Safe robot affordance-based grasping and handover for Human-Robot assistive applications

A crucial aspect of human-robot collaboration involves the robot's ability to safely perform handovers of objects to be used by the operator. In this study, we introduce two complementary frameworks designed to execute every stage of a robot-to-human handover process, using only RGB-D input data. The first framework employs a machine learning model, trained on a custom real-world dataset, to detect objects and their parts' affordances. Affordance is encoded using a novel representation based on keypoints, which are utilized to model hazardous sections of objects and plan appropriate grasps. The second framework tracks hand movements to dynamically determine handover locations, while enforcing safety protocols to prevent exposure of dangerous object parts during the movement. Additionally, it ensures correct object orientation upon delivery, presenting the object handle to the human. The effectiveness of the proposed solution has been successfully validated through testing on a real mobile manipulator.