acceptodds
Under review as a conference paper at ICLR 2027

World-Calibrated Proposal-to-Action Flow for Vision-Language-Action Models

Abstract

Flow-based Vision-Language-Action (VLA) policies generate action chunks by transporting samples from a task-agnostic isotropic Gaussian source. As this source is conditioned on neither recent execution nor predicted future evolution, **(i)** it discards the local continuity established by recently executed motion. **(ii)** Even when predictive world representations are introduced, they often only condition the transport dynamics rather than determine where generation starts, how far it may deviate, or along which action directions it may expand. Building on this observation, we introduce **ProAct**, a world-calibrated proposal-to-action framework that makes the generative source itself predictable. **(i)** To preserve motion continuity, a lightweight Proposal Expert converts recent actions into a scene-aware hypothesis via one motion-anchored endpoint flow-matching step, initializing generation near the demonstrated action manifold. **(ii)** To jointly capture intended scene evolution and proposal–future compatibility, a prospective World Expert treats the hypothesis as a soft motion prior while predicting the task-consistent latent future. **(iii)** From this compatibility, the model calibrates a proposal-centered anisotropic source, where a bounded per-step extent controls the allowed deviation and a trace-normalized low-rank geometry under a condition-number budget allocates refinement over coupled translation, rotation, and gripper directions. Compared with , **ProAct** improves performance across simulation and real-world tasks while reducing denoising steps by 50%, inference latency by up to 25.8%, and increasing throughput by up to 34.8%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotic Manipulation, Vision-Language-Action Models, Emboided Intelligence

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ULNyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

As I understand it, the paper proposes a steering approach to guide diffusion/flow policies in Vision-Language-Action (VLA) models. The main argument presented is that VLA policies lack conditioning. The author proposes PROACT, which uses two components: a motion expert that proposes a motion prior and a world-modelling component that refines it; it then uses this to steer the noise source for action generation, thus providing a steering algorithm for VLA. They show the approach works quite well on the benchmarks and outperforms baselines

Critical strengths
  • The framing that providing additional conditioning is reasonable and should help a policy smoother and locally continuous in action trajectory. I think their propose-calibrate-refine idea is intuitive indeed.
  • PROACT was evaluated on multiple benchmarks, including real-world manipulation tasks. The results seems quite strong against the baselines.
Critical weaknesses

The paper should've been desk-rejected due to the severe lack of care when it comes to scientific communication, such as a lack of writing quality (grasping a high level of understanding of the contribution to begin with). The paper is very hard to follow, as it is filled with jargon. In the abstract, when I first started reading it, I needed to make multiple attempts to keep my focus on it , because every sentence is coming up with a term that is hard to follow. This goes on and on in the full paper. I like to print paper to read, but it's impossible to follow the figures because they are too tiny. The paper doesn't have a standard paper structure one would have started with a related work section after the introduction; neither they have a preliminary introduction to the topic or a background section. The method section is filled again with jargon for 3 long pages. The experiments section doesn't have a proper description of the baselines. The related work, which appears before the Conclusion section, doesn't discuss VLA, so a general reader was never introduced to the domain. I'm reasonably familiar with this area, and when I started reading the paper, the first paper that came to mind was diffusion steering [1]. The author didn't place themselves in relation to prior works (I couldn't find discussion of policy steering that was well studied in prior works). As it stands, an almost complete disregard for basic effort in writing and communication to a reader.

[1] Wagenmaker, Andrew, et al. "Steering your diffusion policy with latent space reinforcement learning." arXiv preprint arXiv:2506.15799 (2025).

Questions for authors

None

Reviewer_2Z9fOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ProAct, a flow-based vision-language-action model that uses recent motion and future-supervised features to guide action generation. It forms a motion proposal, calibrates the source distribution around it, and refines the sampled actions. Experiments in simulation and on real robots show improved average success rates and lower inference latency under the reported settings.

Critical strengths
  1. The paper proposes a well-motivated way to use recent motion and scene information to initialize action generation. It separates the scale of the source noise from its directional structure and constrains the covariance to avoid degenerate estimates.

  2. The paper evaluates the method across simulation benchmarks and real-world manipulation tasks on different robot platforms. The improvements on LIBERO-Plus and physical robots support its practical value under varied conditions.

Critical weaknesses
  1. The main RoboTwin comparison gives an overly positive impression of robustness by focusing on ten selected tasks, where ProAct achieves 60.3% versus 28.6% for π₀.₅. Across all 50 tasks in Table 9, the gain is much smaller: 47.06% versus 43.84%, with performance dropping on 24 tasks. Some drops are large, such as Rotate QRCode (15% versus 56%) and Click Alarmclock (61% versus 93%). Although the full results are provided in the appendix, explaining these failures is necessary to understand when the method helps or hurts.

  2. The claim of a 50% reduction in denoising steps needs a clearer comparison. Table 7 reports latency and success under default settings but does not give the inference step counts. Could π₀.₅ achieve similar success with fewer steps? Comparing success and total latency over several step counts would help answer this question. The timing should include the additional experts and attention operations, and the throughput metric should be defined.

  3. The ablation study leaves the contribution of source calibration to generalization unclear. Table 5 removes calibration while retaining the World Expert, but its LIBERO / real-world results (97.8% / 76.0%) differ from the fixed-isotropic source in Table 6 (96.9% / 72.7%). The difference between these settings needs clarification. The ablations also do not cover LIBERO-Plus, where the improvement is much larger. Testing the source variants there with the same world-feature input would help separate the benefit of calibration from that of the world features.

  4. The reported gains from low-rank geometry are small and lack uncertainty estimates. Table 6 shows that adding geometry to extent-only calibration improves LIBERO by 0.6 percentage points and real-world success by 2.7 points. With six tasks and 25 trials per task, the latter amounts to about four additional successes. Confidence intervals and results across training seeds would help establish whether these gains are reliable.

Questions for authors
  1. What causes the large RoboTwin regressions, particularly on Rotate QRCode and Click Alarmclock? Are they related to the motion proposal or its calibration?

  2. What inference step counts are used in Table 7, and which computations are included in the timing? How does π₀.₅ perform with fewer steps, and what does the throughput metric count?

  3. How does the no-calibration setting in Table 5 differ from the fixed-isotropic source in Table 6? How do the source variants compare on LIBERO-Plus when the world-feature input and inference budget are held fixed?

  4. How many training seeds were used? Can the authors report confidence intervals and variation across seeds for the full model and extent-only calibration?

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

The paper proposes PROACT, a robot policy that starts action generation from a motion proposal instead of generic random noise. The method first uses recently executed actions and the current scene to suggest a likely next movement. A second module predicts a representation of the desired future and estimates how much the proposal may need to change, and in which combinations of action dimensions. The final action generator then refines samples drawn around this proposal. The aim is to use motion continuity without simply repeating movements that are no longer appropriate after a task transitions. Experiments cover many benchmarks such as LIBERO, LIBERO-Plus, RoboTwin 2.0, and six real-world tasks on two robot platforms. The strongest reported improvements are under visual changes and on some difficult manipulation tasks. In particular, on the complete 50-task RoboTwin benchmark, the average improvement over the base pi_0.5 policy is 3.22 percentage points.

Critical strengths
  1. The method is well-motivated. Recent motion is useful information, but directly repeating it can be unreliable when the robot needs to grasp, change direction, or move to the next stage of a task. Treating that motion as a proposal that can be corrected is a sensible way to balance continuity with the current scene.

  2. The ablation studies examine the main components respectively. Tables 4 - 6 compare different motion proposals, remove future prediction and proposal conditioning, and separate the amount of source variation from its direction. I appreciate that these comparisons include real-world tasks rather than only a nearly saturated simulation benchmark. The results support the usefulness of both the motion proposal and the later refinement.

  3. The evaluation is broad and includes two physical robot platforms. The distinction between completing the full task and reaching intermediate checkpoints makes the real-world results more informative. Table 7 also reports measured inference latency, not just a reduction in generation steps, which is important because the added proposal and future-prediction modules also take time.

Critical weaknesses
  1. Analysis of the broad picture of the experiments. Table 2 contains ten selected tasks, and the text reports 60.3% versus 28.6% for pi0.5, a 31.7-point improvement on that subset. However, Appendix B reports 47.06% versus 43.84% over all 50 tasks, a 3.22-point improvement. The complete table shows improvement on 22 tasks, ties on four, and lower performance on 24. The authors do provide these full results, but the overall average and this mixed task-level picture should appear in the main results. I would also like to see an analysis of the tasks where performance decreases, such as Rotate QRCode and Click Alarmclock, to understand when the motion prior becomes unhelpful.

  2. The quality of future prediction needs more direct validation, especially when the proposal is wrong. The World Expert predicts the future from the demonstration, not the future that would result from executing the proposed action. I am curiuos whether the estimated correction remains reliable after a mistaken action history. If we have any experiments on this? Could the authors deliberately introduce incorrect or stale motion proposals and measure whether the predicted correction increases appropriately and whether the robot recovers?

  3. The uncertainty of the reported improvements is not very clear. There are 25 real-robot trials per task, and several ablation differences correspond to only a small number of additional successes. The tables do not report confidence intervals, and I could not find the number of independent training seeds. This matters both for the small gains on standard LIBERO and for the full RoboTwin average.

Questions for authors
  • Section 4 discusses Streaming Flow Policy and A2A, but why the main experiments do not compare against them? The last-chunk and motion-extrapolation ablations are useful, but they are simpler than these related methods.

  • For the efficiency claim, Table 7 gives one default operating point per method, while Appendix A.1.3 leaves the number of integration steps symbolic. I am wondering if authors can report the actual step counts and a success-versus-latency curve for both PROACT and pi0.5, including pi0.5 with the same reduced number of steps.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.