Imitation for motor learning on humanoid robots

Abstract

Motor control in humanoid robots is a challenging task due to the high number of degrees of freedom that must be dealt with; most solutions presented for the motion control area rely, heavily, on domain specific information about the robot or the concrete motor task to be developed, which makes it difficult or not viable to scale up in most cases. Learning by demonstration arises as an easier alternative for programming motor skills in robots, but, until now most of the proposed architectures are only validated using motor tasks which do not compromise the robot stability. This paper presents our ongoing work, the development of an easy to scale learning by demonstration architecture is proposed and validated using the gait of a biped robot as the skill to learn. Our proposed solution takes the motion captured information from a group of human masters executing the task as its input, and the output is computed by a genetic algorithm based on a joint position time series. Beginning with this, the robot can reproduce the assigned task using its own motor repertoire.