UniMate: one unified model animates any 3D skeleton from text, no retraining needed
victormustar · x · 2026-11-05
- UniMate (SIGGRAPH Asia 2026) is a unified text-driven animation model: feed a rigged 3D asset plus a text prompt, and it generates motion without per-skeleton retraining.
- It handles diverse skeleton topologies — humans, animals, and articulated objects — supporting 5–85 joints. The core models use flow matching with graph attention (AdaLN text conditioning), 74M params, trained on the UniML3D dataset; dedicated checkpoints exist for Mixamo (22-joint) and Truebones rigs (73 of 74 species, all but Dragon).
- Weights, paper (arXiv:2609.05415), code, dataset, and an interactive demo are available on Hugging Face under CC-BY-NC-4.0. The poster reports the rigging results "working extremely well" in real tests.
More from Embodied
- Yacine probes mingru hidden states to check if domain randomization captured the real world — yacineMTB · 2026-11-05
- Yacine shares trick: probe mingru hidden states to check domain randomization coverage — yacineMTB · 2026-11-05
- Agility lands on Nasdaq as the first pure-play humanoid robot company to go public — TechCrunch AI · 2026-11-05
- Investor lugs a one-hand-carryable quadruped robot around Buenos Aires as impromptu strength training — StewartalsopIII · 2026-11-05
- China blocks SoftBank's $5.4bn acquisition of ABB's robotics unit — pstAsiatech · 2026-11-05
- Robot Learning's Bottleneck? Insiders Point to Scalable Evaluation in New Community Initiative — micoolcho · 2026-11-05