A novel pipeline for bi-manual handover task

This paper addresses the problem of bi-manual object handover with a humanoid robot, i.e. the task of passing objects from one hand to the other. Bi-manual coordination is fundamental for improving manipulation capabilities of humanoid robots. We propose a novel and effective pipeline that tackles the problem by using visual and tactile feedback to localize the object and maintain grasp stability during the task. Given the object in one of the robot hand (first hand), the object in-hand pose is estimated by a localization algorithm, which makes use of vision and tactile information. Then, the estimated pose is used to automatically choose a suitable pose for the second hand among a set of candidates, a-priori annotated on the object model. The selected pose is finally used to accomplish the handover task. The performance of our approach is assessed on a real robotic system, the iCub humanoid robot, on a set of objects from the YCB dataset. Experiments demonstrate that the proposed method allows performing proper and reliable handovers with different every-day objects.