Fast Human–Robot Handover Using a Vision-Based Pipeline

In this article, we present a fast and robust vision-based pipeline for human-to-robot object handovers. The perception module fuses data from two external RGB-D cameras to estimate the 3-D pose of the human hand and reconstruct a multiview point cloud of the object. From this, an axis-aligned 3-D bounding box is extracted and used by a geometry-based grasp planner that computes the grasp pose based on the user's hand and object size. The robot executes the grasp in a fast and reactive way using an event-driven velocity controller. The system has been validated in three real-world experiments: a user study that follows the CORSMAL benchmark, an evaluation on additional objects, and a cross-platform deployment on a different robotic arm. The proposed pipeline achieves high grasp success rates, low placement errors, and short execution times, outperforming existing methods.