acceptodds
Under review as a conference paper at ICLR 2027

KPE-AC: Kinetic Path Energy for Adaptive Chunking in Vision-Language-Action Models

Abstract

Action chunking enables a vision-language-action (VLA) policy to predict multiple future actions in a single inference, improving efficiency while maintaining temporal coherence. A main challenge is selecting the execution horizon—that is, how many actions from the generated chunk to execute before querying the policy again. Short horizons allow frequent observation updates and more precise control but require more computation, whereas long horizons reduce inference frequency but delay responses to changes in the environment. Recent works have proposed adaptive execution-horizon selection methods, but this problem is still underexplored. In this work, we look inside the generation process of a single action chunk to guide adaptive action chunking. Inspired by Kinetic Path Energy analysis in image generation, we introduce Kinetic Path Energy-guided Adaptive Chunking (KPE-AC). We adapt KPE to action generation by measuring deviations from constant-velocity transport, yielding correction KPE. We analyze how this energy reflects uncertainty in future actions and leverage it to determine how many actions to execute. Beyond looking inside a single action chunk, we assess Cross-Chunk Consistency (CCC) by comparing actions predicted for the same future time steps in the previous and current action chunks. CCC adjusts the execution horizon while leaving the predicted actions unchanged. Experiments in simulation and real-world manipulation demonstrate the effectiveness of KPE-AC across multiple flow-based VLA backbones.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, robotic manipulation, adaptive action chunking, flow matching, inference efficiency

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_d6urOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies adaptive execution-horizon selection for flow-based vision-language-action models, motivated by the trade-off between frequent replanning and efficiently executing longer action chunks. The proposed KPE-AC looks inside the denoising trajectory of a single generated action chunk and introduces a correction version of Kinetic Path Energy (KPE), measuring deviations from constant-velocity transport rather than raw path energy. The method focuses on late-stage correction energy as a proxy for action uncertainty, uses it to select an initial execution horizon, and then further shortens this horizon according to Cross-Chunk Consistency (CCC), which measures disagreement between temporally aligned predictions from consecutive chunks. Importantly, KPE-AC does not require additional action samples or policy retraining. The paper evaluates the approach on RoboCasa, LIBERO with three flow-based VLA backbones, and two real-world manipulation tasks, and reports consistent improvements over the considered adaptive baselines.

Critical strengths

The paper addresses a practical issue in VLA deployment, namely how many actions should be executed before replanning. The proposed method is relatively simple and computationally appealing because it reuses intermediate velocity information that is already available during flow-based action generation. The experimental evaluation is also fairly comprehensive, covering multiple VLA backbones, several simulation benchmarks, and real-world robot experiments. The ablations on KPE correction, denoising windows, and CCC are useful and generally support the main design choices.

Critical weaknesses

My main concern is that the connection between correction KPE and action uncertainty is not fully convincing. The theoretical analysis is based on a simplified Gaussian setting, while the empirical correlation between KPE and action variability is relatively modest, with an AUROC of 0.601. In addition, several important components, including the late denoising window, adaptive threshold, and CCC-to-horizon mapping, are somewhat heuristic, and the paper provides limited sensitivity analysis for these choices. The gain over the best fixed-horizon baseline on RoboCasa is also fairly small, and no statistical significance is reported. The real-world experiments are encouraging, but the number of trials is limited and there is no direct comparison with other adaptive-horizon methods.

Questions for authors

I have three main questions for the authors. First, how sensitive is KPE-AC to the choice of the late denoising window, hmin⁡h_{\min}, threshold design, and CCC mapping, and how were these hyperparameters selected? Second, given the relatively weak empirical correlation between KPE and action variability, can the authors provide stronger evidence that KPE captures task-relevant uncertainty, such as action error or failure probability, rather than only stochastic variation between samples? Third, can the authors report confidence intervals or statistical significance for the main results and, if possible, include at least one adaptive-horizon baseline in the real-world experiments?

Reviewer_xsnCOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

KPE-AC is a training-free rule that decides the execution horizon of a flow-based VLA before replanning. It reuses the velocities the flow model computes while denoising one chunk and measures how much each future action still changes in the late denoising steps. It stops execution at the first action whose change is unusually large for that chunk, and shortens the horizon further when the new chunk disagrees with the unexecuted part of the previous one. A Gaussian model shows this energy grows with action uncertainty, and the authors report higher success than fixed horizons and than AAC and AutoHorizon on simulation and real-world manipulation tasks.

Critical strengths
  1. The signal is well chosen for the problem. It comes from a single generation the policy performs anyway, so it adds no sampling and no training. The threshold is parameter-free. Theorem B.1 is correct and gives the signal a clean interpretation in the Gaussian case, where the mean action cancels exactly and the energy scales linearly with the conditional standard deviation.
  2. On LIBERO, all backbone-suite combinations improve over the default horizon, with every baseline rerun on the same checkpoint over three environment seeds and 6,000 episodes per method and backbone, which makes even the small gains credible. The ablations isolate each part of the method, the two signals, the mean-velocity subtraction, and the denoising window.
Critical weaknesses
  1. KPE-AC is compared with the default fixed horizon instead of the best one. On RoboCasa, the best fixed horizon is h=30h = 30, which comes close to KPE-AC. Both single-signal variants fall below it, as shown in Table 1 and Table 4. The gain reported in Table 4 is measured against the worst fixed horizon, h=5h = 5. On LIBERO, only the default horizon is evaluated, and on the real robot only h=10h = 10 and h=50h = 50 are compared. The paper therefore never shows that KPE-AC beats a well-tuned fixed horizon. It also never tests a control that picks horizons at random from the same distribution as KPE-AC’s choices.
  2. The uncertainty signal is weak. Energy from a single generation is only a modest predictor of which action positions vary most across repeated samples. In its Gaussian setting, the energy of a single generation depends on the random initial noise as well as the uncertainty. Its proof implies the same for the late-window energy the method uses. The initial noise is known at inference, but the method does not account for it.
Questions for authors
  1. The late window measures deviation from the full-path mean velocity. Did you compare the late-window mean or the terminal velocity as the reference, and what justifies the full-path mean beyond the cancellation of μ\mu in the Gaussian model?
  2. In the scalar Gaussian model of Theorem B.1, one generation’s energy is (1−π/4)σξ2(1 - \pi/4)\sigma\xi^2, so it also scales with the squared initial noise. Does normalizing each position’s energy by its squared initial noise ∥ϵi∥2\|\epsilon_i\|^2 raise the AUROC and the success rates in Tables 1 and 2?
  3. What do fixed horizons of 10, 25, and 50 for π0\pi_0, 10 for π0.5\pi_{0.5}, and 4 and 8 for GR00T reach on LIBERO? What is hminh_\text{min} for GR00T, and what does a control that samples horizons from KPE-AC’s own selected-horizon histogram, independently of the signal, reach on RoboCasa? Can you please also add the three-seed spread to Table 2?
  4. Are the thresholds and execution bounds of AAC and AutoHorizon configured per backbone and chunk length? What are their selected-horizon distributions on π0\pi_0, whose 50-step chunk is where both baselines lose the most against their own backbone?
Reviewer_PjvEOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KPE-AC, which adaptively sets the execution horizon of flow-based VLA policies. It uses the correction kinetic path energy of the denoising trajectory as a proxy for action uncertainty, plus a cross-chunk consistency check that compares overlapping predictions. Because the signal comes from velocities the sampler already computes, the method needs no retraining and adds almost no overhead. It is backed by a theoretical analysis and by empirical evidence linking correction energy to prediction variance, and it improves success rates across π0, π0.5 and GR00T N1.5 in simulation and on real robots.

Critical strengths
  1. Training-free and nearly free to compute. The uncertainty signal comes from the denoising velocities that flow-based VLAs already compute, so the method can be added to existing pretrained models without retraining or extra sampling. This makes it practical to adopt.

  2. Two complementary signals that leave the actions unchanged. KPE measures uncertainty within a single chunk, and CCC checks whether consecutive chunks agree. Both only decide when to replan, never what the actions are, which keeps the method simple and backbone-agnostic. It gives consistent gains on π0, π0.5 and GR00T N1.5.

Critical weaknesses
  1. KPE-AC beats the best fixed horizon (h=30) on RoboCasa by only 1.2 pp, which is within noise, and the LIBERO gains are 0.6–2.2 pp on saturated suites. No seeds or confidence intervals are reported.

  2. The paper is motivated by compute cost, yet it reports no query counts, executed horizons or wall-clock time.

  3. The paper is motivated by efficiency, yet the conclusion and experiments report only success rates. The limitations section also leaves out the small gain over a tuned fixed horizon and the inability to anticipate later changes in the scene. (correct me if i am wrong)

Questions for authors
  1. Since the method needs no extra samples or retraining, could you quantify its runtime overhead compared with the base policy?

  2. Also, since it is limited to flow-based generators, how might it extend to diffusion-based or autoregressive VLAs, for example by using denoising updates or token-level confidence?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.