PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yanyan Zhang, Chaoda Song, Vikash Singh, Xinpeng Li, Kai Ye, Zhe Hu, Zhongzhu Pu, Yu Yin, Vipin Chaudhary

arXiv:2605.11459v2cs.ROcs.AIcs.CVcs.LG

TL;DR

Chunked-action VLA models are structurally blind to temporal dynamics, limiting manipulation in changing environments. The paper introduces a training-free Pace-and-Path Correction wrapper and reports success-rate gains of up to 28.8% and 25.9% in dynamic-only and mixed environments.

  • Problem

    Most chunked-action VLAs generate open-loop actions from a single frame, leaving them blind to scene changes and limiting precise characterization of dynamics-specific failures.

  • Method

    Pace-and-Path Correction is a backbone-agnostic, training-free wrapper that uses closed-form orthogonal pace compression and path offsets to compensate for dynamics within action chunks.

  • Results

    Success rates improve by up to 28.8% in dynamic-only environments and 25.9% in mixed environments, while PPC outperforms training-free wrappers and dynamic-adaptive methods.

  • Takeaways & Limitations

    PPC provides a deployable way to improve foundational VLA performance under diverse motion regimes without retraining or architectural modification.

  • Takeaways & Limitations

    The closed-form correction assumes approximately constant disturbance velocity and direction within each chunk, becoming suboptimal when motion changes rapidly.

Abstract

from arXiv · show

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 finetuned 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, a training-free, closed-form inference-time operator that wraps any chunked-action VLA. From a single quadratic cost, joint minimization yields a unified solution that decomposes orthogonally into two distinct 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. Empirical results demonstrate that our framework consistently outperforms state-of-the-art training-free wrappers and dynamic-adaptive methods and improves success rates by up to 28.8% and 25.9% in absolute terms over foundational VLA models in dynamic-only and static-dynamic mixed environments, respectively.

1 Introduction

Most VLAs generate open-loop action chunks from single-frame observations, making them structurally blind to dynamics during execution. The paper proposes PPC, a training-free inference wrapper, and evaluates it on MOVEBENCH, where it improves success rates by up to 28.8% and 25.9%.

  • Motivation: Most current VLAs predict fixed-length future-action chunks from a single visual frame and execute them open-loop, leaving policies structurally blind to dynamics.Each chunk is generated from an initial static snapshot without object-motion supervision and executed before the next chunk is generated.
  • Related work: Existing motion-aware approaches inject temporal cues through optical flow, visual prompting, memory banks, or motion predictors, but require retraining and backbone-specific changes.Extraction latency and forecasting hallucinations further limit reliability during dynamic interaction.
  • Method: Pace-and-Path Correction (PPC) is a closed-form, training-free, inference-time wrapper that uses external velocity signals to correct dynamics inside action chunks.It requires no learnable parameters, backbone modification, or backbone specification.
  • Evaluation: MOVEBENCH systematically isolates and evaluates VLA performance across diverse motion patterns and speeds.The benchmark is dedicated to evaluating performance when motion is the controlled variable.
  • Results: 28.8% and 25.9%: PPC improves success rates by up to these absolute amounts in dynamic-only and mixed environments, respectively.PPC outperforms state-of-the-art training-free wrappers and dynamic-adaptive methods across motion families and foundational VLAs.

2 Related Work

Related work spans VLA policy architectures, predominantly static manipulation benchmarks, and dynamics-aware remedies. Existing remedies either add temporal or predictive cues through retraining and architecture-specific integration, or reduce inference latency while retaining single-frame observations.

  • Vision-Language-Action Models: VLAs map multimodal observations and language to action sequences using pretrained vision-language backbones, with newer policies emitting continuous actions through diffusion- or flow-matching experts.Early autoregressive designs instead decode discrete text tokens, which limits action resolution and incurs token-by-token decoding costs.
  • Dynamic Manipulation Benchmarks: Most VLA benchmarks use static settings, while DOM and DOMINO recently introduced moving-object evaluations.LIBERO, CALVIN, ManiSkill, RoboCasa, and VLABench measure planning, language grounding, or skill transfer with stationary objects.
  • Dynamics-Aware Vision-Language-Action Models: Dynamics-aware methods inject temporal or predictive cues through optical or scene flow, visual traces, episodic memory, or future-state forecasting.Examples include FlowVLA, PUMA, LaMP, TraceVLA, MemoryVLA, DreamVLA, WorldVLA, 4D-VLA, FUTURE-VLA, and SC-VLA; these approaches require retraining and architecture-specific integration.
  • Dynamics-Aware Vision-Language-Action Models: Other approaches reduce inference latency while retaining the single-frame paradigm through smaller backbones, parallel decoding, tokenization compression, token caching, discrete diffusion, or asynchronous inference.Representative methods include DynamicVLA, PD-VLA, and FASTer.

3 Methodology

The method wraps a chunked VLA with a training-free, closed-form correction that separates temporal pace adjustment from orthogonal spatial path correction. A stateful instability stabilizer supplements the exact quasi-stationary solution by detecting sustained direction shifts and limiting execution under irregular motion.

  • Chunked VLA formulation: The VLA executes the first K ≤ H actions of a chunk open-loop before re-querying, motivating inference-time correction over the executed window.Each action combines an end-effector displacement with rotation and gripper commands.
  • Correction operator: The corrected action at step k is uk = α∆p + δk, combining temporal compression α ≥ 1 with per-step spatial offsets δk.The formulation introduces a residual disturbance and cumulative spatial offset to express waypoint-tracking error over the chunk.
  • Pace-and-path decomposition: The convex quadratic has a closed-form minimizer whose pace and path channels decompose orthogonally, allowing sequential derivation without loss of optimality.The pace channel absorbs the disturbance component aligned with the planned direction, while the residual lies perpendicular to it.
  • Path channel: The path profile distributes perpendicular displacement gradually across the executed window, with δK = 0 ensuring that the next chunk starts unbiased.For large K, the first-order perpendicular endpoint contribution approaches approximately 0.618 v d̂⊥.
  • Closed-form dynamics correction: The corrected delta is fully determined by chunk geometry and the dynamics signal, with no learnable parameter.Under second-order disturbance, the path recurrence adds a Lucas-polynomial branch to the Fibonacci first-order branch.
  • Irregular-regime stabilization: The stabilizer triggers on sustained direction instability using a direction-shift threshold and cascaded outer and inner exponential moving averages.When chronic instability drives the sticky state toward one, the latch caps executed chunk length under sustained irregularity.

4 Experiments

Experiments on MOVEBENCH show that PPC consistently improves foundational VLAs under dynamic motion while preserving static performance. Gains are largest for accelerated motion and remain robust across baselines, motion difficulty, and perception noise.

  • Benchmark: MOVEBENCH isolates target motion as the controlled variable in an xArm6 pick task spanning uniform, accelerated, irregular, and static regimes.The benchmark is built on ManiSkill with the SAPIEN engine and varied object shapes.
  • Main results: +16.4 to +28.8 absolute points: PPC raises the dynamic-only average of all four foundational VLAs, with π0.5+PPC reaching 72.9% dynamically and 73.6% overall.PPC preserves static capability because it becomes the identity when v = 0.
  • Motion regimes: +32.8: PPC’s largest per-family improvement occurs on accelerated motion, followed by uniform (+18.2) and irregular (+12.6).The accelerated-motion advantage is attributed to the δ channel absorbing perpendicular residuals accumulated under sustained acceleration.
  • Ablations: −28.3 points: removing α causes the largest ablation collapse below full PPC’s 66.1% overall success, while removing δ costs 24.2 points overall.The α loss is broadly distributed across motion families, whereas δ’s loss concentrates on accelerated motion.
  • Baseline comparison: 44.9%: DynamicVLA’s low-latency policy underperforms π0.5 at 56.9%, showing that replanning frequency does not resolve intra-chunk motion blindness.DynamicVLA also scores 70% statically versus 81% for its SmolVLA backbone.
  • Motion difficulty: +38.5: PPC’s gain reaches this value at the hardest uniform-speed tier and remains around +30 across the acceleration range.The gain grows monotonically with target speed, indicating that intra-chunk compensation scales with dynamic difficulty.

5 Conclusion · A Full Closed-Form Mathematical Derivation · A.4 Derivation of α⋆(Pace Channel)

Pace-and-Path Correction is a closed-form, training-free inference-time wrapper that compensates for dynamics in chunked-action VLAs without retraining or architectural changes. Its derivation uses a single quadratic cost to separate pace correction along the planned direction from orthogonal path correction, with the pace solution governed by the disturbance’s directional projection.

  • 5 Conclusion: PPC is backbone-agnostic, introduces no learnable parameters, and can wrap any released VLA without retraining or architectural modification.The paper also introduces MOVEBENCH to isolate motion regime as the sole evaluation axis for chunked-VLA behavior.
  • A Full Closed-Form Mathematical Derivation: The appendix derives the complete Pace-and-Path Correction operator from the single quadratic cost, adding no assumptions beyond A1–A3.The assumptions are quasi-stationary plans, slowly varying disturbances, and small rotations; violations reduce optimum accuracy without destabilizing the operator because chunk resets refresh signals.
  • A Full Closed-Form Mathematical Derivation: The objective penalizes only per-step spatial offsets δ_k, not α, because α compresses execution along the planned direction while δ_k creates off-plan deviation.This asymmetric penalty satisfies uniqueness, baseline degeneration at v = 0, and the cosine-projection structure in α⋆.
  • A.4 Derivation of α⋆(Pace Channel): At the optimum, rotational invariance forces every δ_k to align with the residual disturbance A, reducing the pace stationarity condition to a scalar projection equation.The condition is (α −1)∥∆p∥² = v( d̂ · ∆p), which yields the pace solution from the disturbance component parallel to the planned direction.
  • A.4 Derivation of α⋆(Pace Channel): The residual A⋆ = v d̂ − v cos θ ĉ_∆p equals the perpendicular disturbance v d̂⊥ and lies entirely orthogonal to the planned direction.Thus, the pace channel absorbs the parallel component while the path channel receives the perpendicular residual.
  • A.4 Derivation of α⋆(Pace Channel): When cos θ < 0, α⋆ is clamped to 1, sending the full disturbance to the path channel; when α⋆ > T/K, K_exec absorbs the overflow.The dynamic horizon shrinks from T toward K as α increases and replaces K in the Fibonacci profile normalization.
  • A Full Closed-Form Mathematical Derivation: Under affine acceleration, the pace solution gains a coupling term that scales linearly in K, and setting a = 0 recovers the first-order formula.The corresponding perpendicular acceleration component contributes an additive Lucas-profile correction to the path channel.

A.5 Derivation of δ⋆ k (Path Channel, Fibonacci Profile) · B MOVEBENCH Details · B.1 Object Pool

The path channel derives a Fibonacci-profile correction through a 2D recurrence, eigenstructure, particular solution, and boundary conditions. MOVEBENCH specifies a controlled simulation protocol and a heterogeneous nine-object YCB pool for grasp-and-lift evaluation.

  • A.5 Derivation of δ⋆ k (Path Channel, Fibonacci Profile): The path channel addresses A⋆, which α⋆ cannot absorb, by differencing condition (4) and applying σj+1 − σj = δj to obtain a 2D linear recurrence.
  • A.5 Derivation of δ⋆ k (Path Channel, Fibonacci Profile): The recurrence’s companion matrix has characteristic polynomial λ2 − 3λ + 1 = 0, producing eigenvalues tied to the golden ratio.
  • A.5 Derivation of δ⋆ k (Path Channel, Fibonacci Profile): The constant particular solution is (δp, ep) = (A⋆, 0), while the homogeneous solution is formed from eigenvectors w+ = (φ, 1)⊤ and w− = (1, −φ)⊤.
  • A.5 Derivation of δ⋆ k (Path Channel, Fibonacci Profile): Boundary conditions e0 = 0 and δK = 0 determine the homogeneous coefficients and enforce closure at the chunk boundary.
  • A.5 Derivation of δ⋆ k (Path Channel, Fibonacci Profile): The resulting profile uses Fibonacci ratios F2k+1/F2K+1, which increase monotonically from approximately 0 to φ−2 ≈ 0.382 as K → ∞.
  • B MOVEBENCH Details: MOVEBENCH runs at 20 Hz in ManiSkill’s SAPIEN backend with 512×512 RGB from overhead and wrist cameras, 7-DoF proprioception, and language instructions.
  • B MOVEBENCH Details: Each episode allows at most 200 environment steps over 10 s wall-clock and succeeds when the target is grasped and lifted by ≥3 cm.
  • B.1 Object Pool: MOVEBENCH samples targets uniformly from nine YCB household objects spanning heterogeneous primitives with dominant-axis sizes of 25–103 mm, all reachable by the xArm6 gripper.

B.2 Motion Regimes … C.1 Cross-Backbone Comparison of Inference-Time Wrappers

The benchmark isolates object motion through deterministic per-episode update rules while holding robot, workspace, observations, and success criteria fixed. Supplementary analyses expose PPC diagnostics and compare inference-time wrappers across π0.5 and SmolVLA.

  • B.2 Motion Regimes: Seven dynamic environments instantiate three motion families, each governed by a deterministic update rule sampled once per episode and applied at 20 Hz.The families are uniform translation, accelerated motion, and irregular motion.
  • B.2 Motion Regimes: Uniform translation uses a fixed planar direction and tiered constant speeds of [2], [2] [4], and [4] [8] cm/s.No within-episode randomization occurs, so each regime is characterized by its initial speed magnitude.
  • B.2 Motion Regimes: Accelerated motion fixes the base speed at v0 ∈[2, 3] cm/s while tiering acceleration magnitudes across [2] [3], [3] [5], and [5] cm/s2.This decoupling attributes cross-tier gaps to the second-order signal rather than initial-speed shifts.
  • B.2 Motion Regimes: Irregular motion probes non-stationarity through Random Walk direction changes, Stop-and-Go pauses, and Teleport behavior within chunk-scale timescales.The passage specifies 5 cm/s Random Walk motion with 5–12-tick direction resampling and 7 cm/s Stop-and-Go motion alternating with pauses.
  • B.3 Episode and Workspace Configuration: The shared setup uses a 7-DoF xArm6, planar tabletop, two RGB cameras at 20 Hz, and observations combining image pairs with 7-D proprioception.The cameras are static overhead and wrist-mounted views.
  • B.3 Episode and Workspace Configuration: Success requires contacting and lifting the target at least 3 cm within 200 steps, while lift-only and approach-only events count as failures.The episode cap corresponds to 10 s wall-clock.
  • B.4 Demonstration Generation: Oracle demonstrations use full simulator state to generate smooth 20 Hz end-effector trajectories in the learned policy’s 6-D delta plus gripper action space.Episodes whose oracle rollout fails the lift criterion are discarded.
  • B.5 Diagnostic Statistics: PPC runs release per-episode α, Kexec, vt, ˆdt trajectories, chunk-boundary timing, re-inference cadence, and minimum gripper–object distance for diagnostics.These signals support Section 4.5 analysis but are not used in headline numbers.

C.2 Additional Analysis on Experimental Results

Additional analyses show that PPC benefits weakest dynamic backbones most, while competing wrappers can fail through stale actions or unnecessary re-inference. Its gains vary by motion regime, with a defined lower operating boundary on very slow motion.

  • Per-backbone gain decomposition: +28.8 dynamic-only is PPC’s largest absolute gain, achieved on GR00T N1.6, followed by π0 (+21.1), π0.5 (+18.6), and SmolVLA (+16.4).Gain ordering inversely correlates with baseline dynamic performance: GR00T N1.6 combines an 88% static score with a 37.3% dynamic score.
  • Inference-time wrappers on SmolVLA: +16.4 points is achieved by PPC on SmolVLA, whereas ACT and BID degrade dynamic-only success rate by −7.0 and −0.8 points, respectively.PPC defaults to the baseline when v = 0.
  • ACT’s teleport collapse: 1% is ACT’s Teleport score on π0.5, a 59-point drop from the bare backbone’s 60%.Temporal ensembling averages overlapping chunks, so stale pre-teleport actions drag the end-effector away from the new target; longer buffers prolong the stale signal.
  • DynamicVLA’s static regression: 70% is DynamicVLA’s static-environment score, 11 points below SmolVLA’s 81% backbone score.Re-inference every 2 env-steps injects inter-chunk discontinuities and micro-jitter into otherwise stable 200-step trajectories.
  • Failure mode: SmolVLA+PPC on Uniform-Easy: −7% is SmolVLA+PPC’s result on Uniform-Easy, where 1–2 cm/s object speeds approach the velocity estimator’s noise floor.Sign noise in α⋆ oscillations near 1.0 can trigger unnecessary compression, marking the lower boundary of PPC’s effective operating range.

D Supplementary Experimental Details

The appendix specifies the hardware, deployment settings, PPC configuration, implementation details, and evaluation protocol used for MOVEBENCH experiments. Evaluations use standardized rollout budgets and define success by grasping and lifting the object at least 30 mm.

  • Hardware: Foundational VLAs are fine-tuned on one NVIDIA H200, while all evaluations run on one NVIDIA RTX A6000.Fine-tuning takes at most 48 H200-hours; evaluating each method over 1,000 trials takes approximately 2 A6000-hours.
  • Comparison Baseline Deployment: Comparison wrappers ACT and BID run on the released π0.5 checkpoint with chunk-boundary execution horizons capped at 10 env-steps.ACT uses exec_horizon = 1, whereas BID uses exec_horizon = 10; DynamicVLA uses its released SmolVLA-based MOVEBENCH checkpoint.
  • PPC Configuration: PPC uses fixed wrapper hyperparameters across all reported runs, including T = 16, K = 2, Heff = 10, βin = 0.3, and Vmax = 1 m/s.For native chunks longer than 16, only the first 16 model-steps enter the wrapper; the velocity bound matches the xArm6 published maximum TCP velocity.
  • Implementation Specifics: The dynamics signal is computed from one-step finite differences of simulator ground-truth object positions, capped at Vmax, rather than the simulator’s velocity field.The position finite difference drives α⋆ and δk, while the reported velocity field is used only by the simulation-consistency gate.
  • Evaluation Protocol: Each method is evaluated with 100 trials in each of 10 MOVEBENCH environments, totaling 1,000 trials per method.Trials use seeds 0–99, run for at most 200 env-steps, and count success after grasping and lifting the object at least 30 mm above its initial height.

E Supplementary Ablation Studies

Supplementary ablations show that PPC adds negligible computation, adapts compression to motion difficulty, selectively bypasses inconsistent simulator dynamics, and remains sensitive to the physical velocity bound. These studies support real-time use while characterizing when the wrapper engages and how physical-specification choices affect performance.

  • Wrapper Computational Overhead: 0.07 ms mean latency and under 0.12 ms P99 latency show the closed-form wrapper adds < 0.2% overhead to typical VLA inference.Measured over 42 chunk-resets on MOVEBENCH-MovingMedium using a single CPU thread; GR00T N1.6 chunk inference is approximately 64 ms.
  • Adaptive α⋆ Engagement: α⋆ remains near 1 at the median, while its P90 increases with difficulty and reaches the cap T/K = 8 on AccelHard.The compression factor is bounded by 8 and degenerates to 1 without environmental motion.
  • Adaptive α⋆ Engagement: 10.5% of AccelHard chunks reach α⋆ = 8, whereas α⋆ ≡ 1 throughout Teleporting because the νt gate clamps the wrapper to identity on A1-violating chunks.This indicates aggressive compression is concentrated in difficult acceleration regimes, while Teleporting is bypassed under the specified contradiction condition.
  • Sim-Consistency Bypass Rate: 82.7% of Teleporting chunks activate the νt gate, compared with 0% in all other regimes.The gate detects simulator-reported zero velocity alongside non-zero observed displacement and acts as a regime-specific bypass.
  • Vmax Robustness: Doubling Vmax from 1 m/s to 2 m/s reduces AccelHard success rate from 33% to 27%, indicating that exceeding the hardware ceiling causes overshoot.The study therefore treats Vmax as a physical specification rather than a hyperparameter.

F Qualitative Visualization · G Limitations and Broader Impact

The qualitative visualizations show that PPC redirects end-effector trajectories toward moving objects, while its adaptive pace responses differ across uniform, accelerated, and irregular motion. No passage describing limitations or broader impact is supplied.

  • F Qualitative Visualization: 30–100 mm: baseline TCP trajectories terminate short of the moving object without grasping after veering toward the static target’s expected position.
  • F Qualitative Visualization: PPC-equipped TCP trajectories terminate at grasp while sharing the baseline’s backbone and random seed.
  • G Limitations and Broader Impact: The supplied passages contain no description of limitations or broader impact.
  • F Qualitative Visualization: PPC redirects chunk-interior TCP paths to track moving targets across all four motion regimes, unlike the bare baseline trajectories.Figure 8 compares identical seeds and backbones across representative dynamic environments.
  • F Qualitative Visualization: Uniform motion produces α⋆ profiles that remain approximately 1, reflecting small and roughly constant observed velocity.
  • F Qualitative Visualization: Accelerated motion produces a monotonically rising α⋆ profile as the object’s velocity accumulates.
  • F Qualitative Visualization: Irregular motion produces transient-spiking α⋆ profiles, distinguishing its wrapper response from uniform and accelerated regimes.

G.1 Limitations · G.2 Broader Impact

PPC has important limitations involving disturbance assumptions, simulation-only validation, task and object scope, and reliance on external tracking. Its broader impact includes safer dynamic manipulation potential, alongside unresolved safety risks for human-facing deployment.

  • G.1 Limitations: Rapidly changing disturbances within an executed chunk violate PPC’s quasi-stationarity assumption and make its correction suboptimal; teleportation provides an extreme violation.The hierarchical latch mitigates chronic violations by capping Kexec, but cannot fully correct rapid within-chunk changes.
  • G.1 Limitations: PPC’s wrapper and velocity-estimation pipeline have been evaluated only in ManiSkill simulation, leaving sim-to-real effects from depth noise, occlusion, and tracker latency unvalidated.These effects may degrade external velocity estimates beyond the regime where the Bayesian confidence gate compensates.
  • G.1 Limitations: MOVEBENCH uses a single pick-task family, leaving transfer of PPC’s gains to place, push, and multi-step assembly unresolved.The controlled design isolates motion regime, but chunk-internal dynamics may interact differently with task semantics elsewhere.
  • G.1 Limitations: PPC depends on reliable external tracking or depth sensing for velocity, and degenerates to the baseline VLA when clutter or occlusion causes object-tracking failure.Using an external signal avoids the ego-motion confound but introduces this perception dependency.
  • G.1 Limitations: The formulation tracks one target object per chunk; independently moving multi-object scenes require target selection or a multi-channel cost-function extension not addressed here.The current work does not provide either mechanism.
  • G.2 Broader Impact: PPC may enable safer, more capable robot manipulation in dynamic manufacturing, household-assistance, and human-robot-collaboration settings without retraining or additional data collection.As a training-free wrapper with no learnable parameters, PPC does not introduce new data privacy concerns.
  • G.2 Broader Impact: More capable dynamic manipulation could lower barriers to autonomous systems near humans, but PPC remains simulation-only and is not safety-validated for real-world human-facing deployment.Deployment should include safety mechanisms independent of the policy layer.
Loading 2605.11459v2…