A Deep Reinforcement Learning Approach for UAS Conflict-Avoidance Maneuvers with Flight-Path Recapture

This paper presents a deep reinforcement learning approach for generating conflict-avoidance maneuvers that aim to maintain well-clear separation from an intruder aircraft while enabling subsequent recapture of the planned flight path. The proposed method is applicable to automated Detect-and-Avoid functions in lost-link scenarios as well as in autonomous flight. Pairwise encounters are used to train the neural-network policy, and the performance of the trained policy is compared with that of a conventional heuristic path-stretch algorithm. Results show broadly comparable conflict-avoidance performance, although the heuristic method is more efficient and the learning-based approach in the 40 kt case results in a small number of well-clear violations. The learned policy also exhibits a distinctive “wait-it-out” maneuver pattern that was not observed in the heuristic baseline.