Target Following for a Snake-Like Robot Using Rolling Locomotion via Deep Reinforcement Learning

The turning motion of snake-like robots based on center of gravity shift has superior energy efficiency compared to the undulatory locomotion that simultaneously drives many joints, but precise guidance control has been difficult due to its nonlinear dynamic characteristics. In this paper, we propose a control method using deep reinforcement learning for following target points in all directions in the turning motion using center of gravity shift and ground reaction forces. The target direction is defined as a normalized vector in the body coordinate system. This allows us to present target direction information to the agent that is independent of changes in self-orientation, and we show that it can acquire a policy for accurate following in all directions. The results of motion analysis confirmed that the learned policy not only skillfully utilizes the ground reaction forces at the time of contact to make dynamic trajectory corrections, but also achieves target following by flexibly switching the turning strategy according to the target direction.