UniMate: One Unified Model to Animate Diverse Skeletons
Paper info
- Title: UniMate: One Unified Model to Animate Diverse Skeletons
- Authors: Linzhan Mou, Jiahui Lei, Zhiyang Dou, Chenyue Cai, Chaoyue Song, Adam Finkelstein, Szymon Rusinkiewicz
- arXiv ID: 2609.05415
- Venue: SIGGRAPH Asia 2026
- Area: Computer Vision / Graphics
- Supports 7 skeleton topologies: bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid bodies
- Zero-shot cross-topology transfer: no fine-tuning or retraining needed for new skeletons
- Outperforms state-of-the-art baselines in quality, generalization, and efficiency
- Core innovation: a topology-aware diffusion transformer that injects graph structure directly into attention
- Human skeleton: spine → pelvis → two legs/arms → hands and feet
- Cheetah skeleton: spine → four legs → tail
- Snake skeleton: a chain of vertebrae, no limbs
- Spider skeleton: a central body → eight legs
- 13,006 motion sequences across 7 topologies: 1. Bipedal (humans, humanoids) 2. Quadrupedal (dogs, cats, horses, cheetahs) 3. Avian (eagles, pigeons, chickens) 4. Marine (fish, dolphins) 5. Insectoid (spiders, ants, scorpions) 6. Serpentine (snakes, eels) 7. Articulated rigid (robot arms, robots)
- Canonicalization: unified coordinate systems, naming, hierarchies, and frame rates
- Text pairing: each sequence has a natural-language description, enabling text-conditioned generation
- Quality: surpasses specialized single-topology methods on all topology types in fluency, naturalness, and physical plausibility
- Generalization: existing methods degrade sharply on unseen topologies; UniMate remains stable
- Efficiency: no hours-to-days of per-topology fine-tuning; faster generation
- In-betweening: automatic transition frames between key poses
- Expansion: extending a short clip into a long loop (e.g., 2 walking steps → 100)
- Text-guided editing: modify existing animations via instructions like "make this cat jump higher"
- Democratizing animation: professional-quality motion without expensive mocap rigs or specialist teams — text description plus a 3D model in, animation out
- Embodied intelligence: by encoding graph structure and physical constraints, UniMate approximates a "body intuition" that AI normally lacks
- Toward generalist robot control? If one model can generate motion for many morphologies, it hints at a unified "motion brain" shared by humanoids, quadruped robots, and manipulators
- Visual plausibility does not guarantee full physical correctness (mass distribution, inertia, ground reaction forces are not modeled)
- Only skeletal animation; secondary motion (muscle jiggle, skin deformation) requires additional processing
- Single-character focus; multi-character and environment interaction remain open
- Real-time generation for games/VR (currently seconds to minutes per generation)
- Finer style control ("running tiredly," "jumping joyfully")
- Cross-modal extensions: speech-driven lip sync, music-driven dance, video-based motion imitation
- Mou, L., Lei, J., Dou, Z., Cai, C., Song, C., Finkelstein, A., & Rusinkiewicz, S. (2026). UniMate: One Unified Model to Animate Diverse Skeletons. *arXiv preprint arXiv:2609.05415*.
- SIGGRAPH Asia 2026
- Project page: https://linzhanmou.com/unimate/
- Dataset: UniML3D (13,006 motion sequences, 7 topologies)
---
Key Findings
UniMate proposes the first unified foundation model for zero-shot cross-topology animation generation:
---
Background: Topology as an Invisible Shackle
In computer graphics, topology describes an object's connection structure:
Existing animation methods are locked to specific topologies. The root cause lies in the Transformer's self-attention mechanism: skeletons are not sequences but graphs. Current approaches either flatten skeletons into sequences (losing structure), build per-topology specialized models (not general), or require retraining for new skeletons (impractical).
---
Method: Teaching AI to Understand "Body Structure"
UniMate's Topology-Aware Diffusion Transformer uses three key mechanisms:
1. Graph-Aware Attention Bias
Attention weights are adjusted based on pairwise joint relations and geodesic distances in the skeleton graph. Physically adjacent joints get stronger attention connections; functionally related joints (e.g., diagonal limb pairs in gaits) are also linked. Traditional Transformers see a skeleton as a string of independent tokens; UniMate sees it as a body that knows "the shoulder connects to the arm" and "the left leg mirrors the right."2. Spectral Rotary Position Embedding
Instead of sequence-index positional encodings, UniMate uses the graph Laplacian: its eigenvectors reveal the skeleton graph's intrinsic structure and provide a numbering-independent positional representation. The analogy: traditional positional encodings are like street addresses (change cities and they're meaningless), while spectral encodings are like GPS coordinates — valid everywhere.3. Global Topological Conditioner
Global topology information is extracted from the rest-pose skeleton and compressed via attention pooling into a "topology summary." This conditions the diffusion process so the model knows which kind of body it is animating — crucial since "walking" means entirely different things for a human, cheetah, spider, or snake.Diffusion Backbone
UniMate follows the standard diffusion paradigm: noise is added to motion sequences in a forward process, and a network learns to denoise random joint configurations into coherent motion. UniMate's innovation is enforcing topological constraints during denoising, ensuring the generated motion is physically and topologically plausible (knees don't bend backwards; snakes have no legs).---
UniML3D Dataset
To train UniMate, the authors built UniML3D:
---
Results
Zero-Shot Cross-Topology Transfer
Trained only on bipedal and quadrupedal data, UniMate generates reasonable animations for birds, snakes, and insects at test time — e.g., coordinated eight-legged spider crawling or serpentine slithering — without any fine-tuning.Comparisons with SOTA
Applications
---
Broader Significance
---
Limitations and Future Directions
Current limitations:
Future directions:
---
Conclusion
The phrase "One Unified Model" in the title echoes a recurring pattern in AI history: Transformers unified NLP architectures, GPT unified NLP tasks, Stable Diffusion unified image generation. UniMate aims to do the same for skeleton topologies — suggesting that the mathematics of motion may be largely independent of the body that performs it.
---