Establishing baselines for dyadic visual motion prediction using the QUB-PHEO dataset

We present a fully reproducible benchmark protocol for short-horizon motion prediction in collaborative assembly using the QUB-PHEO dataset. The protocol fixes data splits, coordinate normalisation, evaluation metrics and training/evaluation scripts, enabling like-for-like comparison and easy extension to new models or modalities. Comparisons across three input regimes (hands only, hands + surrogate + gaze, and full 214-D aerial context) using standardised metrics reveal how architectural choices and added context affect robustness, beyond headline accuracy. All code, data splits, and pretrained weights are released, validating QUB-PHEO’s suitability for vision-based motion prediction and providing a standardised baseline for the dyadic HRI community.