Learning-based Dynamic Robot-to-Human Handover

TL;DR

This work tackles dynamic robot-to-human handover by allowing the receiver to move during the exchange and demonstrating that motion-adaptive strategies can improve efficiency and comfort. It introduces a learning-based pipeline that uses a nonparametric trajectory generator trained on 1,000 human–human handovers, combined with CoSpar-based preference learning to tune impedance parameters and release thresholds. The method integrates Cartesian impedance control to ensure safe, compliant interaction and uses a SPGP-based trajectory similarity measure to generate real-time robot motions. Through simulation and real-world experiments, including user studies with the UR5e, the approach shows faster handovers and higher user comfort compared to static methods, with the optimization of handover parameters further enhancing user satisfaction.

Abstract

This paper presents a novel learning-based approach to dynamic robot-to-human handover, addressing the challenges of delivering objects to a moving receiver. We hypothesize that dynamic handover, where the robot adjusts to the receiver's movements, results in more efficient and comfortable interaction compared to static handover, where the receiver is assumed to be stationary. To validate this, we developed a nonparametric method for generating continuous handover motion, conditioned on the receiver's movements, and trained the model using a dataset of 1,000 human-to-human handover demonstrations. We integrated preference learning for improved handover effectiveness and applied impedance control to ensure user safety and adaptiveness. The approach was evaluated in both simulation and real-world settings, with user studies demonstrating that dynamic handover significantly reduces handover time and improves user comfort compared to static methods. Videos and demonstrations of our approach are available at https://zerotohero7886.github.io/dyn-r2h-handover .

Figures (6)

• Figure 1: Overview of dynamic handover system. In A. Learning Phase, the robot trajectory generation module $\textit{g}(\xi^\text{r}_o;D)$, which predicts the robot trajectory on the observed receiver trajectory $\xi^{\text{r}}_{o}$, is trained using the human-to-human handover dataset $D$. In B. Fine-tuning Phase, preference learning adjusts the handover parameters $\boldsymbol{\Theta}$ including impedance variables, based on human feedback from real-world handover trials. Finally, in C. Dynamic Handover Phase, the system utilizes the learned $\textit{g}$ and ${\hat{\boldsymbol{\Theta}}}$ to select desired force $\mathbf{F}$ at end-effector in real-time, ensuring smooth and responsive handover performance.
• Figure 2: Human-Human data collection. 1,000 pairs of receiver-giver trajectory are obtained using the 3D body pose estimation and motion capture system.
• Figure 3: Sample efficiency. The values within each marker represent the RMS error (mean $\pm$ std) of the trajectory generation modules for different training data sizes.
• Figure 4: Relationship between RMS error and inference time. The values within each marker represent the RMS error and inference time for IAAL and varying inducing rates for the proposed method.
• Figure 5: Real world dynamic handover demonstrations. Following the instruction in Fig. \ref{['fig:instruction']}, real-world dynamic handover demonstrations in Fig. \ref{['fig:dynamic_handover']} are implemented with the UR5e in a real-world setting. In the dynamic handover scenarios, the left figure represents the ID scenario, the middle figure illustrates the OOD scenario with humans moving unpredictably, and the right figure shows the OOD scenario where humans alternate between stopping and moving.