Task-grasping from a demonstrated human strategy

Task-grasping is a challenge in robot grasping because a higher-level understanding of the entire task-context is required for performing the grasp. Learning-from-observation (LfO) is a framework for robot teaching, where a demonstrator teaches manipulative operations as well as contexts. To utilize the LfO approach for the task-grasping problem, we classified grasps based on the force-exertion required in a subsequent task. The classification based on force-exertion was defined by observing grasps from both the human-end perspective and the robot-end perspective, and a lazy-closure was newly defined as one of the types. We demonstrated that one general policy per force-exertion-type is sufficient for handling different grasp shapes. Experimental results show that the appropriate grasp for a task sequence can be executed by obtaining the force-exertion-type from a one-shot human demonstration and then by executing the exertion policy. Real-robot execution results are shown in two task sequence scenarios: (1) picking up a cup and placing it right side up in a basket and (2) opening a refrigerator.