Towards Efficient Mapless Navigation Using Deep Reinforcement Learning with Parameter Space Noise

This paper presents a deep reinforcement learning framework that is capable of training the mapless motion planner end-to-end by taking the laser range findings as input and the continuous steering commands as output. Major improvements are introduced in our Deep Deterministic Policy Gradient algorithm (DDPG): parameter space noise is used to encourage exploration and an efficient exploration strategy is designed to boost navigation performance. The proposed learning framework is implemented to train the mobile robot in the Gym-gazebo simulation environment. The simulation study shows that the proposed mapless motion planner can navigate the nonholonomic mobile robot effectively without collisions.