16 자유도 방사형 4족 로봇의 험지에서의 안정성과 에너지 효율성 향상을 위한 보행 학습

We herein propose learning locomotion with uneven terrain stability and energy efficiency using deep reinforcement learning for a sprawling-type quadruped robot with 16 degree-of-freedom (DOF) leg mechanisms. Our aim is to improve both locomotion stability and energy efficiency by dynamically adjusting the body height based on the locomotion speed. The learning strategy in this study comprises three steps: 1) The workspace of the proposed 4-DOF leg, unlike the 3-DOF leg structure of the existing robot, is compared and analyzed. 2) A body-height reward based on the robot's locomotion speed is designed using the Froude Number. 3) In Fractal Noise Terrain, rapid learning and generalizability for increasing adaptability to various environments are achieved via transfer learning. Finally, the policy learned by Isaac Gym is verified by comparing and analyzing locomotion stability and energy efficiency in various terrains and conditions via the Gazebo Sim-to-Sim process.