Deep Reinforcement Learning Applied to a Spherical Robot for Target Tracking

摘要

The complexity of today’s mobile robots, equipped with multiple sensors and actuators, makes linear control strategies much more challenging. This article presents and tests a deep reinforcement learning approach to control a real spherical robot based on position, velocity, and heading. The robot’s motion is achieved by driving an electric motor for rotational dynamics and two servos for longitudinal displacement. Three deep reinforcement learning controllers are obtained with different sets of control signals of increasing complexity. The best performance is obtained by using its position and heading to control the linear and angular velocities to reduce the time to reach the target. This control technique is recommended for this type of system as it learns an optimal control law that interacts with the robot’s dynamics and the environment, natively capturing the multivariable and nonlinear characteristics of the system. The article also includes previously designed controllers for the robot, comparing their performance in a series of position control tests in simulation and on the experimental platform.