Machine Learning based Joint Torque calculations of Industrial Robots

Getting closed-form inverse dynamics solution for a manipulating robot is desirable for real-time torque computation. Some powerful and well-established mechanics based tools like Newton Euler method (N-E), Lagrangian methods are available for developing a mathematical model for manipulating robots. However, the coupled equation being highly non-linear and incomplete (joint friction, dimensional inaccuracies arising due to manufacturing error are very difficult to model) and hence difficult to apply in real life situation which requires accurate joint torque computation. We believe learning based machine intelligence tools can more efficiently and appropriately be utilized for joint torque computations in the inverse dynamics paradigm. In this paper, K-nearest neighbor (KNN) algorithm has been proposed for finding joint torques from the dataset created by solving forward dynamics equation for the manipulating robots, which is comparatively straight forward and rather less complex. However, since computational complexity of KNN is high, we used K-dimensional tree (K-D tree) for decreasing the computational complexity. The simulation result for two-link manipulator shows the proposed method of KNN based joint torque calculation coupled with K-D tree is simple, robust and accurate, which can be emulated for a further higher degrees of freedom robot having six links.