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
- Published: 2026-09-04
- Field: Computer Vision / Graphics
- Supports 7 skeleton topologies: bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid bodies
- Zero-shot cross-topology transfer: no finetuning or retraining needed for new skeletons
- Outperforms existing state-of-the-art baselines in quality, generalization, and efficiency
- Key innovation: a topology-aware diffusion transformer that incorporates graph structure directly into the attention mechanism
- 13,006 motion sequences
- Covering 7 skeleton topologies: bipedal (humans, humanoids), quadrupedal (dogs, horses, cheetahs), avian (eagles, pigeons), marine (fish, dolphins), insectoid (spiders, ants, scorpions), serpentine (snakes, eels), articulated rigid (robot arms, machines)
- Canonicalization: unified coordinate systems, skeleton naming, hierarchies, and frame rates
- Text pairing: each sequence has a natural-language description, enabling text-conditioned generation
- Trained only on bipedal and quadrupedal data, UniMate generates plausible animations for birds, snakes, and insects without any finetuning — e.g., coordinated eight-legged crawling for an unseen spider, or serpentine slithering for an unseen snake.
- 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 per-topology finetuning; faster generation than methods requiring hours-to-days of adaptation
- In-betweening: automatic interpolation between keyframes
- Animation expansion: extending short clips into longer loops (e.g., a 2-step walk into 100 steps)
- Text-guided editing: modifying existing animation via instructions (e.g., "make the cat jump higher")
- Democratizing animation: professional-quality motion without motion-capture rigs or specialist teams
- Embodied intelligence: by learning graph structure and physical constraints, the model approximates a form of "body intuition"
- Toward universal robot control: a single motion model that adapts to humanoids, quadruped robots, and manipulators suggests a path to unified controllers
- Visual plausibility does not guarantee full physical correctness (mass, inertia, ground reaction forces are not explicitly modeled)
- Output is skeletal animation; secondary motion (muscle, fat, skin deformation) requires additional processing
- Multi-character interaction and environment interaction remain open challenges
- Real-time generation (sub-16ms) for games and VR
- Finer style control ("tired running", "joyful jumping")
- 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)
Core Contribution
UniMate is described as the first unified foundation model for zero-shot cross-topology animation generation:
Background: The Topology Bottleneck
Different creatures have different skeletal connectivity structures (topologies): a human has a spine, pelvis, two legs and two arms; a cheetah has a spine, four legs and a tail; a snake is a chain of vertebrae with no limbs; a spider has a central body with eight legs. Existing animation methods are typically locked to a specific topology:
1. Ignoring topology: flattening skeletons into sequences loses structural information 2. Fixed topology: specialized models per skeleton type do not generalize 3. Finetuning required: retraining for every new skeleton is impractical
The root cause lies in the Transformer's self-attention, which treats inputs as sequences — but skeletons are graphs, not sequences.
Method: Topology-Aware Diffusion Transformer
UniMate embeds the skeleton's graph structure into attention via three mechanisms:
1. Graph-Aware Attention Bias: attention weights are adjusted using pairwise joint relations and geodesic distances in the skeletal graph, so physically adjacent and functionally correlated joints (e.g., diagonal limb pairs in a gait) receive stronger connections. 2. Spectral Rotary Position Embedding: eigenvectors of the graph Laplacian provide a numbering-agnostic positional representation — functionally corresponding joints (a human elbow vs. a cheetah's fore-knee) receive consistent encodings regardless of their index in a sequence. 3. Global Topological Conditioner: a rest-pose skeleton is summarized via attention pooling into a global "topology summary" that conditions the diffusion process, letting the model know what kind of body it is animating.
UniMate is a diffusion model: it learns to iteratively denoise random joint configurations into coherent motion, with topology constraints ensuring physically and structurally plausible results (e.g., knees do not bend backwards, snakes have no legs).
UniML3D Dataset
To train UniMate, the authors built UniML3D:
Experimental Results
Zero-Shot Cross-Topology Transfer
Comparisons with SOTA
Applications
Broader Implications
Limitations and Future Work
Limitations
Future directions
Conclusion
The title's phrase "One Unified Model" echoes historical AI unifications — Transformer in NLP, GPT across tasks, Stable Diffusion in image generation. UniMate attempts the same for animation: making skeleton topology no longer a barrier. It hints that motion, whatever body performs it, may share a unified mathematical structure.