On Efficient and Flexible Autonomous Robotic Insertion Assembly in the Presence of Uncertainty

Abstract

This paper presents a general approach for autonomous tight-clearance assembly of complex-shaped parts using a general-purpose robot manipulator equipped with a force/torque (F/T) sensor. Such autonomous assembly is challenging in the presence of relative part pose uncertainty due to inaccuracies in part and robot modeling, sensing and perception, and robot motion control, especially when this uncertainty exceeds the clearance between assembly components. Our approach uses a geometry-invariant representation of each assembly part and integrates a learned uncertainty prediction model that is general and independent of specific parts into search and optimization to efficiently determine the best estimate of any contact configuration caused by misalignment between assembly parts and to ensure successful assembly. Our approach allows the manipulator to autonomously pick and assemble rigid parts at different workspace locations and accommodates different insertion directions from pickup directions. Our extensive experimental results demonstrate the effectiveness and efficiency of this approach in achieving successful assemblies of complex parts even though the pose uncertainty can be more than 10 times the task clearance between the assembled parts.