Recognition of Assembly Motions based on Manipulated Object Affordances

Abstract

In this work, we propose a recognition system for assembly motions performed by humans, with the aim of generating and re-using robot motions stored in a database. An affordance model for each manipulated object is proposed to efficiently search for matching actions. Based on this model we design motion templates containing a set of basic motions associated to the manipulated objects. Recognition of motion data is done using the proposed motion templates and Hidden Markov Models. We verify the validity of the proposed method using three different assembly tasks.