Teaching a robot manipulation skills through demonstration

Abstract

An automated software system has been developed to allow robots to learn a generalized motor skill from demonstrations given by a human operator. Data is captured using a teleoperation suit as a task is performed repeatedly on Leonardo, the Robotic Life group's anthropomorphic robot, in different parts of his workspace. Stereo vision and tactile feedback data are also captured. Joint and end effector motions are measured through time, and an improved Mean Squared Velocity [MSV] analysis is performed to segment motions into possible goal-directed streams. 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 demonstrations to span his workspace, using the object position as the motion blending parameter. An analysis of the motions in the object coordinate space [with the origin defined at the object] and absolute world-coordinate space [with the origin defined at the base of the robot], and motion variances in both coordinate frames, leads to a measure [referred to here as objectivity] of how much any part of an action is absolutely oriented, and how much is object-based. A secondary RBF solution, using end effector paths in the object coordinate frame, provides precise end-effector positioning relative to the object. The objectivity measure is used to blend between these two solutions, using the initial RBF solution to preserve quality of motion, and the secondary end-effector objective RBF solution to increase the robot's capability to engage objects accurately and robustly.