Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing
Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing
This paper introduces Selected Diffusion Noise (SDN), a training-free method that enhances the reliability of Vision-Language-Action (VLA) policies at test time. By treating diffusion noise as a controllable degree of freedom, SDN samples multiple candidate action chunks and applies a two-stage selection process: first filtering candidates based on task-relevant visual grounding using a counterfactual negative set, then selecting the kinematically most stable chunk based on JerkRMS. Experiments across three VLA policies in both simulation and real-world environments demonstrate improvements of 8-10% and 18.33%, respectively, highlighting SDN's effectiveness for test-time improvement.
The authors present a training-free method called Selected Diffusion Noise (SDN) to improve the test-time reliability of Vision-Language-Action (VLA) policies by filtering action candidates for visual grounding and kinematic stability, reporting specific percentage improvements in simulation and real-world environments.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ironments show improvements of 8-10% and 18.33%, respectively, demonstrating SDN's effectiveness for test-time improvement.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org