Run-time robot planning

We (1993) have developed a neural network architecture which learns a forward model of a redundant manipulator (via self-supervised training) as a map of normalized radial basis neurons and inverts the model by means of run-time gradient descent of a task-related potential field. In this paper, we propose a distributed model for the computation of the field, which is consistent with the model-inversion map; and we discuss the problem of self-synchronization between the gradient-descent process and a process for the generation of virtual trajectories of the end-effector.