Extracting human strategies for use in robotic assembly

In many circumstances, it would be valuable to extract the human operator's assembly strategy and implement that strategy on a robotic system. However, it remains unclear how to analyze the human data quantitatively to extract the strategy, or how to represent that strategy on the robotic system. We assume here that the way humans approach a problem like an assembly task is by using a series of uniform motions terminated by events, which we refer to as "behaviors." We further assume that the sequence of behaviors will be a simple linear progression without loops, which eliminates the possibility of limit cycles developing from the sequence of behaviors. Based on these assumptions, which we call the "behavioral framework," we have developed methodologies for analyzing records of human assemblies and extracting the human's assembly strategy. Experiments have shown that the behavioral framework and the methodologies based on it can be used to create successful algorithms applicable to several different assembly tasks, including square and circular peg-in-hole assemblies with large initial position error, block-in-corner assembly, and the assemblies of multiple convex shapes with very small clearance and small initial position error. The resulting algorithms are comparable to the original human assemblies in terms of success rate, assembly times, and peak forces and moments, and have assembly times superior to those of a benchmark blind search algorithm.