PaperlayerView exact revision on arXiv

Source-linked AI summary

Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies

Jiabing Yang, Yuan Xu, Yixiang Chen, Qisen Ma, Peiyan Li, Kai Wang, Chenhang Cui, Tao Yu, Zichen Wen, Wen Ye, Bowen Fang, Ziheng He, Xiangnan Wu, Yingda Li, Yan Huang, Liang Wang

arXiv:2610.32102v1cs.RO

TL;DR

Flow-matching VLA deployment usually treats solver depth and execution horizon as separate choices, despite their joint effect on trajectory dispersion and open-loop commitment. This paper studies their coupling and explains it through dispersion contraction around an invariant centroid. It reports a sign reversal: shallow sampling can hurt rapid replanning yet improve extended open-loop execution.

  • Problem

    Standard deployment workflows treat solver depth and execution horizon as independent choices, although solver depth governs action dispersion and horizon sets how long it compounds before feedback.

    Source: §I · p. 1

  • Method

    The paper investigates how early Euler truncation contracts stochastic trajectory dispersion around an invariant centroid across frozen VLA checkpoints.

    Source: §I · p. 1

  • Results

    Shallow sampling degrades task success under rapid replanning but outperforms deep sampling under extended open-loop execution, with the LIBERO-Long π0 contrast crossing zero between s=5 and s=10.

    Source: §I · p. 1

  • Takeaways & Limitations

    Sampling depth can serve as an operational variance regulator, enabling training-free joint adjustment of solver depth and execution horizon to improve success while amortizing inference latency.

    Source: §I · p. 2, C. Coupling Mechanism: Dispersion… · p. 3

Abstract

from arXiv · show

Flow-matching Vision-Language-Action (VLA) models generate expressive continuous trajectories for robotic manipulation, yet balancing real-time responsiveness with execution stability remains a persistent challenge. Standard deployment workflows typically treat numerical solver depth and execution horizon as decoupled choices. The horizon amortizes inference latency, while solver depth is kept high to preserve integration fidelity. In this paper, we reveal a fundamental compute-execution coupling in flow-matching VLAs governed by a systematic sign reversal. While coarse numerical integration degrades precision under rapid replanning, it markedly improves task performance under extended open-loop commitment. Geometrically, numerical truncation contracts lateral action dispersion around an invariant trajectory centroid, serving as an implicit regularizer that suppresses compounding open-loop drift. Across LIBERO, RoboTwin 2.0, and real-world Franka manipulation on representative VLA backbones including $\pi_0$, $\pi_{0.5}$, and GR00T, exploiting this joint coupling consistently achieves superior task success while drastically lowering inference latency, entirely training-free and without architectural modifications. These findings establish numerical sampling depth as an operational variance regulator, shifting its role from a static precision budget to an active deployment knob in robot learning. Project page: https://jiabingyang01.github.io/Horizon-Depth-VLA/

I. INTRODUCTION

Flow-matching VLA deployment conventionally treats execution horizon and solver depth as separate choices, although both shape how action uncertainty interacts with feedback. The paper reports that coarse sampling reverses from harmful under rapid replanning to beneficial under extended open-loop execution.

I. INTRODUCTION

  • Prior deployment practices tune horizon and solver acceleration separately, typically holding one fixed while adjusting the other.Receding-horizon methods sweep execution horizons, whereas acceleration methods distill or truncate solvers under standard short horizons.Source: §I · p. 1
  • Solver depth governs generated-chunk dispersion, while execution horizon determines how long that dispersion compounds before sensory feedback intervenes.The proposed account is that shallow Euler truncation contracts trajectories around an invariant centroid, acting as an implicit regularizer against open-loop drift.Source: §I · pp. 1–2
  • Shallow sampling degrades task success at s=5 but outperforms deep sampling at s=20, with the contrast crossing zero between s=5 and s=10 on LIBERO-Long with π0.RoboTwin 2.0 experiments replicate the behavior on π0.5, with boundaries tied to predictive horizons and baseline task capacity.Source: §I · p. 1
  • Diagnostic interventions support redraw stochasticity as the primary driver: freezing sampling noise removes the deep solver’s short-horizon advantage, while scalar rescaling fails to recover the effect.Scalar rescaling also distorts mean trajectories.Source: §I · p. 1
  • Jointly adjusting solver depth and execution horizon yields training-free improvements in task success while amortizing inference latency.The paper frames sampling depth as an operational variance regulator rather than a fixed precision budget.Source: §I · pp. 1–2
  • The paper evaluates this phenomenon across LIBERO and RoboTwin 2.0 simulation benchmarks and real-world Franka manipulation on representative VLA backbones, identifying operational boundaries for compute–execution co-design.Source: §I · p. 2

II. RELATED WORK

Related work addresses action chunking, sampling acceleration, and compute scheduling, but typically fixes sampling depth or execution commitment while optimizing the other. This paper studies their interaction on frozen, unmodified backbones.

Action Chunking and Horizon Scheduling

  • Prior action-chunking strategies add dynamic replanning, trajectory scoring, staleness monitoring, variable-horizon training, or asynchronous execution, while generally holding numerical sampling depth fixed.Examples include bidirectional decoding, reactive diffusion, and VLA-Corrector [13].Source: §II · p. 2

Accelerated Sampling for Visuomotor Generation

  • Sampling-acceleration methods distill or truncate integration, stream partially denoised latents, straighten velocity fields, or adapt solver steps, generally to reduce latency.The paper instead studies how sampling depth interacts with execution horizon.Source: §II · p. 2

Joint Computational and Execution Scheduling

  • Unlike frameworks that use learned meta-controllers or variance monitors to allocate computational resources, this work shows that numerical truncation naturally regularizes trajectory variance on frozen, unmodified backbones, offering a training-free foundation for compute–execution co-design.Source: §II · p. 2

III. METHODOLOGY

The method generates action chunks by integrating a learned probability-flow velocity field backward from Gaussian noise to a conditional action distribution. Forward Euler sampling depth sets both the number of integration steps and the per-query function evaluations.

A. Flow-Matching Vision-Language-Action Policies

  • The policy conditions on synchronized camera streams, proprioceptive state, and language, and predicts an action chunk spanning horizon H and control dimension Da.Conditional flow matching transports a tractable noise distribution onto the demonstration manifold.Source: A. Flow-Matching… · p. 2
  • The method defines an optimal-transport interpolant between conditional target samples x0 and Gaussian source samples x1 as xτ = (1 − τ)x0 + τx1.The flow-time convention assigns τ=0 to the conditional target and τ=1 to standard Gaussian noise.Source: A. Flow-Matching… · p. 2
  • A neural network regresses the velocity field vθ of the probability-flow ODE, which guides the trajectory from noise toward the conditional action distribution.Source: A. Flow-Matching… · p. 2
  • Deployment starts from fresh Gaussian noise and integrates backward with N uniform forward Euler steps of size 1/N; the terminal state forms the emitted action chunk.Each velocity-field evaluation is a complete VLA-backbone forward pass, so N equals the number of function evaluations and per-query latency scales with N.Source: A. Flow-Matching… · p. 2

B. Receding-Horizon Action Chunking

Receding-horizon chunking balances feedback frequency against inference latency by executing an open-loop prefix before replanning. The figure connects that commitment choice to the effect of sampling depth on action dispersion and drift.

B. Receding-Horizon Action Chunking

  • Executing the entire predicted chunk open loop risks compounding tracking errors, while querying the VLA at every control step incurs prohibitive latency and motor chattering.Receding-horizon execution instead commits to a prefix of length s before replanning from a fresh observation.Source: B. Receding-Horizon Action Chunking · p. 2
  • The figure’s lower panel contrasts shallow and deep sampling across short and extended execution horizons: contracted dispersion hurts rapid replanning but suppresses compounding drift during extended commitment.For the illustrated settings, s=5 exploits sample diversity, whereas s=20 benefits from coarse integration, with up to 5.6× inference speedups stated in the caption.Source: Figure 2 · p. 3

C. Coupling Mechanism: Dispersion Contraction vs. Open-Loop Drift

Solver depth controls action-chunk dispersion, while execution horizon controls how long that dispersion compounds before feedback. Shallow integration contracts variance around an invariant centroid for N ≥2, helping extended open-loop execution but potentially hurting rapid replanning.

C. Coupling Mechanism: Dispersion Contraction vs. Open-Loop Drift

  • During reverse flow, coarse solvers capture the conditional centroid early while truncating the late-stage reexpansion of plausible trajectory branches.The learned velocity field initially contracts noise toward the conditional mean, then resolves modes near the end of the flow.Source: C. Coupling Mechanism: Dispersion… · p. 3
  • For N ≥1, Euler truncation preserves the action-chunk centroid and contracts covariance by ρ(N), with ρ(1)=0 and ρ(N) approaching 1 as N grows.For N ≥2, the asymptotic recovery rate is ρ(N)=1−c/N+O(N^-2).Source: C. Coupling Mechanism: Dispersion… · p. 3
  • The operational variance gain g(N)=√ρ(N) describes how injected noise scales in the emitted chunk.The paper predicts monotonic variance contraction with g(N) for N ≥2, while single-step Euler structurally diverges.Source: C. Coupling Mechanism: Dispersion… · pp. 3–4
  • Reducing integration depth reverses its operational effect with horizon: at s=5, coarse integration (N=2) penalizes frequent replanning versus deep integration (N=10); at s=20, it improves extended open-loop execution.Source: C. Coupling Mechanism: Dispersion… · p. 4
  • Rapid replanning can exploit diverse trajectories because feedback arrives before deviations compound; extended open-loop execution instead benefits from shallow sampling that suppresses drift.Each query redraws initial noise at s=5, whereas at s=20 one draw guides the robot for twenty steps.Source: C. Coupling Mechanism: Dispersion… · p. 4

D. Geometric Diagnostics of Action Clouds

The geometric diagnostics compare shallow and deep action clouds on fixed visual observations. They measure directional bias, empirical dispersion contraction, and departures from simple scalar variance change.

D. Geometric Diagnostics of Action Clouds

  • Centroid shift δµ measures directional bias between shallow N=2 and deep N=10 action-cloud means.The comparison uses matched initial Gaussian draws on fixed visual observations.Source: D. Geometric Diagnostics of Action… · p. 4
  • Cloud radius R(N) measures each action cloud's spread around its sample mean, while R(2)/R(10) measures empirical variance contraction.The clouds are generated with shallow N=2 and reference deep N=10 solvers.Source: D. Geometric Diagnostics of Action… · p. 4
  • Linear fit residual measures the shallow cloud variance unexplained by the best linear map of the deep cloud, diagnosing geometric changes beyond scalar contraction.Source: D. Geometric Diagnostics of Action… · p. 4

IV. EXPERIMENTS

The evaluation spans 55 simulation tasks and five physical robot manipulation environments.

IV. EXPERIMENTS

  • The horizon–depth coupling is evaluated across 55 simulation tasks and five physical robot manipulation environments.Source: §IV · p. 4

A. Simulation Experiments

Simulation experiments evaluate horizon–depth choices across VLA backbones and manipulation benchmarks. The results show a sign reversal across execution regimes and a balanced horizon–depth setting that reduces latency while maintaining or improving success.

A. Simulation Experiments

  • Across the LIBERO-Long horizon–depth grid, π0's optimal sampling depth shifts shallower as the execution horizon expands.Source: Figure 3 · p. 4
  • At s=20 on LIBERO-Long, π0 reaches 90.5% success with shallow N=2 sampling and a +6.4% interaction contrast over deep sampling.The short-horizon comparison at s=5 slightly favors deep sampling.Source: A. Simulation Experiments · pp. 4–5
  • The simulation benchmark covers π0, π0.5, and GR00T-N1.7 across LIBERO and RoboTwin 2.0, with different action-head limits shaping reachable horizons.π0.5 has a 10-action prediction head; GR00T-N1.7 supervises 16 actions.Source: A. Simulation Experiments · p. 4, Figure 3 · p. 4
  • On RoboTwin 2.0, π0.5's full-chunk shallow-sampling advantage is +1.6%, with a +8.5% interaction contrast across tasks.The interaction is positive on 37 of 50 tasks; at s=10, coarse sampling loses ground on 36 of 50 tasks.Source: A. Simulation Experiments · p. 5, Table II · p. 5
  • Moving from (10, 10) to (25, 2) cuts amortized step latency 3.5×; π0 rises from 37.0% to 45.7%, while π0.5 records 70.8% versus 73.8%.The paper identifies s=25 as an operational sweet spot for compute savings while maintaining or improving task completion.Source: A. Simulation Experiments · p. 5

B. Mechanistic Analysis

Mechanistic analyses show that shallow Euler sampling contracts action dispersion around nearly preserved trajectory centroids, while solver depth and execution horizon jointly shape how that dispersion affects performance. The benefit depends on regime: shallow sampling can stabilize long open-loop execution, whereas reactive control may require deeper integration and stochastic redraws.

Geometric Contraction

  • R(2)/R(10) = 0.51, with shallow sampling preserving centroid direction while contracting lateral dispersion across π0 and π0.5 backbones.For π0.5, radius ratios are 0.43–0.45; the shallow centroid projects onto the deep-cloud mean at 1.000, and centroid shifts remain small.Source: B. Mechanistic Analysis · p. 6
  • N=1 breaks the contraction regime: it retains 99% unexplained variance and trails N=10 by 26.1 ± 3.2 points at s=2.On LIBERO-Long, N=1 only crosses zero past s=35; at the trajectory level, N=10 produces a 2.1× wider lateral fan than N=2 while center paths remain aligned within 3% of total path length.Source: B. Mechanistic Analysis · p. 6

Causal Interventions

  • Endpoint-matched prior-noise rescaling fails to reproduce the sign reversal and nearly doubles centroid drift from 0.208 to 0.382.The intervention yields a negative interaction contrast (I = −1.5), supporting structured ODE contraction rather than arbitrary variance manipulation as the relevant mechanism.Source: B. Mechanistic Analysis · p. 6
  • Freezing noise redraws costs the N=10 solver 39.6 points at s=2 on π0.5, versus 11.3 points for N=2.At s=5 on π0, N=10 success falls from 81.2% to 42.3%, while N=2 declines by 5.6 percentage points.Source: B. Mechanistic Analysis · p. 6

Real-World Evaluation

  • Across four quasi-static Franka tasks, shallow sampling degrades success at s=10 but systematically outperforms deep sampling at s=H; moving-target tracking instead requires high feedback rates.The dynamic turntable task makes open-loop horizon the dominant failure mode.Source: Figure 5 · p. 7

Solver Dynamics

  • At π0.5 and s=10 with 2 NFE, Euler (95.1% ± 0.6%) and midpoint (94.9% ± 1.2%) similarly outperform stochastic sampling (88.3% ± 0.9%) and Heun (27.6% ± 2.8%).Intermediate-stage evaluations can extrapolate into out-of-distribution regions lacking training support, causing error compounding; Euler evaluates gradients along sequential on-path flow lines.Source: B. Mechanistic Analysis · p. 7

Task Geometry

  • Coarse transport succeeds better with extended commitment, reaching 82% versus 18% for placing both moka pots, while narrow-funnel insertions require rapid replanning.The supported heuristic is to use extended open-loop horizons for gross transport and frequent replanning for tight contact interactions.Source: B. Mechanistic Analysis · p. 7

C. Real-World Evaluation

Physical Franka evaluations across three VLA backbones reproduce the horizon-dependent sign reversal on four quasi-static tasks. Full-horizon commitment with shallow sampling improves success, while the moving-target task marks a boundary where open-loop execution fails regardless of solver depth.

Results

  • At s=H, shallow sampling gains +10.0% on π0, +6.9% on π0.5, and +11.3% on GR00T-N1.7 across the reported quasi-static task results.The gains include 90.0% on Cover Block for π0 and 97.5% on Lift Corn for π0.5; 2,400 hardware rollouts were conducted across the four configurations.Source: C. Real-World Evaluation · p. 7
  • At s=10, shallow sampling incurs a modest success drop across π0, π0.5, and GR00T-N1.7, whereas the four quasi-static tasks unanimously replicate the sign reversal.The evaluation used a 7-DoF Franka Research 3, controlled at 25 Hz, with 40 matched physical trials per configuration cell.Source: C. Real-World Evaluation · p. 7
  • On the 5 RPM turntable task, moving targets make full-horizon open-loop execution fail catastrophically regardless of solver depth.The target moves at approximately 8 cm/s, requiring high-frequency closed-loop feedback.Source: C. Real-World Evaluation · p. 8
  • On π0, co-designing s=50 and N=2 reduces compute burden from 1.0 to 0.04 NFE per step, a 25× reduction that eliminates edge GPU thermal throttling.Source: C. Real-World Evaluation · p. 8

V. CONCLUSION

The conclusion frames numerical sampling depth and execution horizon as coupled deployment choices in flow-matching VLA policies. Co-designing shallow sampling with extended commitment improves manipulation success while reducing inference cost.

Conclusion

  • Across 55 simulation tasks and 2,400 physical trials, co-designing shallow sampling with extended commitment yields up to 5.6× inference speedups alongside improved manipulation success.The conclusion presents early Euler truncation as an intrinsic variance regularizer that contracts stochastic dispersion while preserving trajectory centroids.Source: §V · p. 8

Related papers

Loading 2610.32102v1…