Learning Safe Flow Matching Robot Policies from Binary Feasibility Oracles
Learning Safe Flow Matching Robot Policies from Binary Feasibility Oracles
This research proposes an oracle-guided differentiable safeguard for safe flow-matching robot policies, using binary feasibility queries to correct unsafe actions during forward passes and extract constraint-aware signals during backward passes. The method enables forward-backward aligned policy learning without explicit constraint models, achieving 100% constraint satisfaction in obstacle-aware manipulation tasks.
This paper addresses the challenge of ensuring safe robot policies in environments with complex physical constraints, by introducing an oracle-guided differentiable safeguard that operates without explicit constraint models. The method is validated on obstacle-aware manipulation tasks, achieving 100% constraint satisfaction while maintaining task performance.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org