Abstract
The introduction of robots to innovative tasks requires discovering skills, or knowledge which leads us to the success of the target tasks. This report proposes a general framework for supporting the skill discovery of robotic tasks by human operators; the presented framework expects human operators to complement where computers fail. Diminishing the role of human operator in the current framework provides a perspective toward complete automatization of the skill discovery process.
We define skill as a continuous function from input space with sensory information and time variables to output space with actuator control information. We approximate skill functions with polynomial expressions and reduce the skill discovery to a combinatorial optimization problem among the coefficients of monomials in the expressions. A genetic algorithm is applied to the optimization because of its all-around applicability and its robustness against local optima. The current framework relies on human experts for the strategies of making efficient search, such as partitioning search space into smaller ones and designing better fitness for GA. The procedure to extract basic mechanism explicitly out of the acquired robot behavior is called skill understanding. Our framework enables this procedure because the skill represented by polynomials is partitioned into monomials and which monomials dominate control information can be analyzed by human operators. In order to verify the validity of the proposed framework, two examples are examined; a visual servo problem of catching a flying ball by a robot baseball player and a problem to find open-loop control of a mechanical system with nonholonomic constraints. The first example illustrates the skill understanding procedure and the second one exhibits a unified manner to solve robotic problems with nonholonomic constraints.
