Effects of Training Difficulties on Reinforcement Learning Based Outdoor Robot Navigation System

摘要

This paper presents a map-based navigation system for outdoor mobile robots and results from different training difficulties on reinforcement learning implementations. The proposed navigation system can navigate the robots using maps in the form of 2D binary images. Navigation maps can be processed from conventional map services such as Google Map. The navigation system includes the path planning segment and the navigation segment. A-star search algorithm is used to plan paths on the map. Q-learning is applied in the navigation segment to train the robot to follow planned paths on the map. Location differences between the robot and the A-star generated path are used as states for q-learning. Experiments include navigation tests of two robots which are trained under different training difficulties. Success rate of reaching the goals is used to evaluate the navigation system. Simulation results display better navigation performances of the robot trained in the training settings with more difficulties.