RoMo is a large-scale, curated in-the-wild 3D human motion dataset with taxonomy-aware filtering and hierarchical annotations for fine-grained evaluation. It enables state-of-the-art motion generation with improved fidelity, diversity, and text understanding.

We introduce Manual-PA, a transformer-based framework that leverages diagrammatic assembly manuals to guide both the selection and 6D pose estimation of furniture parts, enabling efficient and realistic 3D assembly by aligning parts with instructional illustrations.

We introduce a new approach to simultaneously localize a sequence of instructional diagrams in videos by modeling their mutual relationships and temporal order, rather than handling each step independently.

We introduce a new framework and dataset (IAW) for aligning assembly diagrams from instruction manuals with real-world video segments, enabling cross-modal retrieval and step-level alignment between illustrated instructions and assembly actions in videos.