Inverse kinematics of a kinematically redundant parallel mechanism based on deep reinforcement learning

Abstract

When a mobile platform of a kinematically redundant parallel mechanism moves along a desired path, the redundant features of the mechanism can be utilized to achieve impressive multi-motion capabilities. This study explores the inverse kinematics of a 3-PRPR mechanism using deep reinforcement learning, considering singular avoidance and end-effector pose accuracy. A degree of closeness to singularity and training strategy based on the SAC algorithm are established, and a fixed joint training strategy method based on SAC algorithm is proposed to enhance the accuracy of the end-effector pose. A simulation study is performed to demonstrate the effectiveness of avoiding singular configuration and pose errors, and the proposed method is compared to the original training strategy. The simulation results show that, the proposed fixed joint method required fewer iterative convergence episodes and had fewer pose errors than the original strategy. The results also indicate that the original training strategy with a single-agent is not ideal for the cooperative relationship of the actuators. However, while consuming the redundant characteristics, the fixed joint method provides an acceptable motion index for this mechanism.