Improvements on action parsing and action interpolation for learning through demonstration

Abstract

Programming humanoid robots with new motor skills through human demonstration is a promising approach to endowing humanoids with new capabilities in a relatively quick and intuitive manner. This paper presents an automated software system to enable our humanoid robot to learn a generalized dexterous motor skill from relatively few demonstrations provided by a human operator wearing a telemetry suit. Movement, end-effector, stereovision, and tactile information are analyzed to automatically segment movement streams along goal-directed boundaries. Further combinatorial selection of subsets of markers allows final episodic boundary selection and time alignment of tasks. The task trials are then analyzed spatially using radial basis functions [RBFs] to interpolate between demonstrations using the position of the target object as the motion blending parameter. A secondary RBF solution, using end-effector paths in the object coordinate frame, provides precise end-effector positioning and orienting relative to the object. Blending of these two solutions is shown to both preserve quality of motion while increasing accuracy and robustness of object manipulation.