Cooperative transportation of tether-suspended payload via quadruped robots based on deep reinforcement learning

Transportation of payloads in cooperative manner is common in human life. It involves the policy to suppress payload swing during legged walking, which is quite different from the wheeled transportation. To promote the field of swinging dynamics in multi-robotics system, we discussed the characteristic of two quadruped robots to deal with a tethersuspended payload based on proximal policy optimization reinforcement learning algorithms. Instead of relying on complex model-based control frameworks, the deployment of our method was relatively simple for dynamic walking. The control problem was modeled in simulation and a universal quadruped robot motion control reward mechanism was developed. Domain Randomization was adopt to minimize the Sim2Real gap. Two locomotion methods including omnidirectional walking and directional walking were demonstrated to tracking a rectangle path and an eight pattern path. And the policy was deployed on Go1 quadruped robot in real experiments.