Uncovering Success in Manipulation

Experiments were performed with the NASA Robonaut to determine if manifold learning could discern successful and unsuccessful teleoperation trials in an unsupervised manner. Repeated teleoperation of drill-mating and chisel-pickup tasks were performed by a skilled teleoperator. Spatio-temporal Isomap (STI) was used to embed data from the robot’s sensory-motor state-space (SMSS) to uncover underlying structure and separability between successful and unsuccessful trials. We present results from embedding SMSS data from repeated teleoperation performances, visualized in 3 dimensions, where success and unsuccessful trials are discerned. Our results are further evaluated by out-of-sample projection and comparison with Support Vector Machines for classifying the success of new teleoperation trials.