware Evaluation). A nominally trained, frozen latent world-action ensemble supplies three ranking signals: action-consequence surprise, the local mass of progress-improving action alternatives, and a penalty for uncertain residual transport. The observed transition residual anchors local counterfactuals without fitting the verifier to the target pool. With rollout pool, query count, replay mixture, and policy-update schedule held fixed, EDGE reaches 58.9±0.7% on LIBERO-PRO and 56.2±0.7% on PhysicalShift. A development-frozen G-bandpass center remains about 3pp behind on ten new paired validation seeds. More importantly, a same-snapshot PhysicalShift rollback study holds the learned WAM, verifier, 16 candidate actions, G, and budget fixed while changing only residual transport. Constant anchoring reduces soft-T error and raises downstream success from 52.0 to 55.5%; kernel decay further improves ranking and reaches 56.2% overall, with gains concentrated in friction, mass, damping, and gain shifts and no gain under delay. This factor-wise pattern connects the local transport assumption to where it succeeds and fails. Five-seed mechanism inference remains limited, but the empirical chain now spans counterfactual prediction, acquisition ranking, and equal-budget post-training.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, active learning, data acquisition, imitation learning, robot post-training, residual transport, world models

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

A failed robot trajectory is not automatically a useful training example. Under a fixed correction budget, the key question is which logged states deserve new expert actions. This paper studies the offline acquisition problem for vision-language-action policies and introduces EDGE (Experience Discovery via Gap-aware Evaluation). A nominally trained, frozen latent world-action ensemble supplies three ranking signals: action-consequence surprise, a teachability score that measures the local mass of progress-improving action alternatives, and a confidence weight that discounts uncertain residual transport. The observed transition residual anchors local counterfactuals without refitting the nominal models to the target pool.

Critical strengths

This paper studies which failed states should receive expert corrections under a fixed correction budget for post-training vision-language-action models. The problem is important, and the experimental design has some controlled elements.

Critical weaknesses

The current evidence is not strong enough to support a robust advantage over strong baselines, and the method’s novelty and generalization remain limited.

  1. The statistical evidence for the main comparison is not strong enough. The main experiments use only five random seeds, so statistical power is limited. On the PhysicalShift benchmark, although EDGE outperforms G-bandpass in all five paired runs, the advantage is not significant after multiple-comparison correction, and the smallest possible two-sided result is still around 0.06. In other words, the current advantage may be driven by seed variation. The authors should add more independent seeds, tasks, or stratified resampling, and they should fix the main comparator in advance rather than choosing the strongest baseline after seeing the results.

  2. Method novelty and theoretical guarantees are limited. EDGE mainly combines residual anchoring, distance decay, ensemble uncertainty, and a progress verifier. Similar ideas already exist in prior work. The authors also admit that the confidence weight is not a strict guarantee, that it cannot handle shared bias, and that the local smoothness assumption fails under action delay. The paper needs a clearer explanation of its irreplaceable contribution relative to prior work, and it needs either clearer theoretical conditions or stronger negative controls.

  3. Hyperparameter tuning and comparison fairness are not transparent enough. The center of G-bandpass is tuned on development tasks and then frozen, but it is not fully clear whether EDGE’s several key hyperparameters are determined on the same development set, with the same budget and the same amount of tuning. If EDGE uses more validation information or more tuning opportunities, the main table advantage may be overestimated. In addition, matching budget by the number of action chunks ignores different action horizons and expert time costs, and cross-backbone comparisons also need more caution.

  4. Generalization and deployment value remain limited. All experiments are concentrated in the same simulator family, and there is no real-robot or cross-simulator validation. The PhysicalShift benchmark mainly uses hand-designed single-factor perturbations, and the set of compound conditions is small. Kernel decay does not help under action delay, which suggests that the so-called support edge may depend heavily on smooth dynamics assumptions. Therefore, the current conclusions should not be directly generalized into a universal acquisition ru

Questions for authors

Q1. The statistical evidence for the main comparison is not strong enough. A1. We agree that five seeds limit statistical power, and we will not overstate significance. The key evidence is a consistent paired advantage: on PhysicalShift, EDGE is about 3.0 percentage points above the frozen G-bandpass in all five paired seeds, with a 95% t interval of roughly 1.6 to 4.4 points. Under the conservative 24-comparison Bonferroni adjustment, the result is not significant at the 0.05 level, and the exact sign-flip sensitivity is also limited by the small number of seeds. In revision, we will (i) pre-specify G-bandpass as the primary comparator, (ii) report effect sizes and seed-level results more prominently, (iii) add seeds if compute permits, and (iv) soften claims that require a multiplicity-robust win on both benchmarks.

Q2. Method novelty and theoretical guarantees are limited. A2. We agree that the individual components are not new. Our contribution is the formulation and controlled study of corrective-label acquisition under matched supervision and policy-update budgets, together with the same-snapshot rollback test that isolates residual transport. We will clarify the distinction from DAgger-style gating, ThriftyDAgger, DataMIL, and CheckVLA. We will also state the theoretical limits more directly: the local bound is conditional, the confidence weight is a heuristic rather than a certificate, and kernel transport fails under action delay. The negative controls in Appendix Q already illustrate this; we will make them more prominent in the main text.

Q3. Hyperparameter tuning and comparison fairness are not transparent enough. A3. All selectors share the same pool, query budget, temporal exclusion window, replay mixture, and update schedule. The G-bandpass center is tuned on disjoint development tasks and frozen before validation. EDGE hyperparameters are fixed from nominal calibration and are not refit to the target pool. In revision, we will add a table listing tuning data, search ranges, and fixed values for EDGE and all baselines. We will also clarify that matching by action-chunk count is a within-backbone budget and does not equate expert time across different action horizons.

Q4. Generalization and deployment value remain limited. A4. We acknowledge that all main results are within one simulator family. We will soften universal claims and add cross-simulator and real-robot validation as explicit future work. We already include SmolVLA, leave-one-factor-out acquisition, compound shifts, independent pools, and imperfect correction channels as robustness checks; we will present them as evidence of scope rather than as proof of broad generalization.

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

The paper proposes EDGE, which states to send for expert correction when post-training a VLA under a fixed correction budget. A frozen VLA collects a pool of rollouts in LIBERO-PRO and in PhysicalShift, a dynamics-shift suite the authors built. A selector picks 2,000 failed states. A scripted oracle with privileged simulator state restores each state and returns one corrective action chunk. The policy is then fine-tuned on these corrections plus 20% nominal replay. EDGE ranks states by the product of three scores from a frozen latent world-model ensemble. G is a rank of how surprising the observed transition is. T is the soft fraction of nearby candidate actions whose predicted progress beats the observed one. C is a confidence weight based on ensemble spread and action distance. EDGE reports 58.9% on LIBERO-PRO and 56.2% on PhysicalShift, about 3 pp above the strongest simple baseline, a band-pass filter on G. A simulator-rollback study attributes most of the benefit to using the observed residual (+3.5 pp) and a smaller part to distance-based decay (+0.7 pp).

Critical strengths
  • The protocol is carefully controlled. All selectors share the pool, the budget and the update schedule, and the statistics are reported transparently, including non-significant results. This strength does not outweigh the weaknesses below.
Critical weaknesses
  • The problem and the method are hard to understand. The paper never plainly states the task it solves. The introduction jumps from VLA background to “acquisition” without saying who provides corrections, when, or how a past state is revisited. The key assumption, that a simulator snapshot is restored and a scripted oracle returns one chunk, only appears in Section 3 and Appendix D. G, T and C appear in the abstract and introduction only as coined names (“action-consequence surprise,” “teachability,” “confidence weight”). They give no intuition, and their meaning becomes clear only through dense equations in Section 4. Many other coined terms are used without definition, such as “support edge,” “residual transport,” “anchoring” and “mechanism seeds.” Figures 1 and 2 are overloaded, and the text is dominated by hedges and caveats rather than clear claims. A reader cannot tell what problem is solved or why the score should work, which on its own falls below the bar for ICLR.

  • The setting has unclear practical relevance and the evidence is weak. The formulation only works in simulation. It needs an exact past state to be restored, the expert is a scripted controller with privileged state, and no real-robot or cross-simulator results are given. Over the strongest simple baseline the gain is only about 3 pp. On PhysicalShift this gain is not significant after the authors’ own multiplicity correction (p = 0.108). Kernel decay, the method’s distinctive component, adds only 0.7 pp. Established acquisition or data-valuation methods (e.g., BADGE, DataMIL) are cited but not compared against.

Questions for authors
  • What is the concrete deployment scenario? How would a human provide a correction at a logged past state on a real robot without restoring it?
  • Please give intuitive, self-contained definitions of G, T and C early in the paper. Why is the score a product with unit exponents, and how sensitive are the results to this choice?
  • Given that the gain over the G band-pass baseline is not significant on PhysicalShift, and that kernel decay adds 0.7 pp, what is the main contribution?
  • How does EDGE compare with established acquisition methods, and with interactive DAgger-style correction under the same budget?
Reviewer_VXW2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses the question of which failed robot states should be sent for expert correction when the number of queries is fixed. EDGE uses an ensemble of frozen world-action models to compute scores for all failed states based upon G × T × C (surprise, teachability, confidence). T is calculated using the kernel-weighted observed residuals added to the predicted values for nearby actions. All experiments are in simulation, on LIBERO, a four-family subset of LIBERO-PRO, and PhysicalShift. Compared to the best-performing baseline (G-bandpass), the gains were 3.8 and 3.0 pp on the two shifted benchmarks. What I found most interesting was the same-snapshot study: most of what residual transport provides is due to constant anchoring (+3.5 pp), and kernel decay adds only 0.7 pp on top.

Critical strengths
  1. All selectors use the same pool, query budget, starting checkpoint, and update schedule; therefore, the only difference in performance among selectors is due to which states each selects.
  2. In the same-snapshot study (Section 6.5), the only thing that was changed was how the residual was transported; thus, it is clear what caused the improvement. Additionally, the authors showed where this approach did not provide benefits (e.g., kernel decay with action delay).
Critical weaknesses
  1. The scorer uses logged data only. However, what it ranks are states, one correction each, and this is only possible as an oracle can go back to the snapshot and get privileged state; see Section 3 and Appendix D. I can’t imagine any real-world robot channel that would provide a label for the specific state that was being scored. Section 7 places real-world robot validation in the category of “future work”. My concern is different: it is about the acquisition unit and not about validation.
  2. The difference between EDGE and G-bandpass was only 3.8 and 3.0 percentage points across five seeds; on PhysicalShift, there was no statistically significant difference between the two after Bonferroni correction (p = 0.108, Table 9).
Questions for authors
  1. Corrections are made in two ways: through online interventions or by gathering new demonstrations on a failure case. They are not made by labeling a previously recorded state. Which of these two does EDGE correspond to? I am especially concerned with whether same-state residual anchoring will be preserved when EDGE is mapped onto either channel, as neither provides labels for the states that were scored.
  2. Please provide the paired 95% confidence interval for the difference between EDGE-K and EDGE-C in Table 5; 0.7 pp is very small compared to the standard deviations of each row.
  3. σ_A = 1 determines how quickly the kernel decays and is thus responsible for the 0.7 pp. How was this value determined? A quick sweep of EDGE-K vs. EDGE-C over several σ_A values would be helpful.
  4. Appendix F states that the task-generalization and compounded-perturbation modes of LIBERO-PRO were excluded from the overall score, with only a short comment in Section 2 regarding matching the leaderboard protocol. I would like to receive a better explanation, and to know if EDGE was ever tested on those modes.

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

Related papers

Discussion (0)

Sign in to comment.