Generalizing Demonstrated Manipulation Tasks

Abstract

Captured human motion data can provide a rich source of examples of successful manipulation strategies. General techniques for adapting these examples for use in robotics are not yet available, however, in part because the problem to be solved by the robot will rarely be the same as that in the human demonstration. This paper considers the problem of adapting a human demonstration of a quasistatic manipulation task to new objects and friction conditions (Figure 1). We argue that a manipulation plan is similar to a demonstration if it involves the identical number of contacts and if the applied contact wrenches follow similar trajectories. Based on this notion of similarity, we present an algorithm that uses the human demonstration to constrain the solution space to a set of manipulation plans similar to the demonstration. Our algorithm provides guarantees on maximum task forces and flexibility in contact placement. Results for the task of tumbling large, heavy objects show that manipulation plans similar to a demonstration can be synthesized for a variety of object sizes, shapes, and coefficients of friction. Experimental results with a humanoid robot show that the approach produces natural-looking motion in addition to effective manipulation plans.