INFERRING OBJECT WEIGHT FROM HUMAN HANDOVER KINEMATICS: INSIGHTS FOR ADAPTIVE HUMAN-ROBOT HANDOVERS

Abstract Object handovers are a fundamental component of both human-human and human-robot interaction, where motion is continuously adapted based on object properties such as weight, size, shape, and fragility. Understanding these motion patterns is critical for designing adaptive and intuitive robotic systems. In this work, we investigate the problem of classifying object weight from kinematic features extracted from the YCB (YaleCMU-Berkeley) Handovers dataset, which contains 2771 segmentedhandoverinstances spanning 27 objects with massesfrom 8 g to 2171 g. Weevaluateboththree-class(light, medium, heavy) and two-class (light vs. heavy) classification formulations using five supervised learning models: Random Forest, Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and k-Nearest Neighbours (k-NN). A central contribution of this work is a systematic analysis of how weight-related information is distributed across interaction roles. We evaluate giver-only, taker-only, and combined giver-taker feature representations. Under train/test evaluation, combined features achieve the best accuracy across all conditions, with XGBoost reaching 82.16% on the two-class task and 61.62% on the three-class task. Under the more rigorous Leave-One-Pair-Out (LOPO) cross-validation, giver-only features prove more robust in macro-F1, with the three-class generalisation gap widening to approximately 6 percentage points relative to train/test, highlighting the participant-dependent nature of interaction-level features.