A Neural Network Architecture for Forward Kinematics Learning of Linked Serial Robots

Forward Kinematics (FK) is arguably the most basic computation in robotics, which computes the Cartesian position of joints given their angular positions. While industrial robots realize this computation using precise models, generic or built-in-house robots require learning an FK model, usually by employing a Neural Network monolithic model that directly maps joint positions into Cartesian coordinates. Despite the success of monolithic approaches, they fail to capture the FK computations' linked nature, in which the estimation of one joint uses the previous one's estimation. Towards improving FK learning, this paper proposes an efficient Neural Network architecture composed of an FK model for each joint, leveraging the linked properties of Denavit-Hartenberg transformations to compose and learn an FK for all joints. The proposed architecture is shown to improve learning efficiency and accuracy compared to the traditional monolithic approach using simulated robotic serial manipulators with up to 7 joints, and its usefulness is demonstrated by implementing a Cartesian velocity estimation and Cartesian impedance controller.