Towards Faster DRL Training: An Edge AI Approach for UAV Obstacle Avoidance by Splitting Complex Environments

摘要

As autonomous Unmanned Aerial Vehicles (UAVs) are becoming more and more prevalent in everyday life, it is paramount that UAVs are equipped with effective obstacle avoidance capabilities. Edge AI, which runs AI on-device (e.g., on-UAV) or on edge servers, offers many advantages to traditional cloud-based AI when applied to the problem of UAV obstacle avoidance. Literature shows that deep reinforcement learning (DRL) applied to robots (e.g., UAVs) is an effective method of obstacle avoidance. One key issue associated with DRL applied to robotics is the time required to train when the environment is complicated. In this paper, we propose a DRL-based UAV obstacle avoidance system that leverages edge AI. Our system distributes the training and inferencing processes of DRL by splitting large environments into multiple smaller environments. Our main goal is to make DRL training faster and more feasible under relatively large and complex environments. We demonstrate the effectiveness of our system in 3D simulation and all our code is open-sourced on GitHub.