Robot programming by human demonstration

Abstract

A method for programming a robotic manipulator by human demonstration is proposed. The human performs the task using a Human Interface Device and a computer monitors the human action and recognizes the task performance strategy. This method of programming is made possible by the recent availability of cheap, light and unencumbering six-degree-of-freedom position and force sensors. A method for learning a task performance strategy from a human demonstrator is developed. A task is modeled as a series of piecewise linear compliant control subtasks. Switching between subtasks is initiated on sensing the termination condition for the current subtask. A least-squares method is developed for recognizing the piecewise linear subtasks, and for characterizing the subtask as a compliant control move with a virtual trajectory. The termination conditions for each subtask is recognized as a sensor that consistently reaches a certain value when the subtask is accomplished. Termination conditions are searched for in the region between the corner of the virtual trajectory and the actual trajectory. Experimentally the method for recognizing task strategies from a human demonstration for a simple constraint following task have been found to be 80% consistent.