Strict safety guarantees for reinforcement learning implementations

Establish strict, formal safety guarantees for reinforcement learning implementations in mobile robotics by proving that learned policies maintain safety under exploration and execution and do not violate state or action constraints.

Background

The paper addresses goal-reaching for large-scale mobile robots operating in harsh environments and emphasizes safety throughout learning and execution. While reinforcement learning has been increasingly applied to mobile robotics, the authors note that achieving strict safety guarantees remains a difficult challenge, especially for heavy robots with complex actuation and slip-prone terrain.

The work proposes a hierarchical framework combining ORB-SLAM3 for pose estimation, a smooth RL motion planner with discrete accelerations, a supervised DNN inverse dynamics model, and a robust adaptive controller with a safety supervisor. Despite these contributions, the authors explicitly state that obtaining strict safety guarantees for RL implementations is still an open problem.