Short- and long-term online combined learning for robotic control

Combined short- and long-term online learning for connectionist robotic feedforward control is presented. The online adjustment of the neural network is achieved by comparison of the actual applied torque with a fictitious torque generated by applying the observed acceleration through the feedforward controller. The online procedure has a time-differentiated learning paradigm that is implemented by a dual learning paradigm. Short-term, fast learning, implemented by a simple adjustable matrix, helps in controlling the system at the beginning of the training procedure or in the presence of perturbations. The neural network model provides for the long-term learning, which is convenient for obtaining maximum dynamic performance from a robot, since the effect of undesirable perturbations required for short-term adaptive schemes is absent