RigMo: Unifying Rig and Motion Learning for Generative Animation

Abstract: Despite significant progress in 4D generation, rig and motion, the core structural and dynamic components of animation are typically modeled as separate problems. Existing pipelines rely on ground-truth skeletons and skinning weights for motion generation and treat auto-rigging as an independent process, undermining scalability and interpretability. We present RigMo, a unified generative framework that jointly learns rig and motion directly from raw mesh sequences, without any human-provided rig annotations. RigMo encodes per-vertex deformations into two compact latent spaces: a rig latent that decodes into explicit Gaussian bones and skinning weights, and a motion latent that produces time-varying SE(3) transformations. Together, these outputs define an animatable mesh with explicit structure and coherent motion, enabling feed-forward rig and motion inference for deformable objects. Beyond unified rig-motion discovery, we introduce a Motion-DiT model operating in RigMo's latent space and demonstrate that these structure-aware latents can naturally support downstream motion generation tasks. Experiments on DeformingThings4D, Objaverse-XL, and TrueBones demonstrate that RigMo learns smooth, interpretable, and physically plausible rigs, while achieving superior reconstruction and category-level generalization compared to existing auto-rigging and deformation baselines. RigMo establishes a new paradigm for unified, structure-aware, and scalable dynamic 3D modeling.

• The paper introduces RigMo, a unified framework that learns rig structures and motion dynamics directly from mesh deformations without requiring manual rig annotations.
• It employs a topology-aware dual-path encoder and Gaussian-based skinning to achieve state-of-the-art reconstruction accuracy and efficient inference.
• The method generalizes across diverse shapes and animations, enabling scalable, controllable, and interpretable generative animation for various applications.

RigMo: Unified Rig and Motion Learning for Generative Animation

Introduction and Motivation

Generative animation in graphics hinges on the interplay between static rigging structures (defining permissible deformations) and dynamic motion trajectories (articulating temporal evolution). Existing paradigms silo these two aspects: auto-rigging systems depend on artist-annotated skeletons and skinning weights, while motion generators operate under the assumption of fixed ground-truth rigs. This separation curtails scalability, interpretability, and cross-category generalization, especially for objects lacking established rigging conventions. Moreover, state-of-the-art motion generators in vertex or Gaussian primitive space yield deformations without structurally meaningful parameterization, impeding control and reusability in downstream tasks.

RigMo (2601.06378) introduces a fundamentally unified framework capable of learning both rig structures and motion dynamics directly from raw mesh deformations, bypassing all rig annotation requirements. The approach is entirely self-supervised and generalizes seamlessly across diverse morphologies and motion patterns.

Figure 1: RigMo-VAE architecture disentangles rig structure and motion dynamics from mesh trajectories into interpretable latent factors and Gaussian bone regions.

RigMo-VAE: Architecture and Representation

RigMo employs a topology-aware dual-path encoder to factorize mesh sequence input into two complementary latent spaces:

• Rig latent: Encodes static geometry, decodes to explicit Gaussian bones and geodesic-aware skinning weights. Each bone parametrizes its region's influence via spatial coordinates, anisotropic scaling, and rotation quaternion, defining a soft ellipsoid in vertex space.
• Motion latent: Encodes per-frame vertex displacements, decodes to temporal SE(3) bone transformations (local and root). These together enable full articulation and motion synthesis.

The decoder reconstructs the mesh by applying Gaussian-based LBS, leveraging geodesic-aware refinement to respect mesh topology and suppress erroneous cross-region influences. Key design elements include sparse anchor sampling for bone tokens and efficient attention mechanisms to preserve spatial coherence and minimize token cardinality without loss of rigging fidelity.

RigMo Motion DiT: Structure-Aware Motion Synthesis

Building upon the learned rig-motion latents, RigMo introduces a Motion DiT—a diffusion transformer conditioned on static rigging tokens and global priors, operating in the motion-latent space. It generates temporally plausible motion trajectories by denoising motion latent tokens subjected to frame-wise sparsity and mask scheduling. This process produces bone transformations and vertex sequences via Gaussian skinning, enabling conditional, controllable long-horizon animation synthesis.

Figure 2: Motion DiT schematic—conditional transformer with cross-attention on anchor and global tokens enabling denoising in RigMo’s motion-latent space.

Experimental Evaluation

Datasets and Topology Normalization

RigMo’s robustness is evaluated on three datasets: DeformingThings4D (organic non-rigid shapes), Objaverse-XL (diverse synthetic objects), and TrueBones (articulated animations). All meshes are normalized to a common vertex resolution using topology-preserving farthest point sampling and geodesic neighborhood maintenance, ensuring architecture resolution-agnosticism.

Quantitative Results: Rigging and Reconstruction

RigMo demonstrates substantial improvements over baselines:

• Generalization: Unlike per-case optimization and static auto-rigging, RigMo sustains fidelity and stability across unseen motion sequences, with markedly lower Chamfer Distances in both training reconstruction and cross-motion transfer.
• Efficiency: RigMo achieves state-of-the-art reconstruction accuracy (CD-L1: 1.73e-2, CD-L2: 1.26e-2) and inference speed (0.74s for 20 frames) surpassing AnimateAnyMesh, Step1X3D, and Hunyuan3D.
• Self-supervised scalability: RigMo trains end-to-end on unlabeled data, eliminating annotation bottlenecks and facilitating learning from large-scale motion corpora.

  Figure 3: Given sparse observed frames, RigMo jointly reconstructs the rig structure and synthesizes missing motion with coherent articulation across categories.

Comparative and Qualitative Analysis

RigMo’s learned rigs are robust and transferable—visually and kinetically—against methods such as UniRig and MagicArticulate. Qualitative comparisons reveal that static geometry-based rigging often collapses under dynamic motion, whereas RigMo produces anatomically coherent bones and stable skinning with no ground-truth supervision required.

Figure 4: RigMo’s rigging generalizes robustly across motions and categories; baseline methods fail under animation, yielding severe artifacts.

Diagnostics and Ablation

• Geodesic refinement: Essential for suppressing topological cross-talk, its removal degrades reconstruction accuracy by approximately 20%.
• Bone token cardinality: 128 tokens offer optimal fidelity, but practical trade-offs favor 48-token configurations yielding near-equivalent performance with improved interpretability and efficiency.
• Resolution robustness: RigMo’s continuous-space bone definition enables stable animation across mesh resolutions, unlike vertex-index-dependent models.

  Figure 5: Side-by-side comparison of skinning weights for 48- and 128-token configurations; 128 tokens yield finer-grained assignments.

Implications and Future Directions

RigMo represents a paradigm shift: joint rig-motion learning unlocks scalable, structure-aware animation modeling for any deformable object, nullifying reliance on category-specific rig annotations. The Gaussian bone abstraction and latent motion synthesis enable explicit, physically plausible, and manipulable asset generation, interfacing directly with downstream tasks (e.g., motion editing, interpolation, cross-object retargeting).

Applications can be extended to human, animal, and non-humanoid shape animation in AR/VR, film, and gaming. With end-to-end differentiability and feed-forward inference, RigMo can be integrated with generative 4D pipelines or unified with conditional controls, such as text-driven animation or physics-guided synthesis. Future investigation may include:

• Differentiable physical priors in rig/motion factorization (e.g., elasticity/material modeling)
• Mesh topology self-discovery for rig-vs-nonrig separation
• Domain-adaptive rig learning for procedural asset creation
• Extension toward multi-agent or crowd animation with shared motion-latent spaces

Conclusion

RigMo advances the field of generative animation through explicit, unified rig and motion learning, eschewing annotation dependencies and demonstrating state-of-the-art accuracy, efficiency, and generalization. The framework establishes a basis for scalable, controllable, and interpretable dynamic 3D modeling, with implications for practical graphics workflows and theoretical exploration of motion/rig-factorized representations in generative AI.