What it means for an object to be 'within reach' depends very much on the morphology and skills of a robot. In this paper, we enable a mobile manipulation robot to learn a concept of PLACE from which successful manipulation is possible through trial-and-error interaction with the environment. Due to this developmental approach, PLACE is very much grounded in observed experience, and takes the hardware and skills of the robot into account. During task-execution, this model is used to determine optimal grasp places in a least-commitment approach. This PLACE takes into account uncertainties in both robot and target object positions, and leads to more robust behavior.
展开▼
