Multiple Robots Path Planning based on Reinforcement Learning for Object Transportation

Abstract

This paper proposes reinforcement learning methods to perform an object transportation task for multiple robots. This task consists of two main subtasks, path planning and motion control task. Double deep Q-learning (DDQN) model is selected to achieve path planning for an unknown environment. To increase the capability of reinforcement learning model, semi-supervised method by A* algorithm is applied during the training process. In motion control task, reinforcement learning model is designed to control a movement of a differential wheeled mobile robot. The actions of mobile robot consisting of linear and angular velocities are computed by agent. The models for motion control task are separately trained for two different purposes. The first agent is trained to deal with the path following task and the other agent is trained to handle the point following task. The agent of the point following task is utilized to control the group of robots to move with a specific formation. Proximal policy optimization (PPO) is selected for the path following task and deep deterministic policy gradient (DDPG) is selected for the point following task. Eventually, the integration of the proposed reinforcement learning models can accomplish the object transportation task for multiple robots successfully both in simulation and experiment.