Self-learning UAV Motion Planning Based on Meta Reinforcement Learning

Motion planning in unknown environments is still a challenge, especially for Unmanned Aerial Vehicles (UAVs) with limited computation resources. Classical motion planning methods often require intricate manual adjustments to the model due to predefined control rules. Meanwhile, Deep Reinforcement Learning (DRL)-based methods also suffer from potential model failures in new scenarios, as agents repetitively engage in trial-and-error interactions within a fixed environment during policy learning. To address these problems, we propose a novel meta reinforcement learning optimization method for motion planning with autonomous obstacle avoidance named SLPlanner. Unlike traditional DRL-based approaches that narrowly concentrate on the interaction between UA Vs and the environment, meta reinforcement learning prioritizes the intrinsic dynamics of interaction, leveraging the advantages of meta-learning to inspire the UA V to make inferences in unknown scenarios. The entire problem is modelled as a Markov Decision Process (MDP) and we introduce a self-learning module within it, which enables the UA V to achieve better decision outcomes in unknown environments while maintaining consistent performance in the original scene. Comprehensive experimental results demonstrate that our approach not only has higher generalization ability in unknown environments compared to related methods, but also effectively enhances collision avoidance efficiency for quadrotor UAVs.