Synthesizing the optimal gait of a quadruped robot with soft actuators using deep reinforcement learning

Quadruped robots have the advantages of traversing complex terrains that are difficult for wheeled robots. Most of the reported quadruped robots are built by rigid parts. This paper proposes a new design of quadruped robots using soft actuators driven by tendons as the four legs. Compared to the rigid robots, the proposed soft quadruped robot has inherent safety, less weight and simpler mechanism for fabrication and control, but the corresponding challenge is that the accurate mathematical model applicable to model-based control design of the soft robot is difficult to derive by dynamics. To synthesize the optimal gait controller of the soft-legged robot, the paper makes the following contributions. First, the flexible components of the quadruped robot are modeled with different finite element and lumped parameter methods. The model accuracy and computation efficiency are analyzed. Second, soft actor–critic methods and curriculum learning are applied to learn the optimal gaits for different walking tasks. Third, The learned gaits are implemented in an in-house robot to transport hand tools. Preliminary results show that the robot can walk forward and correct the walking directions.