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
• Field: Computer Vision / Graphics
• Project page: https://linzhanmou.com/unimate/

Key points

• UniMate is described as the first unified foundation model for zero-shot cross-topology animation generation, covering 7 skeleton topologies: bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid bodies.
• It enables zero-shot transfer to unseen skeletons without any fine-tuning or retraining, and reportedly outperforms state-of-the-art baselines in quality, generalization, and efficiency.
• Core innovation: a topology-aware diffusion transformer that injects graph structure directly into the attention mechanism.

The problem: attention does not understand joints

Existing animation generation methods are "locked" to a specific skeleton topology, for three main reasons:

1. Topology ignored: skeletons are flattened into sequences, losing structural information. 2. Topology fixed: dedicated models are built per skeleton type. 3. Fine-tuning required: new skeletons demand retraining.

Transformers treat inputs as sequences, but a skeleton is a graph—the "hand" of a human and the "front paw" of a cheetah are functionally equivalent yet may appear at entirely different sequence positions.

Method: making AI understand body structure

UniMate's topology-aware diffusion transformer relies on three mechanisms:

1. Graph-aware attention bias — attention weights are adjusted using pairwise joint relations and geodesic distances within the skeleton graph, so physically adjacent and functionally correlated joints (e.g., diagonal limb pairs) receive stronger connections. 2. Spectral rotary position embedding — eigenvectors of the graph Laplacian provide an indexing-agnostic positional representation, analogous to GPS coordinates rather than street addresses. 3. Global topological conditioner — a rest-pose skeleton is summarized via attention pooling into a global topology descriptor that conditions the diffusion process, letting the model know "what kind of creature" it is animating.

The generation backbone is a diffusion model: motion sequences are progressively noised, and the network learns denoising from random joint configurations to coherent motion—now constrained by topology.

UniML3D dataset

• 13,006 motion sequences across the 7 topologies.
• Canonicalization: unified coordinate frames, joint naming, hierarchies, and frame rates.
• Text pairing: every sequence has a natural-language description, enabling text-controlled generation.

Results

• Zero-shot cross-topology transfer: trained only on bipeds and quadrupeds, the model generates plausible animations for birds, snakes, and insects—e.g., coordinated eight-legged crawling for spiders it never saw during training.
• Quality: beats specialized single-topology methods across all topology types in smoothness, naturalness, and physical plausibility.
• Generalization & efficiency: baseline methods degrade sharply on unseen topologies and need hours-to-days of fine-tuning; UniMate stays stable and needs no adaptation.

Applications

• In-betweening: automatic interpolation between keyframes (e.g., "stand" to "jump").
• Expansion: extending a short clip into a longer loop (e.g., a 2-step walk into 100 steps).
• Text-guided editing: modifying existing animations with instructions like "make the cat jump higher".

Limitations and future work

• Generated motion is visually plausible but not guaranteed physically exact (mass, inertia, ground reaction forces are not modeled).
• Output is skeleton-level animation; secondary motion (muscle jiggling, cloth/skin) needs extra processing.
• Multi-character interaction and environment interaction remain open challenges.
• Future directions: real-time generation (<16 ms for games/VR), finer style control ("run tiredly"), and cross-modal generation (speech/music/video to motion).

Significance

By unifying all skeleton topologies in a single model, UniMate lowers the barrier to professional-quality animation for indie developers and small studios, and hints at a step toward general robot control—one "motion brain" adapting to arbitrary embodiments.

References

• 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*.
• Project page: https://linzhanmou.com/unimate/
• Dataset: UniML3D (13,006 motion sequences, 7 topologies)