POMDP based robot teaching for high precision assembly in manufacturing automation

Abstract

Robot teaching is a necessary process for setting up a robot for a new task from a simulation environment to real environment or improving the robot performance with new batch of workpieces. Typically robot teaching process is performed by human beings with a teach pendant. This will increase the operational cost and reduce the efficiency. In this paper we present a framework for autonomous robot teaching using one camera without hand-eye calibration. A mobile robot with a camera is an “adult” performing teaching tasks in a production line. To overcome the problem brought by single uncalibrated vision sensor, the concept of “View Cone” is utilized to provide an effective observations of the underlying metric information. Geometrical model of the “View Cone” is built to describe the relationship between block property and the underlying metric information. The robot teaching process is modeled as a Partial Observable Markov Decision Process(POMDP) and solved by the Successive Approximation of the Reachable Space under Optimal Policies (SARSOP) algorithm. Simulations were performed and the results verify the effectiveness of the proposed method.