Learning-based adaptive motion control of a snake robot

Energy efficient locomotion of a wheel-less snake robot is very crucial for potential applications of untethered snake robots. For this reason, the optimal gait parameters for the energy efficient locomotion of the snake robot are investigated with two different multi-objective optimization methods based on symbiotic organism search algorithm in this study. The obtained results demonstrate that both proposed methods achieve satisfying stable results regarding power consumption reduction with optimal forward velocity for lateral undulation motion. However, it is seen that fast non-dominated sorting multi-objective symbiotic organism search algorithm provides an advantage on obtaining a uniformly distributed solution set with a good diversity only in a single run. The effectiveness of the proposed method in finding the optimal gait parameters for adaptive locomotion of the snake robot is proved by testing at nine different environments changing in a fairly wide friction range from glass to wood. Moreover, four different controllers based on deep reinforcement learning algorithms which stand out with their ability to adapt to changing environmental conditions unlike traditional control methods, are developed for trajectory tracking control of the snake robot joints in this study. The obtained results indicate that TD3 algorithm combined with transfer learning presents best performance in terms of the convergence speed and generalization capability of the controller. Finally, for the energy efficient locomotion of the snake robot, the movement patterns similar to biological snake gaits are achieved by generating adaptive joint trajectories with a speed controller trained using TD3 algorithm.