Augmented Neural Network for Full Robot Kinematic Modelling in SE(3)

Abstract

Due to the increasing complexity of robotic structures, modelling robots is becoming more and more challenging, and analytical models are very difficult to build. Machine learning approaches have shown great capabilities in learning complex mapping and have widely been used in robot model learning and control. Generally, the inverse kinematics is directly learned, yet, learning the forward kinematics is simpler and allows computing exploiting the optimality of the controllers. Nevertheless, the learning method has no knowledge about the differential relationship between the position and velocity mappings. Currently, few works have targeted learning full robot poses considering both position and orientation. In this letter, we present a novel feedforward Artificial Neural network (ANN) architecture to learn full robot pose in SE(3) incorporating differential relationships in the learning process. Simulation and real world experiments show the capabilities of the proposed network to properly model the robot pose and its advantages over standard ANN.