A Novel Solution to the Inverse Kinematics Problem in Robotics Using Conditional Invertible Neural Networks

Solving of the inverse kinematics(IK) problem is a challenging topic in the field of robotic arm controlling, due to the nonlinear, high computational costing, or ill-posed characteristic of IK problem. Various solutions have been proposed and most of them are still time-consuming or may cause unrealistic poses. This paper proposes a method based on Conditional Invertible Neural Networks(cINNs), designed to serve as a computational tool for solving the IK problem of 6 degrees of freedom (DOF) robotic arms. The bidirectional mapping characteristics of Invertible Neural Networks(INNs) are utilized and the joint positions of the previous state of the robotic arm are introduced as conditional constraints, enabling precise backward prediction of the positions of each joint of the robotic arm. Experimental results show that the method based on cINNs is more efficient than traditional inverse kinematics solution(IKS), and the accuracy rate remains stable between 99% and 97%. In contrast, the accuracy of IKS fluctuates considerably, ranging from 99.5% to 85%, especially when the target position is close to the operating boundary of the robotic arm, and then the error will significantly increase. These experimental results highlight the enormous potential and value of cINNs in solving inverse kinematics problems and emphasize the clear advantages of cINNs in terms of stability and efficiency.