A geometrically validated approach to autonomous robotic assembly

Abstract

The paper discusses the employment of different sources of information to support robotic assembly operations. During component interaction, part of the wrist sensed force and torque information is found to be geometrically dependent. This enables the real-time sensorial data retrieved from the assembly scene to be combined with the information on the geometry of the component and the history of the insertion itself. As a result, an intelligent control architecture is developed to perform simple peg-hole assembly operations emphasising the aspects which relate to learning an appropriate state-action mapping without requiring an a priori defined set of manipulative skills. A real time peg in hole experiment involving a PUMA 761 industrial manipulator is detailed to support the theoretical results.