UniMate: One Unified Model to Animate Diverse Skeletons

UniMate. One Unified Model to Animate Diverse Skeletons. By Linzhan Mou and six others, from Princeton University, the University of California at Berkeley, and two other institutions. Published on the fourth of September. Its code repository is worked out: about one hundred and fifty source files, with a readme and a licence.

Automatic rigging now delivers animation-ready 3D assets at scale, which moves the bottleneck downstream. Generating the motion to drive those assets is where it sits instead. Existing learned animators are topology-constrained, and that constraint takes two forms. Either they rely on category-specific templates, or they require per-skeleton fine-tuning together with reference motions supplied at inference.

UniMate is a unified foundation model that synthesizes articulated motion for arbitrary skeletons, taking a rigged 3D asset and a text prompt, with no test-time optimization and no per-skeleton retraining. One model, any skeleton, and nothing asked of the user at inference beyond the asset and the words.

The mechanism is a topology-aware diffusion transformer, which integrates skeletal topology into attention in three ways. A graph-aware attention bias is built from pairwise joint relations and geodesic distances, so attention knows how far apart two joints are along the body rather than in space. A spectral rotary position embedding generalizes RoPE to arbitrary kinematic trees by way of the graph Laplacian, which is what stands in for a fixed ordering of joints. And a global topological conditioner is attention-pooled from the rest-pose skeleton, giving the model a summary of the whole body before it animates any part of it.

Training data was curated for the purpose. UniML3D holds thirteen thousand and six motion sequences, spanning bipedal, quadrupedal, avian, marine, insectoid and serpentine skeletons as well as articulated rigid objects, all under unified canonicalization with text pairing. Seven kinds of body across thirteen thousand sequences is what makes an arbitrary skeleton testable rather than aspirational. Trained on it, UniMate outperforms state-of-the-art baselines in quality, generalization and efficiency.

Beyond matching baselines, the model supports things the topology constraint used to rule out: zero-shot cross-topology transfer, in-betweening, expansion, and text-guided editing. A skeleton it has never seen before stops being a retraining job.

UniMate. One Unified Model to Animate Diverse Skeletons. By Linzhan Mou and six others, from Princeton University, the University of California at Berkeley, and two other institutions. Published on the fourth of September. This is the long version: the whole paper rather than the abstract.

The pipeline for making 3D content has been automated end to end except for one step. Generating assets works. Automatic rigging works, delivering skeleton-ready characters across humans, animals and articulated rigid objects at scale. And then the motion that drives those rigs still has to be made by hand, through keyframing, motion capture, or slow per-asset optimization. Animation is the laborious final bottleneck in an otherwise automated process, and the reason is not that motion models are bad. It is that they are topology-constrained.

State-of-the-art motion generators rely on fixed category-specific skeleton templates, which is fine if your character is a human or a quadruped and useless otherwise. Topology-agnostic models relax that, and then require per-skeleton fine-tuning or reference motions supplied at inference. Mesh-based methods avoid skeletons entirely, at the cost of expensive per-asset distillation or kinematically unconstrained deformation. Two challenges stand behind all of this. The modelling one is that real skeletons are genuinely heterogeneous, with different kinematic trees, joint counts and motion patterns, so a general model has to take topology as an explicit input rather than baking it into an architectural prior. The data one is that text-paired motion corpora spanning diverse topologies barely exist.

The architecture reasons over rest-pose kinematics and motion in a shared token stream, with three mechanisms carrying the topology. A graph-aware attention bias is derived from pairwise joint relations and geodesic distances, so anatomically nearby joints attend more strongly while the model keeps its capacity for long-range full-body coordination. A global topological conditioner, pooled from the skeleton tokens, modulates every transformer block so that layer-wise feature statistics adapt to whichever skeleton arrived. And the position encoding is replaced with a spectral one that derives its rotary angles from the graph Laplacian, generalising the usual rotary embedding to arbitrary kinematic trees.

The data side is met by curating a heterogeneous corpus of thirteen thousand and six animation sequences, roughly twenty hours, drawn from three existing sources and covering thousands of distinct rigs across bipedal, quadrupedal, avian, marine, insectoid and serpentine skeletons as well as articulated rigid objects, paired with three thousand five hundred and eighty-four unique text prompts. Filtering and a unified canonicalization produce a shared representation across skeleton types, and because the topology distribution is long-tailed, training uses balanced sampling plus four kinematics-preserving augmentations applied on the fly: adding a joint, removing one, pooling the skeleton, and perturbing bone lengths. The whole model trains in about a day on eight accelerators and runs at fifty frames per second.

Beyond the missing baseline, the scale deserves keeping in proportion. Twenty hours of motion is small by the standards of human motion corpora, and it is spread across thousands of rigs, so per-topology coverage is thin by construction. The augmentations exist precisely because of that thinness, which means part of the topological diversity the model sees is synthetic rather than observed. The prompts span a broad action spectrum but a finite one. And the guarantees offered are about the position encoding's structural properties, not about the motion: permutation equivariance says the model will treat a relabelled skeleton consistently, not that what it produces will be anatomically plausible.

The idea generalises past animation. When a model has to handle inputs with genuinely different structure, the tempting move is to normalise them into a common format, which is what a category-specific template does and why it fails outside its category. The alternative is to make structure a first-class input and build the architecture so that structure flows through attention: as a bias on who attends to whom, as a conditioning signal on every block, and as a position encoding derived from connectivity rather than from an ordering somebody chose. Then a skeleton the model has never seen is simply another input, rather than a retraining job.