acceptodds
Under review as a conference paper at ICLR 2027

Overcoming Dynamics-Blindness: Training-Free Pace-and-Path Correction for VLA Models

Abstract

Vision-Language-Action (VLA) models achieve remarkable flexibility and generalization beyond classical control paradigms. However, most prevailing VLAs are trained under a single-frame observation paradigm, which leaves them structurally blind to temporal dynamics. Consequently, these models degrade severely in non-stationary scenarios, even when trained or fine-tuned on dynamic datasets. Existing approaches either require expensive retraining, or suffer from latency bottlenecks and poor temporal consistency across action chunks. We propose Pace-and-Path Correction (PPC), a training-free, closed-form inference-time operator that wraps any chunked-action VLA. Under a local motion model, joint minimization of a single quadratic cost yields an interior solution with two orthogonal channels. The pace channel compresses execution along the planned direction, while the path channel applies an orthogonal spatial offset, jointly absorbing the perceived dynamics within the chunk window. We evaluate our approach on a comprehensive diagnostic benchmark MoveBench designed to isolate motion as the sole controlled variable. Under the controlled MoveBench protocol, PPC outperforms the evaluated training-free wrappers and dynamic-adaptive baselines, raising success by up to 28.8% and 25.9% in absolute terms over foundational VLAs in dynamic-only and static-dynamic mixed environments, respectively. We further validate PPC on a real robot with a backbone fine-tuned only on static demonstrations, increasing dynamic grasping success from 0% to 68.3%.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Vision-Language-Action Models, Inference-time Adaptation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8FXxOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the dynamics-blindness of chunked Vision-Language-Action models, where a policy predicts a sequence of actions from a single observation and may therefore fail to react to object motion during open-loop execution. The authors propose Pace-and-Path Correction (PPC), a training-free inference-time wrapper that uses an external velocity estimate to modify the action chunk through two closed-form components: a pace term that adapts execution length along the planned direction and a path term that compensates for spatial drift orthogonal to it. A hierarchical 2-EMA latch is further introduced for irregular motion. The paper also presents MOVEBENCH, a controlled benchmark covering static, uniform, accelerated, and irregular motion regimes, and evaluates PPC across multiple VLA backbones as well as on a real xArm6 setup. The reported gains are substantial, including up to 28.8 percentage points on dynamic settings and a real-robot improvement from 0% to 68.3% with a backbone fine-tuned only on static demonstrations.

Critical strengths

The paper addresses a practically important weakness of current chunked VLAs with a method that is simple, lightweight, and easy to integrate without retraining or architectural modification. The closed-form formulation is appealing, and the separation into pace and path correction gives a reasonably interpretable view of how motion is handled. Experimentally, the paper is stronger than many training-free wrapper works: it evaluates several foundational VLAs, includes component ablations, equal-velocity-access controls, execution-horizon controls, perception-noise studies, non-oracle tracking, and real-robot experiments. In particular, the matched feedforward and horizon baselines help show that the gains are not explained solely by having access to velocity information or by simply re-observing more frequently.

Critical weaknesses

The main concern is that the strongest headline simulation results use an oracle dynamics signal derived from simulator ground-truth object motion, while most comparison methods do not receive equivalent motion information, so the main table partly conflates the value of PPC’s correction structure with the value of privileged dynamics access. The appendix contains useful equal-access controls, but these are limited to two backbones and relatively simple feedforward baselines rather than all major competing methods. A second concern is that some of the theoretical framing is broader than the regime where the clean orthogonal interpretation strictly holds: the derivation relies on quasi-stationary plan and disturbance assumptions, and when the lower bound on the pace factor becomes active, the full disturbance is passed to the path channel rather than remaining orthogonally decomposed. MOVEBENCH is well controlled but relatively narrow, being centered on a single pick task with fixed scene structure, so the claims about general dynamic manipulation should be interpreted with some caution.

Questions for authors

Could the authors clarify the fairness of the main comparison under oracle velocity access, and ideally report non-oracle RGB-D-tracker results for all major PPC backbones or provide stronger equal-information baselines in the main table? How often is the pace lower bound actually active in MOVEBENCH, and what fraction of the observed gains occurs in the true interior regime where the claimed orthogonal pace-path decomposition holds?

Reviewer_QL3DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PPC, a training-free wrapper during inference for any VLAs. It reads an externally estimated object velocity and corrects the chunk interior through two closed-form channels. The Pace channel accelerates execution along the planned direction and compresses the execution horizon, and the Path channel adds per-step spatial offsets. An EMA latch is applied to shorten horizons under irregular motion. Alongside the proposed wrapper, the paper also constructs MOVEBENCH, a ManiSkill pick benchmark that isolates the target’s motion regime as the primary evaluation axis. Empirical results on simulation and real-robot experiments against foundation VLAs demonstrate PPC's effectiveness.

Critical strengths
  1. The formulation for establishing the core closed-form operator is clean and correct. It introduces no learned parameters and only a set of constants, with the latch’s internal thresholds derived from one free rate and the chunk geometry. This supports the motivation to make it training-free and backbone-agnostic. The λ\lambda-family makes the simulation and hardware variants two instances of one cost.
  2. The experiments are well designed and reported cleanly, with confidence intervals on the pooled comparisons and the real-robot results. In the real-robot experiment, the conveyor result with a static-only fine-tuned π0.5\pi_{0.5} shows cleanly that the signal can be injected into a frozen VLA.
Critical weaknesses
  1. The closed form is not demonstrated to beat a constant offset. The hardware deployment switches to the λ→0\lambda \rightarrow 0 constant-offset limit after a pilot. The constant profile is not ablated in simulation, while the Fibonacci profile does not run on hardware. Linear δ\delta shown in Table 2 is described only as replacing the Fibonacci profile with linear offsets without clarifying what those offsets are.
  2. On π0.5\pi_{0.5}, either replanning every 2 steps or naive velocity feedforward alone already closes much of the gap to PPC, and likewise on GR00T. Frequent replanning fails on GR00T, which the paper attributes to late gripper closure. The two controls are never combined into one baseline that applies the feedforward correction and replans every 2 steps. That is the simplest competitor with both of PPC’s ingredients and no closed form.
Questions for authors
  1. What does the λ→0\lambda \rightarrow 0 constant profile score in simulation, and what does the λ=1\lambda = 1 Fibonacci profile score on the conveyor? What exactly is the “Linear δ\delta” variant of Table 2?
  2. What do π0.5\pi_{0.5} and GR00T reach with naive velocity feedforward at Kexec=2K_\text{exec} = 2 and at Kexec=5K_\text{exec} = 5, that is, with the two controls of Tables 10 and 12 combined, and with second-order extrapolation on the accelerated family, where the Lucas branch currently reads an acceleration the feedforward control is not given? I believe that is the baseline PPC needs to beat to carry its claim.
  3. On the conveyor, what does the bare reach at Kexec=2K_\text{exec} = 2 on the same asynchronous stack?
  4. π0.7\pi_{0.7} and π0.6\pi_{0.6}-MEM receive and consume frame history, so the single-frame premise no longer describes the frontier, and the 10K MOVEBENCH demonstrations already contain motion. Does PPC still help a backbone fine-tuned with the previous frame as an extra image input, and will “structurally blind” be scoped to only the single-frame backbones evaluated rather than VLAs in general?
Reviewer_iuuSOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the limited responsiveness of chunked Vision-Language-Action (VLA) policies to moving targets. It proposes Pace-and-Path Correction (PPC), a training-free method that uses external motion estimates to adapt policy execution. The authors also introduce MoveBench for controlled evaluation of motion robustness and demonstrate PPC’s effectiveness in simulation and real-world dynamic grasping.

Critical strengths

PPC improves the responsiveness of existing VLA policies without additional policy training or architectural changes. Its closed-form formulation yields interpretable, computationally lightweight corrections, while experiments across multiple backbones support its applicability as an inference-time component. The proposed MoveBench further enables systematic evaluation of VLA robustness across different motion patterns and intensities. Overall, the paper is clearly written and well organized.

Critical weaknesses

Although Appendix F provides trajectory comparisons and correction-related signals, it does not directly show how the pace and path channels transform the original action chunk. Time-aligned visualizations of the target trajectory, original and corrected chunks, and tracking error would clarify the contributions of both channels and provide more direct empirical support for the proposed mechanism.

Questions for authors
  1. How would PPC extend to tasks where the moving object must be approached from a specific direction, or where the motion is coupled with the task itself, such as a handover or pouring onto a moving cup?
  2. Could the authors explain why removing the cosine factor in Table 2 lowers success on uniform and accelerated motion but improves it on irregular motion?
  3. Appendix A.7 derives the correction solution for arbitrary λ>0\lambda>0, and Appendix G.4.2 explains the choices of λ=1\lambda=1 in simulation and λ→0\lambda\to0 on hardware. Could the authors additionally report a sensitivity analysis over several λ\lambda values, including the full-compensation limit, using a fixed backbone checkpoint and evaluation setup? This would help quantify the effect of regularization strength and provide further empirical support for these configuration choices.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.