Toward a Computationally Efficient Solution of the Inverse Kinematics Problem Using Machine Learning

The paper reports work-in-progress toward exploiting Machine Learning methods to solve the inverse kinematics (IK) problem in real-time, in order to control the pose (position and orientation) of the end-effector of a robotic arm. The work explores the use of Unit Dual Quaternion to the kinematic of a robot arm of serial architecture with six degrees of freedom (DOFs). Non-linear constrained optimization is applied to solve the IK problem taking into account actuation (motor) limitations in every DOF. The results show that this approach overcomes singularities and achieves a continuous “best realizable” solution, even when the target path is outside the reachable workspace. This sets the basis for further research work, namely, building an off-line dataset that can be used to train a computationally efficient Machine Learning model running in real-time.