acceptodds
Withdrawn from ICLR 2027

PatchVLA: Process-Aware Test-Time Critic-Guided Search for Long-Horizon Manipulation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable action selection is essential for completing long-horizon robotic manipulation tasks. However, comparing candidates from generative vision-language-action (VLA) policies remains challenging because plausible action chunks from the same observation can imply different interaction timelines. To address this challenge, we propose PatchVLA, a process-aware framework for critic-guided action search at test time. A process-conditioned flow policy generates candidate chunks, each paired with temporally aligned forecasts of interaction phase, local progress, and subtask boundaries. A sequence critic evaluates the actions and forecasts together with the current observation. Calibrated critic scores and auxiliary penalties guide MPPI-like updates to the latent proposal distribution while learned weights remain fixed. The controller executes a short prefix of a selected candidate and replans from the next observation. Process forecasts are learned from demonstrations with expert anchors and human-reviewed VLM annotations, while recorded candidate outcomes supervise the critic. On LeHome-Fold and LIBERO-Long, PatchVLA achieves 76.25% and 96.20% success, respectively, exceeding the official π0.5 baselines by 9.17 and 2.80 percentage points. In a separate fixed-model LeHome-Fold study, two-round refinement achieves 73.96% success versus 70.21% for one-round selection. Together, these results show that candidate-aligned process forecasts and critic-guided latent search improve long-horizon action selection. The anonymous project page is available at https://patchvla-anonymous.github.io/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, action-conditioned critics, test-time refinement

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_2pfUOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PatchVLA, which pairs each candidate action chunk with aligned forecasts of interaction phase, local progress, and subtask boundaries, and uses a sequence critic to score the resulting action–process sequence. These scores refine the latent proposal distribution over multiple rounds before one evaluated candidate prefix is executed and the controller replans. Experiments on LeHome-Fold and LIBERO-Long report gains over single-round selection and the process-conditioned actor.

Critical strengths

Test-time scaling for VLAs feels like a very interesting topic.

Critical weaknesses
  1. The evaluation is too narrow to support the claims. There are no real-robot experiments, and only two simulation LeHome-Fold and LIBERO-Long are used. The figures in the paper are also hard to follow, containing many bars and columns whose meaning is unclear.

  2. The motivation is unclear, and the paper's positioning of its own method lacks scientific grounding. The ``process structure'' (phase, local progress, boundary) reads as repackaging task progress into subtask progress plus extra observation features, and the paper does not show what is fundamentally new beyond this.

Questions for authors

Please see the Critical Weaknesses above.

Reviewer_AHF9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PatchVLA, a framework for process-aware, critic-guided test-time search for long-horizon manipulation. A process-conditioned flow policy generates action chunks, and a candidate-conditioned predictor forecasts interaction phase, local progress, and subtask-boundary probabilities aligned with each chunk. A sequence critic evaluates the observation and action–process sequence. Calibrated success scores and optional uncertainty, smoothness, and continuity penalties guide gradient-free updates to a Gaussian latent proposal distribution. The controller executes a short prefix of one evaluated candidate and replans.

Training combines demonstration-based action and process supervision with critic learning from recorded candidate continuations. Process targets are constructed using expert anchors, VLM-assisted annotation, and human review. The headline results are 76.25% success on LeHome-Fold and 96.20% on LIBERO-Long. A separate fixed-model, eight-candidate-budget experiment reports 73.96% success for two-round refinement versus 70.21% for one-round selection. Additional repeated evaluations examine task completion, goal persistence, and command continuity.

The intended contribution is the integration of candidate-aligned process forecasts with sequence evaluation and latent-space proposal refinement, rather than direct action generation or one-shot candidate reranking alone.

Critical strengths
  1. Aligning phase, progress, and boundary forecasts with proposed actions is a sensible way to describe how different chunks may advance or transition between interactions. Local-progress resets are explicitly contextualized by phase and boundary predictions. This representation is potentially useful for long-horizon manipulation, where endpoint-oriented scores can miss important temporal distinctions.

  2. Comparing 4 × 2 refinement with 8 × 1 independent sampling holds the number of candidate evaluations fixed and is more informative than comparing search only against a single sample. The paper also reports process-masked controls, additional checkpoint repetitions, and both failure-to-success conversions and success-to-failure regressions. These analyses expose important limitations rather than hiding them, even though they currently weaken the strongest interpretation of the results.

Critical weaknesses
  1. Can you train an observation–action critic without process inputs using matched capacity, training data, optimization, and candidate budgets? How should the identical full-input and process-masked aggregate success in Table A3 be reconciled with the claim that process forecasts improve sequence evaluation?

  2. Can you evaluate the selected checkpoint on additional synchronized executions of the Process Actor, IID8, and MPPI controllers, without further checkpoint selection? Please provide uncertainty estimates that account for repeated garment assets and repeated initial states rather than treating all episodes as independent.

  3. How are recorded continuations generated, how much of each candidate is executed before continuation, and which continuation policy supplies the outcome labels? Does the critic retain calibration and ranking quality on second-round latent-refined candidates?

  4. What are the annotation volume, human-review effort, agreement statistics, and held-out phase/progress/boundary metrics? Can these metrics be reported separately for expert trajectories and generated successful or unsuccessful candidates?

  5. The primary LIBERO controller uses four candidates, an elite fraction of 0.25, and update rate one. Under Equation 5, the single elite has zero empirical variance, so the updated proposal variance reaches the configured lower bound before any subsequent adjustments. How does this concentrated resampling affect diversity and robustness, and how sensitive are results to using multiple elites or a smaller update rate?

  6. How much of the gain remains with matched execution prefixes, proposal reuse, and smoothness/continuity penalties? Can you provide controlled wall-clock comparisons for Actor, IID8, and two-round refinement on identical hardware and server loads?

Questions for authors

See the weakness.

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

This paper proposes PatchVLA, a framework for process-aware, critic-guided action search at test time. A process-conditioned flow policy generates candidate action chunks, and a candidate-conditioned predictor associates each chunk with temporally aligned forecasts of interaction phase, local progress, and subtask boundaries. A sequence critic evaluates the observation and the action–process sequence. Its calibrated scores, together with optional execution penalties, guide MPPI-like updates to the latent proposal distribution. The controller executes a short prefix of one evaluated candidate and replans, while all learned weights remain fixed during deployment.

The actor and process predictors are trained using annotated demonstrations, while the critic is trained using recorded candidate continuation outcomes. The paper reports 76.25% success on LeHome-Fold and 96.20% on LIBERO-Long. A separate fixed-model LeHome-Fold study compares two rounds of four candidates against one round of eight candidates, reporting 73.96% versus 70.21% success.

Critical strengths
  1. The candidate-conditioned process representation is well motivated and provides a coherent interface between action generation and evaluation.

Representing interaction phase, local progress, and subtask boundaries together is a sensible way to describe how a proposed action sequence would advance a multi-stage task. In particular, phase and boundary information provide context for interpreting local-progress resets. Conditioning these forecasts on individual action candidates is more informative in principle than assigning all candidates the same observation-level process estimate. The overall training and inference procedure is technically coherent: critic learning uses detached features and recorded outcomes, deployment uses predicted process signals, and execution selects an intact evaluated candidate.

  1. The paper includes a useful fixed-model, equal-candidate-budget comparison and informative execution diagnostics.

Table 5 compares (4\times2) refinement with (8\times1) selection while fixing the actor, critic, flow steps, execution prefix, and costs, and disabling proposal reuse. Refinement performs better in both reported executions, providing positive evidence that iterative allocation of a fixed candidate budget can be useful. The paper also evaluates two different manipulation settings and reports repeated executions, same-weights actor comparisons, goal persistence, command continuity, and both recovered failures and newly introduced failures. The explicit disclosure of checkpoint sensitivity and pairing limitations makes the empirical evidence easier to assess.

Critical weaknesses
  1. The independent contribution of candidate-aligned process forecasts to critic-guided selection is not convincingly established.

The “Process” comparisons in Tables 3–4 change the process-aware model configuration, including action generation, and therefore do not isolate the benefit of supplying candidate-specific process forecasts to the critic. The closer fixed-model comparison in Table A3 gives identical overall success for MPPI with full process inputs and with those inputs masked out: 73.96% in both cases. For one-round selection, the masked variant is slightly higher, at 70.63% versus 70.21%.

The full-input MPPI variant does perform better on unseen garments, at 63.75% versus 53.75%, which is potentially meaningful. However, this subset contains only 80 episodes across two executions, and the overall result remains unchanged. Moreover, masking a fixed critic at inference is not equivalent to training a matched action-only critic, so this experiment neither establishes that process inputs are unnecessary nor demonstrates their independent benefit.

This uncertainty is consequential because the structured, candidate-aligned process representation is the paper’s main distinguishing contribution. A comparison between retrained observation–action and observation–action–process critics, using the same actor, outcome data, search settings, and comparable capacity, is needed. Quantitative evaluation of process forecasts and their effect on within-state candidate ranking would also help establish the claimed mechanism beyond successful-rollout visualizations.

  1. The robustness and breadth of the incremental gains over strong actor and selection baselines remain insufficiently established.

Appendix C.2 states that the epoch-2 controller used in the fixed-model LeHome-Fold study was selected after inspecting five checkpoint evaluations. Its original two executions achieve 73.96%, while two additional executions achieve 71.25%, producing a pooled result of 72.60%. The additional executions do not include synchronized Process Actor or IID8 controls. The original equal-budget comparison is useful, but an independent confirmation is needed to establish how reliably its reported advantage persists after checkpoint selection.

On LIBERO-Long, Table 4 reports 96.20% for the full controller versus 95.80% for the Process Actor configuration. The separate four-execution, same-weights comparison with the (16\times2) controller is consistently positive, but its mean improvement is 0.95 percentage points. Figure A1 shows that task 09 alone contributes this net average gain, with improvements and regressions across the other nine tasks cancelling out. These results provide evidence of benefits on some tasks, but weaker support for broadly reliable improvements.

The practical tradeoff is also unresolved. Appendix C.5 reports LeHome-Fold median request durations of 743.0 ms for the controller and 86.9 ms for the actor, while explicitly noting unmatched server loads. A controlled success–latency comparison would help determine whether the additional computation is justified by the incremental gains.

Questions for authors
  1. Isolating the process contribution. With the same process-trained actor, branch data, critic capacity, search budget, costs, and execution settings, how does a retrained observation–action-only critic compare with the full process-aware critic? Can a candidate-independent or temporally shuffled process control establish whether candidate alignment itself improves ranking and closed-loop success?

  2. Checkpoint selection and independent confirmation. Which evaluation data were used to select the epoch-2 checkpoint, and were those data separate from the reported comparison cohort? Please clarify the checkpoint and evaluation relationship between the primary 76.25% result and the separate fixed-model study. Are synchronized Actor, IID8, and refinement results available for an independent evaluation cohort, with uncertainty estimates accounting for repeated initial states or garment assets?

  3. Reliability of the process forecasts and critic. What are the held-out phase, progress, and boundary prediction results on generated candidate branches, including unsuccessful candidates? Does adding process information improve within-state candidate ranking and calibration? Appendix C.3 reports only seven mixed-outcome groups among 240 LIBERO critic test groups; how much evidence does this provide for discriminating successful and unsuccessful alternatives from the same starting state?

  4. Performance relative to computation and closely related selection methods. Can the authors provide matched-hardware latency measurements and success–computation comparisons against the Process Actor and one-round selection? A comparison with the closely related VLA-ATTC framework, already cited in the paper, under a compatible candidate or wall-clock budget would further clarify the value of the additional process supervision and iterative refinement.

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

Related papers

Discussion (0)

Sign in to comment.