UniMate animates any skeleton from text, no per-rig training

Motion, not rigging, is now the bottleneck. That is the opening claim of UniMate, a SIGGRAPH Asia 2026 paper from a Penn and Princeton group led by Linzhan Mou, which argues that automatic rigging already delivers animation-ready 3D assets at scale while the models that move them are still stuck to one skeleton at a time.

The pitch is a single foundation model that takes a rigged asset plus a text prompt and outputs motion, for any skeleton, with "no test-time optimization or per-skeleton retraining." Existing learned animators, the authors write, "are topology-constrained: they rely on category-specific templates or require per-skeleton fine-tuning and reference motions at inference." UniMate replaces that with a topology-aware diffusion transformer that folds the skeleton graph itself into attention through three pieces: a graph-aware attention bias from pairwise joint relations and geodesic distances, a spectral rotary position embedding that generalises RoPE to arbitrary kinematic trees via the graph Laplacian, and a global topological conditioner pooled from the rest-pose skeleton.

To train it, the group assembled UniML3D, "13,006 motion sequences spanning bipedal, quadrupedal, avian, marine, insectoid, serpentine, and articulated rigid objects with unified canonicalization and text pairing." The abstract says the resulting model "outperforms state-of-the-art baselines in quality, generalization, and efficiency, and supports zero-shot cross-topology transfer, in-betweening, expansion, and text-guided editing."

No per-topology numbers, no FID, no user-study scores appear in the abstract. The strong claim — one model, any skeleton, from text — is currently backed only by that summary sentence and a project page.