Using Deep-Learning Proximal Policy Optimization to Solve the Inverse Kinematics of Endoscopic Instruments

There is currently a trend towards small tendon-driven robotic devices targeting endoscopic applications. As these robotic systems aim to navigate within narrow tortuous pathways, their joint arrangement must be longitudinal. Additionally, mechanical constraints such as tendon routing limit the joint range which impacts the overall dexterity of the instruments and complicates the use of iterative inverse kinematics solvers. This article investigates the use of reinforcement learning and proximal policy optimization to solve the inverse kinematics problem of endoscopic instruments while considering joint limits, joint velocity, and whole body configuration. The results show that the proposed approach is able to learn the kinematic model of a 5 degrees of freedom endoscopic instrument and performs better than the Damped Least Square iterative approach in terms of computation time, position error, and orientation error.