Autonomous Target Tracking and Obstacle Avoidance of A Quadrotor UAV Based on Meta-RL Algorithm

This paper investigates the autonomous target tracking and obstacle avoidance problem of quadrotor unmanned aerial vehicles (UAVs). Based on meta reinforcement learning (Meta-RL) theory and probabilistic embeddings for actor-critic RL (PEARL), we propose a PEARL-2Buf alogrithm that improves data utilization efficiency. By introducing a meta-task set, the PEARL-2Buf model adeptly tracks diverse target motion trajectories via end-to-end decision-making, eliminating the necessity for retraining on tasks demanding adaptation. The effectiveness of the algorithm is illustrated through a simulation example.