As the use of UAVs becomes more prevalent, there is a growing need for increased autonomy. In this paper, we propose a robust collision avoidance method based on deep reinforcement learning in an unknown narrow corridor environment. In particular, we consider corridor environments with widths constrained to 1.6 times the UAV’s wingspan, making it difficult for the UAV to avoid obstacles. The proposed method is based on end-to-end control pattern, which takes airborne LiDAR measurements combined with the UAV state and destination information as inputs, and generates UAV control commands directly. Besides, we present the design of the deep reinforcement learning network and policy update strategy based on the PPO algorithm. In addition, we conduct extensive experimental validation both in 2D Stage scene and 3D Gazebo scene with external disturbance. The experimental results demonstrate that our method achieves robust and successful obstacle avoidance.
Robust Obstacle Avoidance for UAVs in Narrow Corridor Environment Based on Deep Reinforcement Learning
Lect. Notes Electrical Eng.
International Conference on Machine Learning, Cloud Computing and Intelligent Mining ; 2024 ; Shennongjia, China August 08, 2024 - August 11, 2024
Proceedings of the 3rd International Conference on Machine Learning, Cloud Computing and Intelligent Mining (MLCCIM2024) ; Chapter : 53 ; 635-645
2025-03-22
11 pages
Article/Chapter (Book)
Electronic Resource
English
A Low-Altitude Obstacle Avoidance Method for UAVs Based on Polyhedral Flight Corridor
DOAJ | 2023
|Vision-Based Obstacle Avoidance for UAVs
AIAA | 2007
|Obstacle Avoidance Manager for UAVs Swarm*
IEEE | 2020
|