UniMate: A Unified Foundation Model for Zero-Shot Skeletal Animation Across Diverse Topologies

Forum topic · 小凯 · 2026-09-08

Summary

UniMate (arXiv:2609.05415, SIGGRAPH Asia 2026) is presented as the first unified foundation model for zero-shot cross-topology animation, covering seven skeleton types: bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid bodies. Its core innovation is a topology-aware diffusion transformer that embeds skeletal graph structure directly into the attention mechanism through three components: graph-aware attention bias using pairwise relations and geodesic distances, spectral rotary position encoding based on graph Laplacian eigenvectors, and a global topological conditioner that summarizes the rest-pose skeleton as a conditioning signal. Training uses UniML3D, a new dataset of 13,006 motion sequences normalized and paired with text descriptions. Experiments show zero-shot transfer to unseen topologies (e.g., generating coordinated spider locomotion after training only on bipeds and quadrupeds), outperforming specialized baselines in quality, generalization, and efficiency, with no fine-tuning required. Applications include in-betweening, animation expansion, and text-guided editing. Limitations include physical accuracy, secondary motion, and multi-character interaction. The authors suggest this unified approach could pave the way toward general-purpose robot locomotion control. Project page: https://linzhanmou.com/unimate/

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 (SIGGRAPH Asia 2026)
Area: Computer Vision / Graphics
Project page: https://linzhanmou.com/unimate/

Key points

UniMate is claimed to be the first unified foundation model for zero-shot cross-topology animation generation, supporting 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.
It reportedly outperforms existing SOTA baselines in quality, generalization, and efficiency.
Core innovation: a Topology-Aware Diffusion Transformer that injects graph structure directly into attention.

Motivation

Existing animation generation models are locked to specific skeleton topologies. The Transformer's self-attention treats inputs as sequences, but skeletons are graphs: a human hand and a cheetah's forepaw are functionally corresponding joints that may occupy entirely different sequence positions. Prior approaches either flatten skeletons into sequences (losing structure), build per-topology models (no generality), or require fine-tuning for each new skeleton (impractical).

Method: three key mechanisms

1. Graph-aware attention bias — adjusts attention weights using pairwise joint relations and geodesic distances on the skeleton graph, strengthening connections between physically adjacent and functionally related joints (e.g., diagonal limbs during locomotion). 2. Spectral rotary position embedding — uses graph Laplacian eigenvectors to provide a topology-invariant positional representation, analogous to GPS coordinates rather than street addresses; a human elbow and a cheetah fore-knee map to corresponding positions despite different sequence indices. 3. Global topological conditioner — extracts global topology from the rest-pose skeleton via attention pooling, producing a "topology summary" that conditions the diffusion process on the type of creature being animated.

The model is a diffusion-based generator: it learns to denoise random joint configurations into coherent motion, with topology constraints ensuring physical and structural plausibility (e.g., knees bend the right way, snakes have no legs).

Dataset: UniML3D

13,006 motion sequences across 7 topologies (bipedal, quadrupedal, avian, marine, insectoid, serpentine, articulated rigid).
Canonicalization: unified coordinate frames, skeleton naming/hierarchy, and sampling rates.
Text pairing: each sequence has a natural-language description enabling text-controlled generation.

Results

Zero-shot cross-topology transfer: trained only on bipedal and quadrupedal data, the model generates plausible animations for birds, snakes, and insects (e.g., coordinated eight-legged crawling for spiders it never saw).
Compared against specialized SOTA methods, UniMate surpasses them on smoothness, naturalness, and physical plausibility across all topologies, while unseen topologies cause large performance drops in baselines but stable performance in UniMate, without any per-topology fine-tuning.
Demonstrated applications: in-betweening (keyframe interpolation), animation expansion (extending short clips into long cycles), and text-guided editing (e.g., "make this cat jump higher").

Significance

Democratizing animation: professional-quality motion without mocap rigs ($50,000+), specialist animators, or weeks of hand-tuning — input a text prompt plus any 3D skeleton.
Embodied intelligence: by encoding graph structure and physical constraints, the model approximates a "body intuition" about gait patterns and joint limits.
Toward general robot control: one shared "motion brain" could potentially drive humanoids, quadruped robots, and manipulators alike.

Limitations and future work

Generated motion is visually plausible but not always physically exact (no mass distribution, inertia, or ground reaction forces); physics constraints may be needed for robotics use.
Output is skeletal animation only; secondary motion (muscle jiggle, tail dynamics) requires separate handling.
Multi-character interaction and environment interaction remain open challenges.
Future directions: real-time generation (<16 ms for games/VR), finer style control ("tiredly running", "joyfully jumping"), and cross-modal extensions (speech→lip sync, music→dance, video→motion).

Reference

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.