A neural network approach to pointwise control of robot manipulators

The use of neural networks as an alternative to the computation-intensive control-theoretic approach to motion planning and execution is discussed. A three-tiered network structure is proposed: one tier is for conversion of task-space coordinates to joint-space coordinates, one is for converting the start and goal positions in joint space to a dynamic sequence, and one is for enabling the manipulator to execute the sequence by generating the necessary control signals. The issues involved in implementation are discussed and some results are presented. It is concluded that the neural-network approach has shown some promise in its ability to tackle nonlinear static problems and as a tool for generating plans. However, its effectiveness in eliminating completely any kind of online computation for precise control seems doubtful