Multiple gait locomotion generation for quadruped robots based on trajectory planning and reinforcement learning

Abstract

Realizing flexible multiple gait locomotion like real quadrupedal animals remains a challenge in the field of quadruped robot motion control. To address this challenge, a multiple gait locomotion generation framework based on trajectory planning and reinforcement learning is proposed for quadruped robots, enabling the generation of various quadrupedal and three-legged locomotion. In this framework, a reference trajectory generator and a trajectory adjustment module are designed. The generator produces the reference trajectory by integrating the policy network, gait temporal information, and foot positions, using a specially designed reference trajectory formula. The adjustment module subsequently refines the trajectory in real time by employing the policy network to adapt the motion to the specific task requirements. The policy network is trained using reinforcement learning in simulation. The proposed framework has been verified both in the simulation environment and on an actual quadruped robot (Unitree A1). Both simulation and physical experimental results demonstrate that the framework enables the quadruped robot to perform a variety of gaits, similar to real animals, including quadrupedal (walk, walkingtrot, amble, trot) and three-legged locomotion (lift any leg). Furthermore, the framework enables the robot to traverse rough terrain and ascend slopes of up to 30 degrees with robustness.