Off-Policy Adversarial Inverse Reinforcement Learning in Mobile Robot Navigation Task

Navigation tasks on mobile robots are getting more important to assist human tasks. However, navigation on mobile robots generally has several challenges, such as performing the task of navigating an unknown environment and being able to avoid obstacles. Conventional methods generally rely only on obstacle maps and cannot learn autonomously. Mapless navigation methods based on Reinforcement Learning (RL) are widely used to learn autonomous mobile robot navigation systems. However, the RL method requires a suitable reward function setting technique, and a simple reward function may not be easy to determine for the robot to perform navigation tasks. Therefore, this research proposed Off-Policy Adversarial Inverse Reinforcement Learning (AIRL) on mobile robot navigation tasks. We choose Off-Policy AIRL because it has strong advantages against dynamic changes so that mobile robots can perform navigation tasks efficiently. We compared the cumulative reward value, success rate, and trajectory length to reach the destination point with a reward designed via the reward shaping method and the reward estimated by Off-Policy AIRL. By applying RL using reward estimation from Off-Policy AIRL, the success rate of the robot reaching the destination point is 14 out of 15 scenarios and the proposed algorithm can outperform the reward produced by manually-reward shaping in 12 out of 15 scenarios.