ContactHandover: Contact-Guided Robot-to-Human Object Handover

TL;DR

The paper addresses robot-to-human object handover by introducing ContactHandover, a two-phase system that (1) predicts 6-DoF grasp poses and a 3D human contact map to choose grasp poses that leave human-reachable contact regions unobstructed, and (2) computes a handover pose that minimizes the human arm torque and displacement while aligning contact points with the receiver’s view. It introduces a contact-guided grasp selection mechanism and a delivery-phase pose optimization, guided by two ergonomic metrics, visibility and reachability. Evaluations on 27 household objects show that ContactHandover achieves higher handover quality than ablations, with an average success rate of 68.5% and demonstrated generalization to unseen objects. The work advances naturalistic human-robot collaboration by leveraging 3D contact affordances to tailor both grasp and delivery to human preferences, with implications for safer and more intuitive handovers.

Abstract

Robot-to-human object handover is an important step in many human robot collaboration tasks. A successful handover requires the robot to maintain a stable grasp on the object while making sure the human receives the object in a natural and easy-to-use manner. We propose ContactHandover, a robot to human handover system that consists of two phases: a contact-guided grasping phase and an object delivery phase. During the grasping phase, ContactHandover predicts both 6-DoF robot grasp poses and a 3D affordance map of human contact points on the object. The robot grasp poses are re-ranked by penalizing those that block human contact points, and the robot executes the highest ranking grasp. During the delivery phase, the robot end effector pose is computed by maximizing human contact points close to the human while minimizing the human arm joint torques and displacements. We evaluate our system on 27 diverse household objects and show that our system achieves better visibility and reachability of human contacts to the receiver compared to several baselines. More results can be found on https://clairezixiwang.github.io/ContactHandover.github.io

Figures (5)

• Figure 1: Contact-Guided Robot to Human Object Handover. We propose a robot-to-human handover system with two phases: (a) contact-guided grasping and (b) object delivery. (a) During grasping, the robot predicts 6-DoF grasp poses and human contact points (denoted in red) for the object, and selects a grasp pose that maximizes stability while minimizing contact points occlusions. (b) During delivery, the robot computes a handover location and orientation that minimizes human arm joint torque and displacements, as well as the distance between contact points and the human.
• Figure 2: Contact-Guided Grasp Selection. (a)§ \ref{['method:human_contact']} The robot takes RGB-D Images from 16 views around the table and construct a $64^3$ voxel representation of the object via TSDF fusion; the occupancy grid is then fed into a trained 3D VoxNet to predict human contact maps. (b)§ \ref{['method:robot_grasp']} The robot takes a partial point cloud observation as input to the pre-trained Contact-GraspNet model to generate a set of 6-DoF robot grasps. (c)§ \ref{['method:GF']} The robot executes the grasp with highest score as computed by Equation \ref{['eq:grasp_filter']}.
• Figure 3: Qualitative Results and Ablations. As shown in (a), ContactHandover predicts the human contact map (red indicate human contact points, and blue non-contact points), picks up the objects while avoiding human contacts, and orients the human preferred contacts towards the human during delivery. In (b), without grasp re-ranking, the robot gripper blocks human contacts, i.e., the handle of pan and hammer. In (c), without handover orientation, the human contacts, i.e. the handles of the scissors, hammer, and pan, points away from the human. More qualitative results can be found on the https://clairezixiwang.github.io/ContactHandover.github.io/.
• Figure 4: Generalize to Unseen Objects. We show ContactHandover's performance on unseen YCB objects. We show both human contact predictions and handover results on these objects. ContactHandover is able to generalize to (a) objects with unseen shapes (e.g. hammer and knife) and (b) objects with unseen types (e.g. spoon, flat screwdriver, fork and phillips screwdriver). It predicts reasonable human contacts (denoted in red points) around the handles of the objects, picks up and delivers the objects to human with respect to the predicted human contacts.
• Figure 5: Clustering bimodal human contacts. For objects with bimodal human contact distributions, ContactHandover clusters the contact points and only optimizes one cluster. In (a) the robot orients one side of the binoculars towards the human. Without clustering, as in (b), both sides are pointed to the human, leaving neither cluster close to the human.