Fault‐Aware Safe Reinforcement Learning and Robust Dynamic Inversion Framework for Scalable Multi‐UAV Formation Control

This paper presents a health‐aware hybrid control architecture for resilient and safety‐critical multi‐UAV formation flight, unifying fault‐aware Safe reinforcement learning with robust generalized dynamic inversion (Safe RL–RGDI). A neural fault and disturbance observer (FDO) provides online estimates of thrust‐efficiency loss and actuator biases, enabling health‐informed constraint adaptation within a distributed CLF–CBF safety filter. These health indicators allow dynamic reconfiguration of collision‐avoidance and actuation constraints under evolving degradation. The Safe RL agent generates nominal formation‐control actions that are minimally adjusted by the CLF–CBF filter to guarantee forward invariance of safety sets, while an RGDI inner loop delivers fast, smooth, and robust attitude and angular‐rate tracking. A Lyapunov‐based analysis establishes uniform ultimate boundedness of all closed‐loop signals and formal safety guarantees despite model mismatch, wind disturbances, measurement noise, and partial thrust loss. High‐fidelity six‐UAV simulations demonstrate strong health‐aware performance. The proposed framework maintains highly accurate nominal formation tracking, limits fault‐induced deviations to 0.19 m under 30% thrust degradation, and restores formation‐tracking errors to within a 2 cm band in 14.8 s following fault occurrence, while preserving a minimum inter‐UAV spacing of 0.8 m and satisfying actuator constraints. These results demonstrate that the proposed Safe RL–RGDI framework provides a computationally efficient, safety‐assured, and degradation‐aware solution for autonomous aerial formations operating under uncertain and fault‐prone conditions.