Deep Reinforcement Learning for UAV Navigation Based on Lightweight Simulation Environment

Autonomous navigation systems have garnered significant research attention in recent years, demonstrating potential applications in warehouse logistics, autonomous vehicles, and robotic technique. While traditional obstacle-avoidance algorithms face critical limitations including high computational overhead, dependency on pre-existing maps, and inadequate adaptability to dynamic environments. The deep reinforcement learning (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$D R L$</tex>) techniques has shown obvious advantage in intelligent navigation research. This work introduces a distributed multi-agent obstacle avoidance framework leveraging Proximal Policy Optimization(PPO), this paper introduces differential artificial potential field for reward plasticity to guide the Unmanned Aerial Vehicle (UAV) to complete the task step by step. Furthermore, we developed a high-fidelity simulation environment that accurately replicates quadrotor UAV dynamics and sensor payload characteristics, enabling scalable training and validation of navigation strategies under complex operational scenarios. Experiments show that the algorithm proposed in this paper has a mission success rate of more than 97 % in a variety of typical scenarios, and the simulation environment in this paper can support more than 100 UAVs to perform the mission at the same time.