GRU-Attention based TD3 Network for Mobile Robot Navigation

摘要

In this paper, we propose a goal-oriented navigation reinforcement learning network called GRU-Attention based TD3 network, which takes lidar measurements, the distance between agent and target, and yaw toward the target as state inputs. The policy in the network outputs continuous actions consisting of forward velocity and yaw angular velocity. Our proposed network can perform obstacle-avoidance navigation without prior knowledge of the environment. We train our network in a simulation environment. To show that our proposed network is better in navigation tasks, we compare its performance with two other networks: the pure TD3 network and the GRU-based TD3 network in several different simulation worlds. The experiments show that the mobile robot with our proposed network can bypass the obstacles safely and arrive at the goal positions as fast as possible. The supplementary video is given at: https:// youtu.be/HkqUZSsT5a0. The implementation is made open source at: https://github.com/Barry2333/ DRL_Navigation.git.