Event
SEVI: A Real-World Reinforcement Learning Safety Evidence Framework
1 reports · 1 sources
The authors propose the SEVI framework, which uses an evidential safety critic to simultaneously estimate future constraint violation risk and epistemic uncertainty in a single forward inference, determining before action execution whether an action is unsafe or has insufficient evidence, and overriding that action with a recovery policy trained offline. The authors evaluate SEVI on two simulated manipulation tasks and one real-world USB insertion task under in-distribution and distribution-shift conditions. The authors report that, compared with recovery-based safe reinforcement learning, SEVI reduces unsafe events by 13.4% to 57.1% while maintaining low-latency runtime filtering.
Report timeline
- SEVI: A Safety Evidence Framework for Real-World Reinforcement Learning
The authors propose the SEVI framework, which uses an evidential safety critic to simultaneously estimate future constraint-violation risk and epistemic uncertainty in a single forward pass, judging before an action is executed whether it is unsafe or insufficiently supported by evidence, and overriding that action with an offline-trained recovery policy.
arXiv cs.RO ·