Neural processing based robot kinematics modeling and calibration for pose control

Neural network approaches to calculating robot kinematics have been studied intensely in the past. Most of them were concerned with learning robot kinematics by feedforward networks with sigmoidal units, and offered only approximate solutions to the kinematics problem. In this paper, we present a method that circumvents this limitation by using a procedure that maps kinematic equations to their neural network representation. This requires the kinematic equations to be pre-processed by symbolic computation algorithms to convert them to a format (as sum of sines) that can be implemented directly by a two-layer, feedforward neural network with sinusoidal units. That is a straightforward observation that such neural network topology limits our method application to manipulators with revolute joints only. The main focus of this work lies in the efficient kinematics calibration with the help of neural networks. The process itself is performed as training the neural networks implementing the robot kinematics.