Princeton's UniMate Animates Dragons, Crabs and Robots With one AI Model
A Princeton-led team released UniMate, a single diffusion transformer that animates any rigged skeleton from text, with code, dataset, and interactive demo.
- Princeton-led team released UniMate, a single model that animates any rigged skeleton from text.
- Topology-aware diffusion transformer uses graph attention bias, spectral RoPE, and rest-pose conditioning.
- New UniML3D dataset: 13,006 clips across bipedal, quadrupedal, avian, marine, insectoid, serpentine, and rigid rigs.
- Zero-shot support for motion editing, in-betweening, and multi-prompt expansion from one checkpoint.
- Dataset and data-processing pipeline live now; training and inference code coming soon.
- Try the interactive demo with your own .fbx or .glb rigs in the browser.
UniMate animates different skeletons with one motion model
UniMate generates articulated motion from a rigged 3D asset and a text prompt. A single checkpoint handles humans, dragons, crabs, snakes, octopuses, cheetahs, and mechanical objects without per-skeleton fine-tuning or test-time optimization. Researchers from Princeton, UC Berkeley, MIT, and NTU describe the system in a SIGGRAPH Asia 2026 paper.
At the time of writing, the GitHub repository had passed 900 stars, with hundreds added during the previous week. The dataset, processing tools, browser demo, and model weights are public, while training and inference scripts remain pending.
The skeleton becomes part of the prompt
Skeleton topology describes how joints connect into a kinematic tree. Rigs can differ in joint count, parent-child relationships, proportions, and rest pose, even when their motions look similar. An auto-rigged creature with an unfamiliar hierarchy can therefore fall outside the templates supported by a category-specific animator.
Many learned animation systems assume a fixed human or animal template, require reference motion, or adapt a model to each new skeleton. UniMate passes the supplied skeleton directly into a topology-aware diffusion transformer, which iteratively refines motion tokens while conditioning on the rig and text.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
