STAR: Steering Action Representations during Inference
STAR: Steering Action Representations during Inference
This research introduces STAR, a training-free inference method for vision-language-action models that addresses a grounding gap in late-layer action representations. By applying an exponential moving average over residual updates and adapting mixing strength based on proprioceptive states, STAR improves object awareness without model updates or architectural changes. Experiments show it increases average task success rates by 4.0 percentage points on robot manipulation benchmarks.
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.
- code: Available Artifact link Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ixing when execution dynamics change rapidly. Experiments on robot manipulation benchmarks show that STAR improves late-layer object awareness and increases average task success rates by 4.0 percentage points. Our code is available at https://github.com/PKU-VaLuE-Lab/STAR.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org