RoMo dataset achieves state-of-the-art fidelity and diversity for motion generation

The long-standing tradeoff in 3D human motion generation—choosing between small, high-fidelity motion capture datasets and large-scale but noisy in-the-wild collections—has finally been addressed. Researchers led by Jiahao Zhang and 11 co-authors present RoMo, a rich, large-scale dataset that aggressively filters out static and artifact-prone sequences using a taxonomy-aware pipeline. Every sequence comes with detailed captions organized by a novel three-level semantic taxonomy, enabling fine-grained per-category evaluation that reveals model strengths and weaknesses obscured by global metrics.

Models trained on RoMo achieve state-of-the-art fidelity and diversity, with a superior grasp of complex, subtle text prompts. To further support reproducible research, the team released the Motion Toolbox, which standardizes metrics, data conversion, and visualization. Accepted at CVPR'26, RoMo establishes a foundation for interpretable and controllable human motion generation, with implications for animation, gaming, VR, and robotics.