SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies

2026-07-05

AI Summary

This work addresses the issue of trajectory incompatibility and motion discontinuity at action chunk boundaries in vision–language–action policies caused by independent Gaussian latent variables. To resolve this without additional training, the authors propose an inference-time smoothing method that leverages the unexecuted tail of the action sequence—conditioned on the executed prefix—as a consistency reference. By incorporating a velocity-guided loss (VLS) mechanism, the method applies closed-form corrections after each Euler integration step to ensure smooth transitions between action chunks. Evaluated on the LIBERO-10 benchmark using the pi_0.5 model, the approach reduces boundary jerkiness by 28% and inter-chunk discontinuity by 27%, while maintaining baseline task success rates and incurring negligible computational overhead.

Abstract

Vision-Language-Action (VLA) policies that execute fixed-length action chunks can exhibit multimodal bifurcation: a cross-chunk inconsistency in which adjacent chunks generated from independent Gaussian latents can converge to incompatible trajectory modes, producing abrupt discontinuities at chunk boundaries. Existing remedies either require backpropagation through the policy at each denoising step, rely on rejection sampling, or require retraining, each trading computational cost or task reliability for smoother transitions. We propose SEAM (Smooth Execution of Action-Chunked Motion), a training-free inference-time method for flow matching VLAs. SEAM exploits a simple synchronous-execution insight: after the robot consumes the executed prefix, the previous chunk's unexecuted tail is already available as an analytic consistency reference. Its core mechanism, Velocity-guided Loss Steering (VLS), derives a time-dependent target from this tail and applies a closed-form correction after each Euler step without backpropagating through the policy network. On LIBERO-10 with pi_0.5, SEAM reduces boundary jerk by 28%, reduces chunk transition discontinuity by 27%, preserves baseline-level task success, and keeps denoising-loop cost near the unguided baseline.

Problem

• Vision-Language-Action
• action chunking
• multimodal bifurcation
• trajectory discontinuity
• chunk boundary

Innovation

• Vision-Language-Action
• action chunking
• flow matching
• smooth trajectory
• inference-time correction