A Mapless Navigation Method Based on Reinforcement Learning and Local Obstacle Map

Recently, research on mapless navigation based on Reinforcement Learning (RL) shows great potential. However, for indoor scenes with "local minimal areas", the results of end- to-end training based solely on RL are not ideal. This paper presents an encoding method for navigation input information, which encodes the single-line lidar information into a Local Obstacle Map (LOMap), and combines the robot speed and target point position information into the neural network. We modify a Safe RL algorithm Constrained Policy Optimization (CPO) for training, and propose a new navigation evaluation metric that considers time-consuming. Simulation experiments on Gazebo show that our proposed method can better adapt to indoor scenes with local minimal areas, and outperforms the CPO method with raw lidar information input in common metrics.