Abstract
This report presents a general method toward discovering unknown skills which are required to achieve some dexterous tasks. We define a skill as a domain specific mapping from sensory input to control output and formulate skill discovery as a strong learning of such skills: strong learning assumes little a priori information of the task. Apart from the context of manipulation or assembly tasks, our definition of skill involves dynamic and reactive behavior. We approximate a skill function by a polynomial function and reduce skill discovery to a combinatorial optimization problem among the coefficients of polynomial terms. Our approach applies genetic algorithm (GA) to this problem because GAs require no a priori knowledge of the problem and are shown to be competent to solve combinatorial problems. Practical issues about how to enable GAs to make efficient search are also discussed.
As a preliminary test case of the proposed framework, we consider a simulation setup to discover a skill to catch a ball flying in 3D space. A mobile robot on a plain is given elevation and azimuth angles and their time derivatives of the ball from the robot viewpoint and is commanded to catch balls projected with various initial velocities. The skill to be discovered is a mapping from this visual information to a 2D driving force vector. Simulation results show that the proposed method is successful in discovering the skill to catch balls and that the discovered skill is tough to some extent against such disturbances as the influence of the wind. Finally, the discovered skill is analyzed and found to be compatible with the strategy of a human ball catcher.
