Towards a geometric identification of compliant motions in learning from demonstration

We propose a method to learn compliant motions from demonstration to reduce the need for positional accuracy in an assembly task. An example of compliant motion in workpiece alignment is depicted in Fig. 1. Impedance control is a natural control method to realize compliant motions, but the drawback is that it makes motion planning difficult. Preimage planning [1] can be used for planning compliant motions in two dimensions. However, in 3-D preimage planning has been shown to be computationally infeasible [2]. In contrast to automatic planning, learning from demonstration (LfD) [3] can be used to transfer skills from a human expert to a robotic system. Most LfD research has focused on position control and only recently LfD has been extended to in-contact tasks and force control [4]. However, the existing methods are unable to learn compliant motions for aligning workpieces. We propose a method for learning compliant assembly motions from demonstrations. The intuition of our method stems from geometry: there is always a certain range of angles from which a human can push an object to make it slide along a surface. Most LfD methods try to reproduce any kinds of motions, and use tools such as dynamic movement primitives [5] or hidden semi-Markov models [6]. However, they are not well equipped to handle motions where interaction forces can appear from a variety of directions, such as inserting a tool into a funnel. We use kinesthetic teaching to collect both position and force data from demonstrations. From one or more recordings we learn the direction which will lead the end-effector through the motion, either directly in free-space or in contact. In addition, we learn the necessary compliant axes for performing the demonstrated task. Our earlier work [7] considered a similar problem under the assumption that the forces exerted by human during demonstration can be directly measured, as is the case for example in teleoperation. In contrast, this work provides a solution to a more general problem where only the interaction forces can be measured. Kronander and Billard [8] proposed learning compliance parameters from human demonstration by halting a trajectory demonstration at intervals and wiggling or tightening the grip on the robot to demonstrate the correct compliance. However, this is not a suitable method for in-contact tasks.