We study human perceptions of a robot that performs robot-to-human (R2H) handovers controlled to grasp, transport, and transfer 34 objects by mimicking human givers in human-human (H2H) handover data. Recognizing the importance of human-like robotic behavior for successful collaboration, R2H studies use models of human behavior or observations of H2H data to plan robot giver motion. However, R2H studies have been limited in object counts. In this work, we use the Human-Object-Human (HOH) dataset, consisting of H2H interactions performed by 20 giver-receiver pairs with 136 objects, to conduct an R2H study with 34 objects. We teleoperate a Kinova Gen3 manipulator to grip an object as grasped by an HOH human giver, and program it to automatically transport and orient the object to a participant by mimicking the HOH giver's trajectory and transfer pose. We survey participants on safety, naturalness, and preferred choice over linear trajectory and random orientation baselines. We find that transfer pose influences perceptions of naturalness, with HOH poses showing higher naturalness ratings. Participants prefer handovers with HOH end poses when asked to pick their preferred interaction.

In this work, we study patterns in the movement times of human participants engaging in object handover during the pre-handover phase, to inform on conducting fluent human-robot handovers. Fluency of object handover between two agents is critical to ensure success of shared overall collaborative goals in the context of larger tasks involving the transferred objects. Human givers and receivers often coordinate their pre-handover movements to ensure fluent transfer, with receivers showing anticipatory behavior and proactive response. Since typical timing patterns of human and robot movements for short-range tasks demonstrate a speedup to a peak speed, and a slowdown to the end point, we analyze relationships between time differences of peak speed attainment relative to start and end (i.e., reach) times, across the giver and receiver. Our work helps inform the design of motion planning algorithms for robots that embody timing patterns during movement that are predictable to the partner.

As robots increasingly integrate into human environments, they must interact safely, intuitively, and effectively. We present HI-Grasp, a novel human-inspired grasp network that generates robotic grasps that closely mimic human grasp behaviors observed in human-to-human object handovers while ensuring stability. Our approach combines a deep grasp prediction network with Transformer-based modules to predict human-like, stable grasps. We contribute the dataset HOH-Grasps to train and evaluate HI-Grasp. HOH-Grasps consists of interaction data from the HOH handover dataset annotated with grasp labels, and is available at https://huggingface.co/datasets/tars-home/HOH-Grasps. Using HOH-Grasps, HI-Grasp learns and reproduces human grasp preferences while maintaining grasp stability. Extensive experiments on HOH-Grasps, along with real-world robot tests, show that HI-Grasp outperforms baselines and ablated variants in terms of human alignment and stability. We propose HI-Grasp-Lift, a robot-to-human object handover strategy built on HI-Grasp's predictions, to showcase the practical applicability of our approach.

The fluency of handover between two agents is important to ensure safety and success of handover. In this work, we study the relationships between the timings of giver and receiver motions in human-human handover interactions, in order to inform human-robot handover. We use giver and receiver hand trajectories from the Human-Object-Human (HOH) handover dataset to study movement during the pre-handover phase, prior to the point of transfer. We find that human receivers adopt a largely proactive behavior, and plan and start motion early in the pre-handover phase. We also find that human receivers spend much of their motion moving in coordination with the giver, rather than after the giver has reached the transfer point. Further, we find that human receivers may predict future movement of the giver from early giver motion, and adjust their start times accordingly to ensure coordinated grasp at transfer. Our findings suggest that robot receivers should adopt a predictive giver-aware approach to plan motion early, and robot givers should recognize that human receivers may expect giver behavior to be human-like and predictable.

We present GraspPC, an approach to perform learning-based synthesis of multiple human hand grasps as point clouds from point clouds of objects. GraspPC benefits human-robot handover approaches by providing hypotheses of human grasp on objects to inform robotic manipulation algorithms on how to bias robotic grasp for safe handover. Existing learning-based approaches to conduct hand grasp prediction require datasets to contain annotated articulated hand models, making them difficult to train on datasets that lack hand model annotations. GraspPC treats the problem of hand point cloud generation from object point clouds as a set-to-set translation problem. We contribute a Transformer architecture to synthesize point clouds via GraspPC. To generate diverse hand grasps, we generate multiple object-dependent queries and train the network using a winner-takes-gradient strategy. We show results of diverse grasps by training and testing on a variety of real-world datasets. We demonstrate how human grasps generated by GraspPC can be used to filter robotic grasp candidates to inform human-robot handover. Our code is available at: https://github.com/Terascale-All-sensing-Research-Studio/GraspPC.

We present the HOH (Human-Object-Human) Handover Dataset, a large object count dataset with 136 objects, to accelerate data-driven research on handover studies, human-robot handover implementation, and artificial intelligence (AI) on handover parameter estimation from 2D and 3D data of two-person interactions. HOH contains multi-view RGB and depth data, skeletons, fused point clouds, grasp type and handedness labels, object, giver hand, and receiver hand 2D and 3D segmentations, giver and receiver comfort ratings, and paired object metadata and aligned 3D models for 2,720 handover interactions spanning 136 objects and 20 giver-receiver pairs—40 with role-reversal—organized from 40 participants. We also show experimental results of neural networks trained using HOH to perform grasp, orientation, and trajectory prediction. As the only fully markerless handover capture dataset, HOH represents natural human-human handover interactions, overcoming challenges with markered datasets that require specific suiting for body tracking, and lack high-resolution hand tracking. To date, HOH is the largest handover dataset in terms of object count, participant count, pairs with role reversal accounted for, and total interactions captured.