Iterative learning neural network control for robot learning from demonstration

摘要

Learning from demonstration is an efficient way for transferring movement skill from a human teacher to a robot. Using a camera as a recorder of the demonstrated movement, a learning strategy is required to acquire knowledge about the \{nonlinearity\} and uncertainty of a robot-camera system through repetitive practice. The purpose of this paper is to design a neural network controller for vision-based movement imitation by repetitive tracking and to keep the maximum training deviation from a demonstrated trajectory in a permitted region. A distributed neural network structure along a demonstrated trajectory is proposed. The local \{networks\} for a segment of the trajectory are invariant or repetitive over repeated training and are independent of the other segments. As a result, a demonstrated trajectory can be decomposed into short segments and the training of the local neural \{networks\} can be done segment-wise progressively from the starting segment to the ending one. The accurate tracking of the whole demonstrated trajectory is thus accomplished in a step-by-step or segment-by-segment manner. It is used for trajectory imitation by demonstration with an unknown robot-camera model and shows that it is effective in ensuring uniform boundedness and efficient training.