Using TRPO to control quadruped gait behaviors

Abstract

Quadruped robot locomotion control is tough and complex due to the redundant DOF and interlocked movement of their four legs, even though a suitable control method has a significant impact on the performance of the control. The following contributions are made by the paper to the creation of the ideal gait controller for the legged robot. The quadruped robot's fundamental mechanical parts are first recreated in a virtual setting. Second, a TRPO model based on KL divergence is created, and the model's accuracy and computation speed are evaluated. Using curriculum learning and actor-critic approaches, the best gaits for various walking tasks are discovered. Finally, the virtual model is updated to incorporate the learnt gaits together with many additional behaviours, including vision and directional variance. According to preliminary findings, the robot can efficiently navigate and correct its walking paths in real time with no processing overhead.

URN:NBN:sciencein.jist.2023.v11.574