Deep Reinforcement Learning for Autonomous Dynamic Skid Steer Vehicle Trajectory Tracking

Abstract

Designing controllers for skid-steered wheeled robots is complex due to the interaction of the tires with the ground and wheel slip due to the skid-steer driving mechanism, leading to nonlinear dynamics. Due to the recent success of reinforcement learning algorithms for mobile robot control, the Deep Deterministic Policy Gradients (DDPG) was successfully implemented and an algorithm was designed for continuous control problems. The complex dynamics of the vehicle model were dealt with and the advantages of deep neural networks were leveraged for their generalizability. Reinforcement learning was used to gather information and train the agent in an unsupervised manner. The performance of the trained policy on the six degrees of freedom dynamic model simulation was demonstrated with ground force interactions. The system met the requirement to stay within the distance of half the vehicle width from reference paths.

Keywords

• trajectory tracking
• reinforcement learning
• skid-steer robots
• deep learning
• dynamic model