A PbD approach for learning pseudo-periodic robot trajectories over curved surfaces

This paper provides a new method for modeling, clustering, and generalizing complex pseudo-periodic motions in a Robot Programming by Demonstration (PbD) framework. Relevant features of the trajectories are extracted by applying a linear mapping off the surface part using Moving Window Principal Component Analysis. A Hidden Markov Model is used for segmentation and temporal clustering of feature data from multiple trajectories. The generalized trajectory is derived by spline fitting through the time-aligned features, followed by a reverse mapping back onto the working surface. The proposed approach is tested with an experiment in a highly manual manufacturing process (shot peen forming), and it is shown that the generalized paths are both more consistent and more effective than those observed in the human demonstrations.