Learning to chase a ball efficiently and smoothly for a wheeled robot

Target chasing control of autonomous mobile robots is important for both civilian and military applications. Efficiency and stability are two important factors in chasing problem. The robot is expected to chase the target quickly to save time. Besides, the driving velocity and direction of the mobile robot should be changed smoothly to avoid wheel slippage or mechanical damage during the chasing process. In this paper, we propose a method for a wheeled robot to learn a policy to chase a red ball efficiently and smoothly. Without any knowledge of motion strategies, the wheeled robot can be trained to chase its target by learning from the given rewards using Deep Reinforcement Learning (DRL). The motion control of the robot is decided not only by the features of the target, but also by the current state of the robot itself. The reward is set according to Multi-Sensor data of the robot for good chasing performance. We use Double Deep Q-Network (Double DQN) to build our model and have obtained good experimental results in simulation environment.