acceptodds
Under review as a conference paper at ICLR 2027

DepthFlow-VLA: Routed Low-Noise Flow Refinement for Compact VLA Models

16 views

Abstract

Compact vision-language-action (VLA) models commonly allocate the same action-decoding budget to every chunk despite substantial variation in prediction difficulty. We present DepthFlow-VLA, a coarse-to-refine flow-matching decoder in which Stage A first produces a coarse action and a lightweight router selectively invokes restart-style low-noise Stage B. Training combines difficulty-weighted velocity supervision, routing supervision, and a Stage-B objective constructed from replayed Stage-A predictions. On LIBERO, Stage-A-only, adaptive, and uniform-refinement modes achieve 95.9%, 98.3%, and 98.5%, respectively, while adaptive routing uses 7.54 rather than 13.0 decoder evaluations per chunk. On SimplerEnv-WidowX, adaptive and uniform refinement achieve 80.2% and 81.3%, and equal-budget random routing reaches 76.0%. A separately trained vanilla flow-matching control reaches 94.8% on LIBERO and 72.2% on WidowX at six decoder evaluations. On a single H100 with batch size 1, adaptive routing reduces mean end-to-end latency from 59.2 to 42.6 ms relative to uniform refinement.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, robot learning, robotic manipulation, conditional flow matching, adaptive computation, dynamic routing, coarse-to-refine action decoding, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_T9XdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DepthFlow-VLA, an adaptive coarse-to-refine flow-matching decoding framework for compact vision-language-action (VLA) models. The method first generates a coarse action prediction through Stage A and then uses a lightweight router to determine whether the current action chunk requires further Stage-B refinement. Rather than simply extending the original decoding trajectory, Stage B re-corrupts the Stage-A coarse prediction with low-level noise and performs a dedicated low-noise correction process.

During training, DepthFlow-VLA combines difficulty-weighted velocity supervision, routing supervision, and a Stage-B refinement objective constructed from replayed Stage-A predictions. Experiments on LIBERO, SimplerEnv, and real-world robotic manipulation tasks show that adaptive routing retains most of the performance of uniform refinement while substantially reducing decoder evaluations and inference latency.

Critical strengths

None

Critical weaknesses

This paper proposes DepthFlow-VLA, which dynamically allocates inference computation to action chunks of different difficulty through coarse-to-refine flow decoding and adaptive routing. The overall approach is clear and straightforward, and it achieves strong results on LIBERO, SimplerEnv, and real-world robotic tasks while substantially reducing decoder evaluations and inference latency. However, the validation of the core mechanism remains insufficient. Stage B is specifically trained on re-corrupted Stage-A predictions, whereas the continuation and no-recorruption baselines only modify the inference path of the same checkpoint. Therefore, the observed performance differences may be partly caused by a train–test mismatch, and the paper appears to lack sufficient validation to rule out this possibility. In addition, the correlation between the router score and the actual refinement benefit remains relatively limited.

Questions for authors
  1. Fairness of the refinement comparison. Stage B is specifically trained on re-corrupted Stage-A predictions, while the continuation and no-recorruption baselines only modify the inference path of the same checkpoint. How does the paper rule out train–test mismatch as the main reason for the lower performance of these baselines?

  2. Router supervision. The router is trained using the midpoint reconstruction residual rather than the actual benefit brought by Stage-B refinement, and its correlation with the refinement benefit is relatively limited. Could the authors explain why this supervision design is appropriate, and to what extent the mismatch between the training signal and the actual refinement benefit may limit performance?

  3. Novelty relative to recent adaptive VLA inference methods. How does DepthFlow-VLA fundamentally differ from recent adaptive-computation approaches such as AC²-VLA and AdaVLA in terms of adaptive inference and compute allocation? AC²-VLA introduces action-context-aware adaptive computation for VLA models, while AdaVLA dynamically adjusts the number of inference steps for flow-matching-based VLAs.

Reviewer_VQoSOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DepthFlow-VLA, a compact VLA decoder that selectively routes coarse actions to refine, balancing the compute and results. The denoise process is split into two stage. Stage A generate coarse actions. Then a small MLP predicts the routing score. Some action chunks with higher score are routing into Stage-B to refine again. Benchmark comparisons examine the design of refinement and routing. On LIBERO, Stage-A-only, adaptive, and uniform-refinement modes achieve 95.9%, 98.3%, and 98.5%, respectively. On SimplerEnv-WidowX, adaptive and uniform refinement achieve 80.2% and 81.3%, and equal-budget random routing reaches 76.0%. A separately trained vanilla flow-matching control reaches 94.8% on LIBERO and 72.2% on WidowX at six decoder evaluations. In addition, selection routing reduces mean end-to-end latency from 59.2 to 42.6 ms relative to uniform refinement, improving the efficiency of VLA.

Critical strengths
  1. The main advantage of this paper is that it propose an efficient adaptive computation allocation mechanism, which significantly reduces the computational overhead of VLA action decoding while maintaining a high robot operation success rate. DepthFlow-VLA uses a coarse-to-refine two-stage design: Stage-A generates rough actions, Router determines whether further optimization is needed, Stage B performs low-noise refinement only on the selected action chunks. This design has strong practical value.
Critical weaknesses
  1. There is an inconsistency between the training objective of Router and the actual refinement benefit. The paper uses the action reconstruction error of Stage-A midpoint as the difficulty proxy, and generates binary labels (Appendix A.1). However, action chunks with larger prediction errors may not necessarily benefit more from Stage-B refinement, so there is a gap between the current supervised signal and the optimal computation allocation target. In addition, Router receives noisy midpoint estimate constructed by ground-truth action during training, but coarse prediction generated by Stage-A multi-step decoding during inference. Train-inference distribution gap may be generated. In Table 4, the Spearman correlation between Router and actual refinement benefit is only 0.36, and the recall is 61.2%. It shows that there is still a large space for recognition improvement of the samples that really need refinement.
  2. While Table 5 compares methods such as single-stage integration, routed continuation, and no-re-corruption, However, these comparisons use checkpoints that have been trained on DepthFlow-VLA. Such a setting may benefit the authors for Stage-B refinement, since the training objective itself specifically optimizes this decoding path.
Questions for authors
  1. Router uses ground-truth-corrupted midpoint estimate at training time, while the coarse prediction actually generated by Stage-A is used at inference time. The two action estimates may have different error distributions. Can the authors explain the rationality of the design and analyze the impact of this distribution mismatch on routing accuracy?
  2. The current difficulty label is constructed by reconstruction error, while the true refinement benefit depends on how much prediction error Stage-B can reduce. Table 4 shows that the agreement between the two is still limited. Can the authors explain the reasons for choosing this proxy and the relationship between it and the true refinement benefit?
  3. It is recommended to train baselines such as single-stage flow matching, routed continuation and no-re-corruption separately, and to match training FLOPs or GPU-hours. And decoder evaluations in their inference phase.
  4. Add more inference efficiency experiments on real robots with more adequate statistical validation. Currently the real robot experiments also only cover three tasks on a tabletop manipulator platform. It is recommended to report the end-to-end control latency, actual control frequency in real deployment.
Reviewer_1gMdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DepthFlow-VLA uses a shared flow-matching decoder to generate coarse action chunks and selectively refine them through a learned router and low-noise restart. This work was evaluated on LIBERO, two SimplerEnv embodiments, and three real-world manipulation tasks.

Critical strengths
  • The controlled experiments provide useful evidence for selective refinement (e.g., Tables 5 and 15, Appendix B.4)

  • The evaluation connects logical decoder cost to measured model-inference latency and tests more than one manipulation setting. The H100 measurements include the backbone and routing overhead.

Critical weaknesses
  • The LIBERO evaluation horizon setting needs more clarification. From Appendix B.1, the experiment setting is allowing 800 environment steps for Spatial/Object/Goal and 900 for Long. However, the official OpenVLA evaluation script (https://github.com/openvla/openvla/blob/main/experiments/robot/libero/run_libero_eval.py) uses 220/280/300/520 steps, respectively, plus stabilization. Unless these steps have different physical meanings, the proposed protocol allows substantially more time to complete each task.

  • The evaluation does not establish whether selective refinement remains effective under distribution shift. Section 4.1 and Appendix C train separate policies for LIBERO and the two SimplerEnv embodiments, while the real-world policies are adapted using task-specific demonstrations. These experiments do not test whether a fixed trained router still identifies useful refinement opportunities when distribution changes.

  • More evaluation on the generalizability is needed. An evaluation such as LIBERO-Plus (https://arxiv.org/abs/2510.13626) can increase the diversity of the benchmarks used in this work. The existing experiments also leave coordinated bimanual manipulation untested, a diverse, randomized suite such as RoboTwin (https://robotwin-platform.github.io/) would better support the breadth of the approach.

Questions for authors
  1. Why are the LIBERO rollout limits 800/800/800/900? How many successful rollouts finish within those limits?

  2. With the LIBERO-trained router, how does adaptive refinement compare with Stage A, equal-budget random routing, and uniform refinement on LIBERO-Plus or equivalent benchmarks?

  3. Does the accuracy-compute trade-off persist on a diverse bimanual benchmark such as RoboTwin?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.