Research map / Video / Video generation

SAMoR: Motion Modelling for Articulated Objects of Any Skeleton and Topology

NeurIPS 2026 · Video › Video generation · approach: Text-to-motion diffusion · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SAMoR, a graph-transformer part VQ-VAE that compresses arbitrary-skeleton motion into K=8 shared part tokens with attention supervision and joint-name dropout, enabling cross-topology reconstruction, transfer, text-to-motion, and part editing.

Problem

Motion generation and tokenization are tied to fixed human skeletons; existing cross-topology methods lack a fixed-size discrete motion vocabulary shared across rigs with different joint counts and connectivity.

Approach

Encode motion into a small fixed number of functional part tokens via cross-attention pooling over a graph-transformer encoder, supervise the pooling attention with joint-name-derived functional-group labels to prevent query collapse, and use joint-name dropout for robustness to missing names.

Key results

• Cross-topology reconstruction MPJPE 2.75e-2, about 5.8x below the strongest adapted baseline MoGenTS (16.14e-2).
• On HumanML3D, SAMoR-H3D achieves FID 0.039 and Top-3 R-Precision 0.812, competitive with MoGenTS (FID 0.033) and MoMask (0.045).
• Removing attention supervision at K=8 gives MPJPE 5.04, nearly matching K=1 (5.49), suggesting supervision rather than query count drives structured routing.

Surprising findings

• K=8 without attention supervision performs nearly as poorly as K=1, so query count alone provides little benefit.
• Joint-name removal at inference costs only modest error increase (about 16% relative), and parts are still routed sensibly without names.

Limitations

• Cross-topology setting operates on joint positions and velocities rather than rotations, requiring IK post-processing for mesh animation with twist ambiguity.
• The K=8 functional grouping is a fixed design choice.
• Only 69% of OXL joints and 81% of Zoo joints receive canonical labels, limiting supervision coverage.

Experimental setup

Datasets: HumanML3D Truebones Zoo Objaverse-XL

Benchmarks: HumanML3D Truebones Zoo Objaverse-XL (held-out characters)

Metrics: FID R-Precision MM-Dist MPJPE bone-length error round-trip MPJPE

Baselines: T2M-GPT MoMask MoGenTS AnyTop SinMDM

Builds on: MoMask T2M-GPT MoGenTS CLIP Puppeteer

Closest papers by meaning

• Learning Implicit Representation for Reconstructing Articulated Objects (ICLR 2024 · similarity 0.90)
• Heterogeneous Skeleton-Based Action Representation Learning (CVPR 2025 · similarity 0.90)
• Universal Skeleton Understanding via Differentiable Rendering and MLLMs (ICML 2026 · similarity 0.88)
• How to Move Your Dragon: Text-to-Motion Synthesis for Large-Vocabulary Objects (ICML 2025 · similarity 0.88)
• Revealing Key Details to See Differences: A Novel Prototypical Perspective for Skeleton-based Action Recognition (CVPR 2025 · similarity 0.87)
• S3O: A Dual-Phase Approach for Reconstructing Dynamic Shape and Skeleton of Articulated Objects from Single Monocular Video (ICML 2024 · similarity 0.87)