A geometric approach for learning compliant motions from demonstration
This addresses the challenge of planning compliant motions in unstructured 3D environments for robotics, though it is incremental as it builds on existing Learning from Demonstration approaches.
The paper tackles the problem of learning compliant contact motions from human demonstrations to align workpieces using interaction forces, even under varying contact directions, and demonstrates that their method successfully reproduces such motions on a KUKA LWR4+ robot.
This paper proposes a method to learn from human demonstration compliant contact motions, which take advantage of interaction forces between workpieces to align them, even when contact force may occur from different directions on different instances of reproduction. To manage the uncertainty in unstructured conditions, the motions learned with our method can be reproduced with an impedance controller. Learning from Demonstration is used because the planning of compliant motions in 3-D is computationally intractable. The proposed method will learn an individual compliant motion, many of which can be combined to solve more complex tasks. The method is based on measuring simultaneously the direction of motion and the forces acting on the end-effector. From these measurements we construct a set of constraints for motion directions which, with correct compliance, result in the observed motion. Constraints from multiple demonstrations are projected into a 2-D angular coordinate system where their intersection is determined to find a set of feasible desired directions, of which a single motion direction is chosen. The work is based on the assumption that movement in directions other than the desired direction is caused by interaction forces. Using this assumption, we infer the number of compliant axes and, if required, their directions. Experiments with a KUKA LWR4+ show that our method can successfully reproduce motions which require taking advantage of the environment.