RSS 2015 Workshop

Bridging the Gap between Data-driven and Analytical Physics-based Grasping and Manipulation

July 16, Rome

"Thin-Plate Splines and Learning to Manipulate from Demonstrations", Pieter Abbeel, UC Berkeley

A key challenge in learning to manipulate is generalization to new situations. I will describe an approach that uses non-rigid registration to warp training scene (where the demonstration happened) onto the test scene (which is a previously unseen situation where the robot has to autonomously succeed). While registration is only concerned with the objects and their environment, we show that it is possible to meaningfully extrapolate the registration and to also warp the robot gripper trajectories from training scene to test scene. This approach is particularly appealing for the manipulation of deformable objects such as rope or cloth, which present the robot with a very high-dimensional state space and large amounts of variability, making them particularly challenging for robots to manipulate. This approach has enabled autonomous knot tying for a wide range of knot-types and starting configurations as well as automation of some simplified suturing tasks.