Our approach in programming a robot to perform a grasping task is to have the robot system replicate the task through observation. This is in line with the Assembly Plan from Observation (APO) paradigm [17] which results in lower requirements of programming expertise and higher ease of programming, since programming would be mostly reduced to directly demonstrating what needs to be done.

The first step in recognizing the grasping task is to identify the grasp itself. Once the grasp is identified, it has to be represented in a struc ture that both the system and human operator can interpret and under stand. We propose a grasp abstraction hierarchy for this purpose. It is important that the human operator be able to understand the results of the recognition by the system in order to intervene and perform diag nostic actions when necessary. The grasp hierarchy comprises the low-level, intermediate-level, and high-level descriptions of the grasp used in the task. This information would be used to select the grasp for the manipulator to effect the task at the task execution stage. Examples and experiments are presented to illustrate the proposed hierarchy. We also describe the relevance and relation of this research to future work on the other modules of our task recognition system.

[17] K. Ikeuchi and T. Suehiro, Towards an Assembly Plan from Observation: Task recognition with polyhedral objects, Tech. Rep. CMU-CS-91-167, Carnegie Mellon University, Aug. 1991.