Recognition of Assembly Tasks Based on the Actions Associated to the Manipulated Objects

摘要

This paper proposes a complete framework to automatically recognize assembly manipulation motions performed by humans, for the purpose of generating and retrieving robot motions from a database. Using the concept of affordance, we can obtain the relationship between the manipulated object and its associated human actions to narrow down the possible actions that each manipulated object can afford. Based on this relationship we design motion templates containing a set of basic motions associated to the manipulated objects and stored them in the database. Recognition of motion data is done by matching it with the existing motion templates on the database using Hidden Markov Models (HMMs). We verify the validity of the proposed method using three different assembly tasks performed by two subjects, which include basic assembly motions such as insertion and bolt screwing.