Full-Actuation Rolling Locomotion with Tensegrity Robot via Deep Reinforcement Learning

Tensegrity robots, entirely composed of elastic cables and rigid rods, have many excellent properties which have a wide application from complex co-robotics to planetary. Nevertheless, it is still difficult to control tensegrity robots because of their unconventional designs and highly coupled dynamics. Deep reinforcement learning algorithms have been used in a lot of robotic tasks due to their strong perception and decision-making capabilities. However, it often needs to collect a lot of samples, which limits its application. Model-based algorithms can learn with fewer samples but have a sub-optimal result because of the model error. In the paper, we proposed a hybrid method to achieve effective control of tensegrity robots. Firstly, we established the simulation platform via the framework of the positional formulation finite element method. And thens, we use a medium-sized neural network to fit the dynamic model and control the tensegrity robot via model prediction control (MPC). The controlled trajectories are used to initialize the parameters and memory of deep deterministic policy gradient (DDPG). We demonstrated that the hybrid algorithm can achieve efficient control of the tensegrity robot. In this work, we have realized full-actuation rolling with a tensegrity robot on a plane and 5° slope surface.