Event
VLA-Stem: A Reinforcement Learning Safety Mechanism Without Fine-Tuning VLAs
1 reports · 1 sources
Bojun Zou and other researchers propose VLA-Stem, a lightweight reinforcement-learning-based safety mechanism: instead of fine-tuning the underlying vision-language-action (VLA) policy, it converts its action chunks into motion references and, at inference time, combines proprioceptive and local obstacle information to adaptively correct execution. Across six constrained simulation settings, VLA-Stem achieved an average collision-free task success rate of 52.8%, versus 23.7% for the strongest safety baseline; in obstacle-free environments it maintains the original policy's performance. Real-robot experiments show reduced collisions under both static and dynamic obstacles.
Report timeline
- VLA-Stem: Adding a Lightweight Safety Mechanism to Vision-Language-Action Models with Reinforcement Learning
Researchers propose VLA-Stem, a lightweight reinforcement-learning-based safety mechanism that does not fine-tune the underlying VLA policy; instead, it converts its action chunks into motion references and, at inference time, adaptively corrects execution using proprioception and local obstacle information. Across six constrained simulation settings, VLA-Stem achieved an average collision-free task success rate of 52.8%, versus 23.7% for the strongest safety baseline, while preserving the original policy's performance in obstacle-free environments; real-robot experiments also showed reduced collisions under static and dynamic obstacles.
arXiv cs.RO ·