Integration of CoG Estimation and Torque Control and Utilizing Reinforcement Learning to Enhance Walking Stability in Bipedal Robots

• Main Authors: Cheng, Yi-Lun, 鄭逸倫
• Other Authors: Yeh, Ting-Jen
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/7yfthz

碩士 === 國立清華大學 === 動力機械工程學系 === 106 === This thesis aims to improve walking stability in bipedal robots through integrate torque control and CoG estimation. In addition, we utilize reinforcement learning to make walking direction of the robot more straight. The cart table model is used as dynamic model of the robot. In order to use ankle torque as torque input of the system, the ankle of the robot is consisting of series elastic actuator. Because of the change of robot’s posture or measurement errors, CoG may not be at the ideal place. So, we use a controller that can automatically estimate the bias of the CoG. Based on the dynamic model and the controller above. We take the mass of the swing leg into consideration and propose a model named "double cart-table model" to generate walking pattern. The double cart table model simultaneously plans the robot’s upper body centroid and the end point of swing leg trajectory by using two equivalent masses. In this way, the robot can walk forward with a trajectory which is close to its own dynamics. Furthermore, the angular momentum caused by swing leg during walking period makes the walking direction of the robot unstable. This study then proposes a reinforcement learning method to make it more stable.