UniMate paper proposes one text-to-motion model for any 3D skeleton

Researchers including Adam Finkelstein and Szymon Rusinkiewicz of Princeton University have described UniMate, a diffusion-transformer model that generates motion for a rigged 3D character or object from a text prompt, without retraining for each skeleton. The paper is dated September 4, 2026, and is listed on Hugging Face Papers.

The authors frame the problem as a gap left by automatic rigging: animation-ready 3D assets are now produced at scale, but generating motion to drive them remains a bottleneck. They say existing learned animators are tied to skeleton topology, relying on category-specific templates or needing per-skeleton fine-tuning and reference motions at inference time.

UniMate's stated answer is a single model that takes a rigged asset and a text prompt and needs no test-time optimization or per-skeleton retraining. According to the abstract, skeleton structure enters the transformer's attention in three ways: a bias built from pairwise joint relations and geodesic distances, a rotary position embedding generalized to arbitrary kinematic trees through the graph Laplacian, and a global conditioner pooled from the rest-pose skeleton.

The team also built UniML3D, a set of 13,006 motion sequences with unified canonicalization and text pairing. It spans bipedal, quadrupedal, avian, marine, insectoid and serpentine bodies as well as articulated rigid objects, and the model was trained on it.

The authors say UniMate outperforms state-of-the-art baselines in quality, generalization and efficiency, and that it supports zero-shot transfer across topologies, in-betweening, expansion and text-guided editing. The saved page gives no benchmark numbers, so those comparisons cannot be checked here.

What is confirmed and what is not: the page confirms what the authors say they built and claim, not that the results hold. This is the authors' own account, mirrored from their arXiv abstract, and a search for an independent evaluation, reproduction or press analysis found none that could be verified. Only one credible source exists for this story. The page also links a project page and a code repository, which were not used as sources here.

Why it matters: if the approach holds up beyond the authors' tests, developers working with varied auto-rigged assets could avoid per-character animation setup. Whether it works on production assets, and with which file formats, is not addressed by the source.

Revised 16 October 2026: Guth re-read the original sources and rewrote this article. Correction: the earlier version said the source does not report comparative animation performance or state animation speed. The source says the authors claim UniMate outperforms state-of-the-art baselines in quality, generalization and efficiency, though it gives no benchmark figures. The earlier version remains available as revision 1.