Backward construction-a decomposed learning method for robot force/position control

A decomposed learning method, where the motor task is partitioned into several stages by virtue of its task characters, is proposed. The controller can be implemented by neural networks which can be trained partially in different stages. Some of these neural networks, which can be called output networks, directly drive the objects to move along the desired trajectories. The others perform some virtual activations to feed the output networks to make them have a proper output; these are called internal networks. Simple samples that can describe the expected partial characters are engaged in each stage. The well-trained parts can be used to help to train the other parts. The approach uses the notion of backward construction to build backwards the interpretation of the desired motor strategy progressively. This approach is demonstrated in detail with an example of robotic manipulator position/force learning control