PaperlayerView exact revision on arXiv

Source-linked AI summary

SMILE: Smooth Motion for Improved Long-Horizon VLA Execution

Jongwoo Park, E-Ro Nguyen, Kanchana Ranasinghe, Cristina Mata, Xiang Li, Michael S Ryoo

arXiv:2608.29432v1cs.RO

TL;DR

Longer execution horizons reduce VLA inference cost but can degrade control because raw action chunks contain temporal inconsistencies. SMILE replaces raw per-step action generation with compact B-spline coefficient prediction while preserving each baseline’s conditioning and denoising backbones. Across heterogeneous VLA experts, simulations, and real-world manipulation, the paper reports improved accuracy–efficiency trade-offs and smoother long-horizon execution.

  • Problem

    Longer execution horizons can reduce amortized inference cost while lowering task reliability because raw action chunks accumulate oscillations, inconsistent directions, and outliers.

  • Method

    SMILE predicts compact B-spline coefficients for robot controls and deterministically decodes them into smooth action sequences, changing the action interface while retaining the original VLA architecture.

  • Results

    SMILE reports consistent accuracy–efficiency gains across SmolVLA, Evo1, VPP, and DAWN on LIBERO, CALVIN, and real-world xArm manipulation.

  • Takeaways & Limitations

    Smooth coefficient-space generation is presented as a general mechanism for improving compact VLA accuracy–efficiency trade-offs without replacing original visual-language conditioning or denoising backbones.

Abstract

from arXiv · show

Vision-Language-Action (VLA) models reduce inference cost by executing multiple actions per call, but longer horizons often degrade accuracy because raw chunks contain jitter and outliers. We introduce SMILE, an architecture-preserving interface that predicts B-spline coefficients and decodes them into smooth action sequences. SMILE changes only the action representation, enabling longer fixed horizons while retaining each baseline's backbone and model scale. We apply SMILE to SmolVLA, Evo1, VPP, and DAWN, improving accuracy and amortized inference efficiency across LIBERO, CALVIN, and real-world experiments. SMILE-Evo1 reaches 98.0% with a 1.1x speedup on LIBERO, while SMILE-VPP reaches an average length of 4.42 with a 1.5x speedup on CALVIN. At a matched execution horizon of 10, SMILE-SmolVLA reduces non-boundary acceleration by 78.6% and velocity sign-change rate by 42.3%. Real-world xArm tests show higher success, fewer drops, and fewer contacts. These results establish smooth coefficient-space generation as a route to accurate, efficient long-horizon VLA execution. Project page: jongwoopark7978.github.io/smilevla

I. INTRODUCTION

SMILE addresses the accuracy loss of long-horizon VLA execution by replacing raw action generation with smooth B-spline coefficient prediction while preserving the main architecture. Across heterogeneous models and benchmarks, it targets a better accuracy–efficiency trade-off at comparable model scale.

  • Motivation: Larger execution horizons amortize expensive model calls over more robot actions, but raw chunks can accumulate oscillations, inconsistent directions, and outliers.These effects can reduce task success as more actions are executed before replanning.
  • Method: SMILE predicts compact B-spline coefficients for translation, rotation, and gripper control, then decodes them into temporally coherent execution actions.Smoothness is imposed through the action representation rather than a larger backbone or separate planner.
  • Architecture-preserving integration: SMILE preserves the original visual-language conditioner and main denoising backbone, modifying only the action-side interface across heterogeneous VLA action experts.The evaluated experts include SmolVLA, Evo1, VPP, and DAWN, while upstream predictive-visual or pixel-motion modules remain unchanged.
  • Evaluation: SMILE is evaluated across LIBERO, CALVIN, and real-world xArm manipulation, with reported gains in accuracy, efficiency, and action smoothness.The paper relates these gains to longer fixed horizons and lower intra-chunk acceleration and oscillatory action changes.
  • Related work: SMILE reparameterizes action generation rather than changing chunk selection or execution, distinguishing it from related action-chunking approaches.The related-work comparison describes BID, AAC, and AutoHorizon as modifying chunk selection or execution without changing the learned action representation.

III. OBSERVATIONS ON LONG-HORIZON EXECUTION

Increasing the execution horizon alone improves amortized efficiency but makes raw-action execution less reliable. SMILE instead improves the temporal structure of action chunks through coupled B-spline coefficient generation.

  • Observation: Increasing H amortizes each model call over more controller steps but commits the controller to a longer locally open-loop prefix before processing the next observation.The longer prefix can contain temporally inconsistent individual actions.
  • Observation: Raw-action baselines lose task accuracy as execution horizons increase, showing that longer execution cannot be obtained reliably by increasing H alone.The observation motivates changing the temporal structure of predicted chunks rather than only increasing their executed length.
  • SMILE response: SMILE operates with longer fixed horizons in cluttered xArm experiments while reducing object drops and contacts with neighboring objects.The passage associates these lower rates with more stable gripper behavior and more controlled approach and transport motions.
  • SMILE response: SMILE addresses the limitation by denoising compact B-spline coefficients whose decoded actions are coupled across time.This changes the temporal structure of the action chunk instead of merely increasing its execution horizon.

IV. METHOD

The method replaces raw action-chunk prediction with compact B-spline coefficient prediction and deterministic decoding, while retaining the baseline’s major computational components.

  • Baseline interface: At each model refresh, a standard action expert predicts a raw action chunk Aj containing F low-level actions.The controller executes only the first H actions before processing the next observation, with H ≤ F.
  • Baseline interface: Increasing H reduces amortized inference latency per action but extends the locally open-loop commitment before replanning.The trade-off arises because individual actions in the longer prefix may be temporally inconsistent.
  • SMILE interface: SMILE replaces the raw chunk with a compact macro-action consisting of B-spline coefficients for translation, rotation, and gripper control.A deterministic decoder maps the coefficients to an F-step smooth action sequence compatible with the original controller.
  • SMILE interface: Within each modality, spline coefficients jointly determine neighboring decoded actions, reducing sensitivity to isolated predictions while keeping H fixed within each rollout.The controller still executes the first H actions of the decoded sequence.
  • Architecture: SMILE retains the baseline VLA backbone, conditioning pathway, and main denoising backbone, changing only action-side projections and coefficient heads.Visual-language, predictive-visual, or pixel-motion features are computed as in the original model.

B. B-spline action representation and decoding

SMILE represents translation, rotation, and gripper trajectories with B-spline coefficients, then deterministically decodes them into temporally coherent native controller actions.

  • B-spline action representation: A fixed B-spline basis maps compact coefficients into translation, rotation, and gripper trajectories over uniformly spaced action times.The basis uses F decoded actions, spline degree P, and K control points.
  • Trajectory decoding: The gripper trajectory is clipped to benchmark- or controller-specific lower and upper command bounds.These bounds are denoted by gmin and gmax.
  • Trajectory decoding: Translation and rotation are decoded as cumulative paths and converted into native actions through finite differences and adjacent relative rotations.The virtual initial cumulative translation and rotation states are both set to zero.
  • Controller compatibility: Decoded actions preserve the baseline’s relative-action semantics for translation and rotation, while gripper commands remain absolute.At inference, the controller executes the first H actions from a decoded chunk of length F.

C. Architecture-preserving denoising objective

SMILE preserves the baseline denoiser and conditioning while changing its denoising variable from raw actions to normalized spline coefficients.

  • Coefficient-space denoising: SMILE changes the denoised variable from a raw action sequence to normalized spline coefficients for translation, rotation, and gripper modalities.The target coefficient tensor is modality-specific, while the baseline multimodal conditioning remains unchanged.
  • Preserved training formulation: The action expert receives noisy coefficient tokens and predicts the target used by the original training formulation.That target is a velocity field for flow-matching experts or noise for diffusion experts.
  • Training objective: SMILE uses equal coefficient-denoising weights for translation, rotation, and gripper control, with λtr = λrot = λgrip = 1.The weighting is shared across all SMILE variants.

D. Coefficient-target construction

SMILE constructs coefficient targets by transforming baseline trajectories and projecting them onto a fixed B-spline basis, then decodes predicted coefficients into coupled controller actions.

  • Target preparation: Coefficient targets use the same expert trajectories as each baseline, collecting the next F native actions at every eligible start step.Translation and rotation become cumulative command-space trajectories, while gripper commands remain absolute.
  • B-spline projection: For translation, rotation, and gripper modalities, target coefficients are obtained by least-squares projection onto the fixed B-spline basis.The projected target is the corresponding cumulative or absolute trajectory.
  • Normalization and masking: Dataset-level coefficient statistics normalize the targets, and validity masks handle samples near episode boundaries.These masks are applied during coefficient-target construction.
  • Inference decoding: At inference, denoised coefficients are decoded through the same basis to generate temporally coupled native controller actions.This replaces raw per-step action denoising with coefficient-space denoising.

V. EXPERIMENTS

Experiments evaluate SMILE across heterogeneous action experts and environments, measuring accuracy, horizon, latency, model scale, and smoothness.

  • Experimental scope: SMILE is evaluated across four heterogeneous action experts on LIBERO, CALVIN, and real-world xArm environments.The study measures task accuracy, execution horizon, amortized inference latency, and action smoothness.
  • LIBERO results: SMILE-Evo1 raises LIBERO overall success from 92.7% to 98.0% while extending the horizon from 14 to 16 and reducing latency from 8.5 to 8.0 ms per action.Both integrations retain the reported model-size scales of their corresponding baselines.
  • LIBERO results: SMILE-SmolVLA raises LIBERO success from 74.0% to 82.6% while extending the horizon from 1 to 10 and reducing latency from 243.9 to 24.2 ms per action.This corresponds to a 10.1× speedup.
  • LIBERO suite results: SMILE-SmolVLA improves all four LIBERO suites, with its largest gain on LIBERO-Long from 42.8% to 63.6%.SMILE-Evo1 also improves Spatial, Object, and Long, reaching 98.2% on LIBERO-Long.

B. CALVIN Benchmark

On CALVIN and real-world xArm tasks, SMILE supports longer execution horizons while improving efficiency, sequence completion, and manipulation reliability across action experts.

  • CALVIN Benchmark: SMILE-VPP increases average sequence length from 4.33 to 4.42 while reducing amortized latency from 19.1 to 13.0 ms per action.Both SMILE variants increase the execution horizon from 10 to 15.
  • CALVIN Benchmark: SMILE-VPP reaches an average sequence length of 4.42 versus DreamVLA’s 4.44, while reducing latency from 193.2 to 13.0 ms per action.This corresponds to a reported 14.9× speedup relative to DreamVLA.
  • Real-world Experiment: In plain xArm settings, average success rises from 26.7% to 78.3% for Evo1 and from 82.5% to 98.3% for VPP.The evaluation covers six object categories under plain and cluttered arrangements.
  • Real-world Experiment: In clutter, average success rises from 8.3% to 56.7% for Evo1 and from 70.8% to 88.3% for VPP.The cluttered setting tests approach, grasping, and transport around nearby distractors.
  • Real-world Experiment: SMILE reduces clutter hit rate by 22.5 percentage points for Evo1 and 13.3 percentage points for VPP.Wrong-object pickup rate also falls by 23.4 and 0.8 percentage points, respectively.
  • Real-world Experiment: SMILE increases the executed horizon from 6 to 16 for Evo1 and from 10 to 15 for VPP, yielding 2.7× and 1.5× speedups.The reported latency changes are 21.3→8.0 ms for Evo1 and 19.1→13.0 ms for VPP.

D. Action Smoothness Analysis

At the matched horizon H = 10, SMILE-SmolVLA substantially reduces intra-chunk acceleration and oscillatory motion, supporting smoother long-horizon execution as its mechanism.

  • Action Smoothness Analysis: Vel. ZCR measures the fraction of adjacent velocity steps that reverse sign, providing a measure of oscillatory motion.Acceleration is computed using finite differences of normalized translation and rotation actions.
  • Action Smoothness Analysis: 78.6% lower non-boundary acceleration and 42.3% lower Vel. ZCR are observed for SMILE-SmolVLA at H = 10.Global and boundary acceleration decrease by 18.6% and 20.5%, respectively.
  • Action Smoothness Analysis: The larger within-chunk reduction indicates that coefficient-space generation primarily suppresses intra-chunk jitter rather than only smoothing model-call transitions.The comparison matches execution horizon and model-refresh frequency between SmolVLA and SMILE-SmolVLA.
  • Action Smoothness Analysis: SMILE-SmolVLA maintains a lower acceleration baseline and suppresses high-frequency fluctuations across both illustrated task types.The traces show frequent high-amplitude spikes for SmolVLA within chunks and near model-call boundaries.

E. Ablation Study

Across multiple long execution horizons, SMILE variants consistently outperform raw-action baselines, demonstrating robustness of coefficient-space generation to horizon choice.

  • Ablation Study: SMILE-SmolVLA and SMILE-Evo1 consistently outperform their corresponding raw-action baselines across multiple long execution horizons.The ablation section evaluates robustness across horizons and identifies B-spline hyperparameters.
  • Ablation Study: Fig. 9 compares LIBERO success against execution horizon for SMILE variants and raw-action baselines.The reported comparison spans the tested horizons rather than a single fixed horizon.
  • Ablation Study: The selected spline configuration uses P = 3 control points and spline degree K = 8.These values are identified in the ablation materials as selected hyperparameters.

VI. CONCLUSION

SMILE replaces raw per-step action generation with compact B-spline coefficient prediction, enabling smoother, longer-horizon execution across diverse VLA experts and environments.

  • VI. CONCLUSION: SMILE is an architecture-preserving B-spline interface that predicts compact coefficients instead of raw per-step actions.It retains the original visual-language conditioning and denoising backbones.
  • VI. CONCLUSION: SMILE improves the accuracy–efficiency trade-off through temporally consistent action chunks, longer fixed horizons, and lower amortized inference latency per action.The conclusion reports consistent improvements across heterogeneous experts, simulation benchmarks, and real-world manipulation.
Loading 2608.29432v1…