Learning from demonstrations through the use of non-rigid registration

Abstract

We consider the problem of teaching robots by demonstration how to perform manipulation tasks, in which the geometry (including size, shape, and pose) of the relevant objects varies from trial to trial. We present a method, which we call trajectory transfer, for adapting a demonstrated trajectory from the geometry at training time to the geometry at test time. Trajectory transfer is based on non-rigid registration, which computes a smooth transformation from the training scene onto the testing scene. We then show how to perform a multi-step task by repeatedly looking up the nearest demonstration and then applying trajectory transfer. As our main experimental validation, we enable a PR2 robot to autonomously tie five different types of knots in rope.