UniMate: One Unified Model to Animate Diverse Skeletons Across Topologies

Forum topic · 小凯 · 2026-09-08

Summary

UniMate is a unified foundation model for zero-shot cross-topology animation generation, introduced in a paper by Linzhan Mou and colleagues at Princeton (arXiv:2609.05415). Unlike prior methods locked to a single skeleton type, UniMate animates 7 skeleton topologies—bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid bodies—without fine-tuning. Its core innovation is a Topology-Aware Diffusion Transformer that embeds skeletal graph structure into attention via three mechanisms: graph-aware attention bias, spectral rotary position embedding based on the graph Laplacian, and a global topological conditioner. The model is trained on UniML3D, a new dataset of 13,006 motion sequences with text annotations across all 7 topologies. Experiments show zero-shot generalization to unseen topologies (e.g., generating coordinated spider or snake motion from a model trained only on bipeds and quadrupeds), outperforming specialized baselines in quality, generalization, and efficiency. Applications include in-betweening, animation expansion, and text-guided editing, suggesting broader implications for democratized animation and unified robot control.

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
Venue: SIGGRAPH Asia 2026
Field: Computer Vision / Graphics
Project page: https://linzhanmou.com/unimate/

Key contributions

UniMate is presented as the first unified foundation model for zero-shot cross-topology animation generation:

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 bakes graph structure directly into attention

Why existing methods fail

Skeletons are not sequences—they are graphs. Standard Transformers treat inputs as flat token sequences, so models are either topology-agnostic (losing structure), topology-fixed (one model per skeleton type), or require retraining for new skeletons. UniMate addresses this with three mechanisms:

1. Graph-Aware Attention Bias: adjusts attention weights using pairwise joint relations and geodesic distances on the skeleton graph, linking physically adjacent and functionally related joints (e.g., diagonal arm–leg coupling in locomotion). 2. Spectral Rotary Position Embedding: uses eigenvectors of the Graph Laplacian to produce a numbering-agnostic positional encoding—like GPS coordinates rather than street addresses—so functionally corresponding joints (human elbow, cheetah fore-knee) align across skeletons. 3. Global Topological Conditioner: extracts a "topology summary" from the rest-pose skeleton via attention pooling, conditioning the diffusion process on the creature type being animated.

UniMate as a diffusion model

UniMate is a diffusion model: it learns to denoise random joint configurations into coherent motion. The difference is that topological constraints (e.g., knees don't bend backwards, snakes have no legs) are enforced during denoising via the three mechanisms above, ensuring physically and topologically plausible output.

UniML3D dataset

13,006 motion sequences across the 7 topologies
Canonicalized to a standard coordinate system, joint naming, hierarchy, and frame rate
Each sequence paired with natural-language descriptions (e.g., "a cheetah sprinting at full speed"), enabling text-conditioned generation

Experimental results

Zero-shot transfer: trained only on bipeds and quadrupeds, the model generates plausible motion for birds, snakes, and insects—e.g., coordinated eight-legged spider crawling and serpentine slithering—without any fine-tuning
Quality: exceeds specialized single-topology SOTA methods in fluency, naturalness, and physical plausibility across all topologies
Generalization: remains stable on unseen topologies where baselines degrade sharply
Efficiency: no per-topology fine-tuning (which takes hours to days for baselines); faster generation

Application demos

In-betweening: automatic transition frames between keyframes (e.g., "standing" → "jumping")
Expansion: extending a short loop (2 walk steps) into long cycles (100 steps)
Text-guided editing: modify existing animation via instructions ("make this cat jump higher")

Limitations and future work

Generated motion is visually plausible but not strictly physically correct (no mass distribution, inertia, or ground reaction forces)—relevant for robotics applications
Output is skeleton animation only; secondary motion (muscle jiggle, fur/tail dynamics) requires extra processing
Single-character focus; multi-character and environment interaction remain open
Future directions: real-time generation (<16 ms for games/VR), fine-grained style control ("run tiredly", "jump happily"), and cross-modal extension (speech→lip sync, music→dance, video→mimicry)

Significance

UniMate aims to do for animation what Transformers and GPT did for NLP: unify across structures. It lowers the barrier for independent developers and small studios—text prompt plus a 3D model of any topology yields professional-quality animation. It also hints at embodied intelligence: by encoding graph structure and physical constraints, the model approximates an understanding of "how bodies move," pointing toward unified robot controllers that drive humanoids, quadruped robots, and manipulators with a single motion "brain."

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*.
SIGGRAPH Asia 2026
Project page: https://linzhanmou.com/unimate/
Dataset: UniML3D (13,006 motion sequences, 7 topologies)