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

摘要

Combined short- and long-term online learning for connectionistnrobotic feedforward control is presented. The online adjustment of thenneural network is achieved by comparison of the actual applied torquenwith a fictitious torque generated by applying the observed accelerationnthrough the feedforward controller. The online procedure has antime-differentiated learning paradigm that is implemented by a dualnlearning paradigm. Short-term, fast learning, implemented by a simplenadjustable matrix, helps in controlling the system at the beginning ofnthe training procedure or in the presence of perturbations. The neuralnnetwork model provides for the long-term learning, which is convenientnfor obtaining maximum dynamic performance from a robot, since the effectnof undesirable perturbations required for short-term adaptive schemes isnabsent