Learning Grasp Affordances Through Human Demonstration

When presented with an object to be manipulated, a robot must identify the available forms of interaction. How might an agent acquire this mapping from object representation to action? In this paper, we describe an approach that learns a mapping from objects to grasps from human demonstration. For a given object, the teacher demonstrates a set of feasible grasps. We cluster these grasps in terms of the position and orientation of the hand relative to the object. Individual clusters in this pose space are represented using probability density functions, and thus correspond to variations around canonical grasp approaches. Multiple clusters are captured through a mixture distribution-based representation. Experimental results demonstrate the feasibility of extracting a compact set of canonical grasps from the human demonstration. Each of these canonical grasps can then be used to parameterize a reach controller that brings the robot hand into a specific spatial relationship with the object.