Transfer of Human Skills to Robots: Learning from Human Demonstrations for Building an Adaptive Control System

Abstract

A new method for building a task-level robot adaptive controller is presented. Based on human demonstration data, an adaptive control law that elucidates skillful human behavior is extracted and used for building a robot control system. This allows us to transfer, directly from the human to the robot, the ability to adapt to unknown, varying environments by monitoring changes and modifying control parameters accordingly. A functional relationship is formulated and identified from demonstration data that represents the relationship between what the human monitors in the task environment and what changes in control are made in response to the change in the task environment. A critical problem in identifying this adaptation law is to determine what a human detects in the task environment in order to make a control decision. Namely, the input space of human perception must be determined. In this paper, an efficient method for determining a sufficient set of input variables is presented. The method, based on the Lipschitz quotient, allows us to detect any lack of essential quantities in the input space without assuming any model or representation of the input-output relationship. The method is applied to robotic deburring, in which feedrate and compliance must be varied depending on the task environment, e.g., burr size and hardness. Data acquired from human deburring demonstrations are analyzed by using the Lipschitz quotient.