HJ Reachability-Based RL Enables Safer Emergency Collision Avoidance

Emergency collision avoidance under extreme driving, like on ice or wet roads, demands safety-critical control that anticipates future risks. Existing methods often only evaluate safety at the current moment, missing dynamic instabilities that emerge over time. To address this, researchers (Yuhong Jiang et al.) introduce a safe reinforcement learning framework guided by a Hamilton-Jacobi (HJ) reachability-based motion safety set. This set is built by combining geometric collision margins and chassis stability limits into a unified signed safety function, then extending it through reachability analysis over a finite horizon. The result is a forward-looking safety supervision that tells the RL agent whether it can maintain safety as its state evolves.

To make the approach practical, the motion safety set is approximated from offline extreme driving data, sidestepping the heavy computation of traditional grid-based HJ solvers. The learned safety set is embedded as a continuous cost in a constrained Markov decision process, with a PID-Lagrangian policy optimization that adaptively adjusts the Lagrange multiplier to enforce safety. In both simulation and real-vehicle tests on low-adhesion obstacle-avoidance scenarios, the method outperformed baselines: it achieved higher goal-reaching rates, produced smoother avoidance maneuvers, and maintained larger unified safety margins. The work, posted on arXiv, points toward safer autonomous driving in hazardous conditions.