PRISM: Primitive Routing via In-context Skill Mixing for Retentive VLA Adaptation
PRISM: Primitive Routing via In-context Skill Mixing for Retentive VLA Adaptation
This paper introduces PRISM, a framework that stores reusable skills as frozen low-rank action-expert adapters in a primitive bank. A per-chunk vision-language model conditioned router composes these primitives during inference by tracking task progress via prefix representations and is supervised by an action-derived oracle. Evaluated on RoboCasa365, PRISM achieves 56.43% atomic-skill success and 20.00% long-horizon composite task success, outperforming a standard two-stage LoRA fine-tune (17.86% atomic, 17.50% composite). The method is also validated on an AgileX PIPER real-robot setup.
The authors propose PRISM, a method using low-rank adapters and an in-context router to enable skill retention and composition in Vision-Language-Action models, validated on RoboCasa365 and an AgileX PIPER robot.
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.
n RoboCasa365, PRISM achieves 56.43% atomic-skill success and reaches 20.00% on long-horizon composite tasks, while a standard $\pi_{0.5}$-LoRA two-stage fine-tune reaches 17.86% atomic and 17.50% composite success. We additionally validate PRISM on an AgileX PIPER real-robot setup.
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org