Autopilot-Preserving Residual Q-Learning with HJB-Inspired Finite-Action Risk Filtering for Fixed-Wing UAV Command Supervision
Researchers have developed a new method for supervising fixed-wing UAV autopilots that aims to improve path-tracking accuracy while maintaining safety. This approach places a learned supervisor above the existing autopilot, selecting residual commands for airspeed, altitude, and heading. The system uses a Hamilton-Jacobi-Bellman inspired critic and a control-Lyapunov barrier to filter these commands, ensuring a safe fallback option. AI
IMPACT Introduces a novel RL-based supervisory layer for UAV autopilots, potentially improving flight control precision and safety.