Implementation of a framework for learning handover grasp configurations through observation during human-robot object handovers

As humanoids work alongside people, there will be many situations where they need to handover objects to people. If humanoids are to fulfill their purpose effectively, it is imperative that they perform handovers properly. When handing over an object, the giver needs to determine where to grasp and how to orient the object properly in order to ensure a safe and efficient handover. We propose and implement a framework for automatically learning handover grasp points and orientations - which we refer to collectively as the handover grasp configuration - by observing how people hand over the objects to the robot. We achieve this using a skeleton tracker and a particle filter based object tracker. Our system requires no additional external cameras, or any markers on the person or the object. As far as we know, this is the first system that offers such capabilities for learning handover grasp configurations. An implementation on an HRP2V robot and an experiment with three different objects verified that our framework is capable of extracting and learning grasp configurations from handover demonstrations, and subsequently using the learned grasp configurations to handover the objects.