Robot inverse kinematics via neural and neurofuzzy networks: architectural and computational aspects for improved performance

Neural networks are universal nonlinear-function approximators and have for a long time been used to implement various practical nonlinear inverse mappings. The choice of network type and structure depends on the mapping type and the degree of generalization required. The use of neural networks for solving the inverse robot kinematics has been extensively studied by many workers, but still some problems related to the complexity and strong nonlinearity of the inverse kinematics process need suitable heuristic and ad hoc techniques and simplifications. The aim of this paper is exactly to contribute towards filling this gap by investigating a number of computational and architectural issues so as to improve the performance of the implementation of the inverse kinematics process. These issues include the generation and preprocessing of training data, the data scaling, the treatment of multiple solutions, the reduction of the approximation error, and the speeding-up of the training process. A generic neural network architecture is proposed which employs multiple radial-basis function network elements and the ‘mixture of local experts’ principle. An algorithm is presented for the training data preprocessing which greatly reduces the training time and overall system error. Two fuzzy-logic solutions are provided and discussed; one employing a fuzzy associative memory, and the other a neurofuzzy cell architecture proposed by the authors. In all cases no knowledge is assumed about the inverse kinematics of the robot at hand, as long as its forward kinematics is known. Some indicative simulation results are included and discussed.