Abstract: Aiming at the difficulties in transferring reinforcement learning policies from a simulated environment to real scenarios, and to improve the pedestrian avoidance ability of a drone using monocular vision without depth information, this paper proposes a cross-sensor transfer learning method based on an asynchronous deep neural network. Firstly, a stable preliminary collision avoidance policy using only a virtual Lidar sensor is trained in the simulation environment based on the deterministic policy gradient (DDPG) deep reinforcement learning. Secondly, a monocular camera and Lidar are used to collect the real-world vision and depth data sets simultaneously and bind them frame by frame, then the aforementioned preliminary collision avoidance policy is used to automatically obtain the labels which are further used to train a new monocular vision collision avoidance policy without Lidar data, thus achieving the cross-sensor transfer leaning from the simulated Lidar to real-world monocular vision. At last, the YOLO v3-tiny network and the Resnet18 network are introduced to form an asynchronous deep neural network structure, which effectively improves the policy performance in pedestrian involved scenarios.

Key words: Monocular vision, Deep reinforcement learning, Deterministic policy gradient, Cross-sensor transfer learning, Asynchronous deep neural network