This project considers the problem of apprenticeship learning, in which a robot first gets access to demonstrations of a task and ought to learn from these demonstrations how to perform that task in new, yet similar, situations. This line of work has already shown significant promise, including in helicopter control where it enabled autonomous helicopter aerobatics at the level of the best human pilots. However, fundamental limitations remain, and robotic capabilities to manipulate deformable objects are currently still well below human level. The approach followed builds on, and extends, non-rigid registration algorithms, which can capture how scenes with deformable objects relate to each other. Such registration is extrapolated to morph a demonstrated manipulation trajectory into a good trajectory for a new scene. New machine learning algorithms are developed to enable choosing the optimal training demonstration and the optimal morphing objective while accounting for external constraints, such as avoiding collisions and satisfying joint limits. Infrastructure is being built for large-scale data collection of demonstrations and theoretical and empirical characterizations are developed for how much data is needed for a given task. Concrete challenge tasks considered are knot tying, cloth and fabric manipulation, surgical suturing, and small surgical procedures. Results will be incorporated into the PI's graduate robotics course and the source code will be shared with the robotics community.
