Training-Free Pace-and-Path Correction for VLA Models

A novel technique known as Pace-and-Path Correction (PPC) tackles the issue of dynamics-blindness in Vision-Language-Action (VLA) models. Typically, VLAs are trained using single-frame data, which limits their ability to perceive temporal dynamics, leading to significant performance drops in non-stationary environments. Current alternatives either necessitate costly retraining or experience issues with latency and temporal inconsistency. PPC operates as a training-free, closed-form inference-time tool that can be applied to any chunked-action VLA. It separates a quadratic cost into two distinct components: a pace channel that streamlines execution in the intended direction, and a path channel that introduces an orthogonal spatial adjustment. This combined minimization effectively integrates perceived dynamics without the need for retraining. The research is available on arXiv with ID 2605.11459.