Benchmarking Off-Policy Deep Reinforcement Learning Algorithms for UAV Path Planning

This paper presents a benchmarking of off-policy Reinforcement Learning (RL) algorithms for Unmanned Aerial Vehicle (UAV) path planning. The focus is on assessing the generalization from simulation to real-world deployment. The benchmark includes optimal simulation environments, a comparative analysis of off-policy RL algorithms (DDPG, TD3, SAC, SoftQ), and a simulation-to-reality transfer procedure. Contributions encompass environment design, algorithm comparison, and a transfer method. Results show DDPG excelling in maximum reward, while TD3 demonstrates superior collision avoidance. Generalization is tested in diverse environments, revealing algorithmic strengths and weaknesses. The study concludes with insights and outlines for future research, including refining metrics, exploring more algorithms, and incorporating vision-based RL for UAV path planning.