UniMo4D: Unified Motion Dataset for Cross-Species

• UniMo4D is a unified dataset that standardizes human and animal motions to a common 25-joint topology, supporting cross-species text-driven motion generation.
• It integrates data from 115 species with 118,663 motion sequences, utilizing a shared skeleton with three virtual tail joints to maintain morphological consistency.
• The dataset underpins the X-MoGen framework, enabling joint training and evaluation with metrics that assess both semantic alignment and anatomical plausibility.

Searching arXiv for UniMo4D and associated work. UniMo4D is a large-scale dataset for unified modeling of human and animal motion, introduced to support cross-species text-driven motion generation under a shared skeletal representation. It standardizes human and animal motions to a common 25-joint topology while retaining species-specific morphology, enabling unified representation learning, cross-species generalization, and direct benchmarking of both semantic alignment and morphological plausibility. In the X-MoGen framework, UniMo4D is the dataset that integrates human and animal motions under a shared skeletal topology for joint training, with 115 species and 119k motion sequences reported at the framework level and 118,663 total motion sequences specified in the dataset description (Wang et al., 7 Aug 2025).

1. Definition and research motivation

UniMo4D was introduced in response to a recurring limitation in motion generation research: human and animal motions had typically been modeled separately because of substantial differences in morphology and skeletal structure. This separation hindered cross-species generalization and made unified representation learning difficult. UniMo4D addresses this by standardizing skeletal topology across species while retaining species-specific morphology, thereby enabling unified representation learning and cross-species transfer within a single dataset (Wang et al., 7 Aug 2025).

The dataset is explicitly motivated by two objectives. First, it supports unified representation learning, in which motion semantics across humans and animals can be aligned in a structured latent space. Second, it supports cross-species generalization, since diverse body morphologies and motion patterns in the same corpus can improve transfer to new or unseen species and to novel actions. This suggests that UniMo4D functions not merely as a data aggregation effort but as an infrastructural benchmark for universal motion modeling (Wang et al., 7 Aug 2025).

Within the X-MoGen study, UniMo4D also serves as the empirical basis for evaluating whether a single model can operate across both humans and animals. The paper describes X-MoGen as the first unified framework for cross-species text-driven motion generation covering both humans and animals, and UniMo4D is the large-scale dataset constructed to support that goal (Wang et al., 7 Aug 2025).

2. Scale, composition, and dataset partitions

UniMo4D covers 115 total species, comprising humans and 114 animal species. It contains 118,663 motion sequences in total, and each sequence has length $L \in (18, 300)$. Every motion is represented with a skeleton of 25 joints ($N_j = 25$), including three virtual tail joints introduced for universality (Wang et al., 7 Aug 2025).

The dataset includes an unseen-species test set consisting of 5 species and 5,179 motion sequences, reserved for out-of-distribution generalization testing. For the remaining species, the train/validation/test split is 80\% / 5\% / 15\% (Wang et al., 7 Aug 2025).

The scale of UniMo4D is central to its role as a benchmark. The dataset description emphasizes that prior datasets were typically limited to only humans or to much smaller sets of animal skeletons, often with incompatible joint layouts and scales. By contrast, UniMo4D provides a single standardized corpus large enough to support both in-distribution evaluation and held-out species evaluation under a common protocol (Wang et al., 7 Aug 2025).

3. Data sources and integration pipeline

UniMo4D is constructed by integrating motions from HumanML3D, KIT-ML, and AniMo4D. The integration process is described in three stages: feature representation, joint standardization, and unified skeleton topology construction (Wang et al., 7 Aug 2025).

For feature representation, the dataset adopts "essential feature groups" as per the cited prior work, using a compact frame representation of $N_j \times 3 + 1$, where each joint is a 3D position and the root joint includes rotation velocity, linear velocity, and height (Wang et al., 7 Aug 2025).

For joint standardization, UniMo4D remaps joints from KIT-ML and reorders those from AniMo4D to match the HumanML3D convention, and then scales bone lengths accordingly. This step aligns heterogeneous source datasets to a common indexing and geometric convention (Wang et al., 7 Aug 2025).

For topology unification, all skeletons are standardized to a 25-joint topology, including three virtual tail joints to accommodate animals with tails. For tailless species such as humans, the virtual tail joints are colocated with the pelvic root in order to prevent motion artifacts while maintaining consistency. As a result, every motion in UniMo4D has the same joint ordering and topology, ensuring interoperability across species (Wang et al., 7 Aug 2025).

Aspect	UniMo4D specification
Data sources	HumanML3D, KIT-ML, AniMo4D
Universal skeleton	25 joints
Special design	Three virtual tail joints

This integration strategy is significant because it makes cross-species comparison and joint training operationally possible. A plausible implication is that the standardization step is as important as the raw scale of the corpus, since a large but topologically inconsistent dataset would not support the same class of unified models.

4. Shared topology and morphological design

The defining structural feature of UniMo4D is its shared 25-joint skeletal topology. This topology is used as the universal template for all species and includes three virtual tail joints. For humans and other tailless species, these joints are initialized to the pelvis so that topological consistency is preserved without introducing spurious motion (Wang et al., 7 Aug 2025).

The dataset description states that this topology provides consistent bone indexing and joint names/order across all species and activities. That consistency enables explicit structural evaluation, including bone length errors and other morphology-sensitive criteria, across species with very different body plans (Wang et al., 7 Aug 2025).

A core challenge addressed by this design is huge morphological diversity. The paper highlights that bone lengths and proportions vary massively across species, as illustrated by the length distribution of key bones. UniMo4D therefore does not eliminate morphology; rather, it standardizes topology while preserving morphological variation as a modeling challenge. This suggests that the dataset is intended to separate representation compatibility from anatomical homogeneity (Wang et al., 7 Aug 2025).

The topology also supports mixed locomotor regimes, including bipedal and quadrupedal motion patterns, and includes species both with and without tails. That breadth is directly relevant to cross-species modeling, since it exposes models to substantial variation in articulation, balance, and motion semantics within one representation space (Wang et al., 7 Aug 2025).

5. Training, benchmarking, and evaluation protocol

UniMo4D serves as the sole training and evaluation dataset for all models in the X-MoGen study, including X-MoGen itself and all baselines. All comparisons, generalization evaluations, and ablations are run on UniMo4D splits, and the unseen-species split is used as a dedicated test of cross-species generalization (Wang et al., 7 Aug 2025).

The benchmark includes standard text-to-motion metrics: R-Precision (Top-1/2/3) for text-motion alignment, FID for distributional similarity between generated and real motion features, MM-Dist for average distance between motion and text embeddings, and Diversity for motion variability. UniMo4D also introduces a dataset-specific metric, Mean Morphological Error (MME), to evaluate morphological plausibility through skeletal bone-length consistency (Wang et al., 7 Aug 2025).

The MME formula is given as

$$\text{MME} = \frac{1}{L \cdot N_b} \sum_{t=1}^{L} \|\hat{\mathbf{b}}_t - \mathbf{b}\|_1,$$

where $\hat{\mathbf{b}}_t \in \mathbb{R}^{N_b}$ denotes bone lengths at frame $t$, $\mathbf{b} \in \mathbb{R}^{N_b}$ is the reference skeleton, $L$ is sequence length, and $N_b$ is the number of bones (Wang et al., 7 Aug 2025).

The introduction of MME is methodologically important because it formalizes a property that becomes especially salient in cross-species generation: a motion sequence may be semantically aligned with text yet still be anatomically implausible. UniMo4D therefore evaluates both semantic fidelity and structural consistency within the same benchmark.

6. Reported benchmark results and ablations

On the UniMo4D test set for seen species, the reported comparison table includes Real, MDM, T2M-GPT, AttT2M, MMM, MoMask, MARDM, AniMo, and X-MoGen. In that table, X-MoGen achieves R-Precision Top-1 = 0.848, FID = 0.050, MM-Dist = 0.742, Diversity = 19.332, and MME = 0.201, while the Real data row reports 0.899, 0.000, 0.470, 19.407, and no MME entry, respectively. The summary states that X-MoGen outperforms state-of-the-art methods on both seen and unseen species on UniMo4D (Wang et al., 7 Aug 2025).

A separate compression and representation study compares quantization-based compressors with an autoencoder on UniMo4D. The reported table gives the following values:

Compressor	Top-1	FID	MM-Dist
VQ (T2M-GPT)	0.581	5.010	4.015
FSQ (MMM)	0.629	4.272	3.414
RVQ (MoMask)	0.700	0.451	2.299
AE (X-MoGen)	0.898	0.008	0.496

The same study reports MME values of 0.389, 0.388, 0.220, and 0.064 for VQ, FSQ, RVQ, and AE, respectively (Wang et al., 7 Aug 2025).

The stated takeaway is that quantization-based compressors struggle with UniMo4D’s diversity, whereas continuous autoencoders suffice for highly diverse unified motion data. Ablation studies further remove word-level CLIP features, the T-pose prior, or morphology-aware losses, and the summary reports that all of these elements are beneficial for performance on UniMo4D’s mixed-species benchmark (Wang et al., 7 Aug 2025).

7. Distinctiveness, limitations addressed, and research significance

UniMo4D is described as the first large-scale dataset to unify humans and over 100 animal species under a single joint topology. Its distinctiveness lies not only in scale but also in the combination of standardized topology, preserved morphological diversity, and an evaluation protocol that includes both conventional text-motion metrics and a new morphology-sensitive metric (Wang et al., 7 Aug 2025).

Relative to prior datasets, the paper identifies three main limitations that UniMo4D addresses. First, earlier datasets often focused only on humans or on a small number of animal species. Second, their skeleton definitions were typically incompatible, which obstructed unified training. Third, they did not provide a shared testbed for direct comparison across model classes or for out-of-distribution species evaluation (Wang et al., 7 Aug 2025).

UniMo4D therefore supports several research directions simultaneously: universal text-to-motion generation, cross-species generalization, morphological consistency modeling, and fair retraining-based comparison of prior baselines under a single data regime. It also enables explicit morphological reasoning, because constant joint ordering and topology make it possible to quantify structural fidelity directly through measures such as MME (Wang et al., 7 Aug 2025).

The broader significance of UniMo4D is that it redefines dataset construction for motion generation from a domain-specific resource to a universal benchmark. This suggests a shift from separate human-motion and animal-motion ecosystems toward a common representation space in which variation in morphology is treated as a first-class modeling variable rather than a reason for dataset fragmentation (Wang et al., 7 Aug 2025).