Path-integral-based reinforcement learning algorithm for goal-directed locomotion of snake-shaped robot

Abstract

This paper proposes a goal-directed locomotion method for a snake-shaped robot in 3D complex environment based on path-integral reinforcement learning. This method uses a model-free online Q-learning algorithm to evaluate action strategies and optimize decision-making through repeated "exploration-learning-utilization"processes to complete snake-shaped robot goal-directed locomotion in 3D complex environment. The proper locomotion control parameters such as joint angles and screw-drive velocities can be learned by path-integral reinforcement learning, and the learned parameters were successfully transferred to the snake-shaped robot. Simulation results show that the planned path can avoid all obstacles and reach the destination smoothly and swiftly.