Robot skill discovery bases on observed data.

Abstract

Should robots be able to acquire skills as humans do, a dramatic improvement in their pelfomnce is expected to come true. In this paper, we present a method for representing and discovering skills by a robot based on observed data. First, the capabilities of a robot to transform a situation from one to another based on available control actions are extracted from the data, and represented in the situation-action space as a feasible situation transition manifold (FSTM). The Multi-resolution Globally Competitive and Locally Cooperative algorithm is formulated to self-organize the FSTM in terms of an union of hyper-ellipsoidal subregions in various sizes and shapes. An optimal sequence of transitions from the initial to the goal situations is searched for as a skill, under the constraint imposed by the FSTM. The search of an optimal sequence is based on the novel Bidirectional Dynamic Path Planning algorithm formulated based on the potential field method. The proposed methodology is applied for the discovery of a nonholonomic motion planning skill of a car-like robot and for a telemanipulation skill.