SIMART¶

Development line: project:simart · thread simart-development
Last event: 2026-05-19 · 2 dated since 2026-03-30 · Researched: 2026-09-04 · confidence: high

What it is¶

SIMART is a ByteDance Seed research implementation for 3D and robotics practitioners who need to convert a monolithic mesh into articulated parts for simulators. It combines a Qwen3-VL-based MLLM with a Sparse 3D VQ-VAE. It predicts part decomposition, physical metadata, joint types, axes, and limits, then outputs URDF-oriented files. The paper reports 70% fewer geometry tokens than dense voxels. The repository is tested on Python 3.10 and provides local checkpoints rather than a hosted inference endpoint. Use it as a local research pipeline for preparing articulated assets, and validate the generated URDF before relying on it in a simulator.

Development line¶

• 2026-03-30 — SIMART project website referenced. On 2026-03-30, the project website established the first public entry point. The available evidence identifies no specific release, capability, or version.
• 2026-05-19 — SIMART Hugging Face repository referenced. On 2026-05-19, the ByteDance-Seed/SimArt Hugging Face repository appeared alongside the project page. The available evidence does not establish repository contents, license, or exact release status.

What changed¶

• 2026-03-24 — The arXiv preprint introduced SIMART’s unified mesh-to-articulated-asset method, Sparse 3D VQ-VAE, and SIMART-Bench, reporting a 70% token reduction against dense voxel tokens.
• 2026-03-30 — The project page made the research presentation and interactive articulated-asset gallery available, describing structured URDF metadata plus decomposed segments.
• 2026-05-19 — Tracked links added the ByteDance-Seed Hugging Face repository alongside the project page. The model card identifies two checkpoint groups, simart_mllm/ and simart_vqvae/.
• 2026-05-28 — An update appeared on ByteDance-Seed/SimArt on Hugging Face. The model card supplies weights without a hosted inference provider.
• 2026-09-05 — The public GitHub README documented local installation, mesh alignment, preprocessing, and inference for GLB inputs.

How to use this¶

From 2026-05-19, practitioners should check the linked ByteDance-Seed/SimArt Hugging Face repository alongside the SIMART project site when looking for project materials; the dated evidence does not confirm which artifacts are usable.

1. Clone the implementation, create a Python 3.10 environment, and install requirements.txt. — https://github.com/ByteDance-Seed/SimArt
2. Download both checkpoint groups from the official Hugging Face repository and place them under ./checkpoints or pass custom paths. — https://huggingface.co/ByteDance-Seed/SimArt
3. Normalize a raw .glb mesh with scripts/process_raw_objects.py; render a preview during preprocessing. — https://github.com/ByteDance-Seed/SimArt
4. Run inference/infer.py --object_path <processed.glb> --debug to generate the articulated structure, URDF and debug visualizations. — https://github.com/ByteDance-Seed/SimArt

Best practices¶

• Use a right-handed coordinate system with +Z up and a consistent forward direction; correct rotations for meshes that use a different convention. — https://github.com/ByteDance-Seed/SimArt
• Render the normalization preview and check that the object is upright and forward-facing before inference, because orientation affects semantic parts such as handles and front legs. — https://github.com/ByteDance-Seed/SimArt
• Use --debug to inspect colored part meshes and joint axes, then verify the produced URDF in the target simulator before using it for robot-policy work. — https://github.com/ByteDance-Seed/SimArt

Superseded by this¶

• 2026-03-24 — Treating SIMART as a paper-only method is obsolete: the repository and checkpoint publication provide a local inference path.
• 2026-05-28 — Assuming SIMART is available through a managed Hugging Face inference provider is unsupported; the official model page states that no provider deploys it.

Still unknown¶

• The official Hugging Face model card does not display its own publication timestamp, so no independently dated change can be assigned specifically to the 2026-05-19 event beyond the presence of that repository link.
• The name is ambiguous on the wider web: an unrelated 6G sensing-and-communication platform is also called SimART. The cited ByteDance Seed project, paper, code and weights consistently identify the articulated-3D-assets subject.

Sources¶

Agent brief¶

• Subject: project:simart, thread simart-development, 2 dated events 2026-03-30 → 2026-05-19.
• Practical note: From 2026-05-19, practitioners should check the linked ByteDance-Seed/SimArt Hugging Face repository alongside the SIMART project site when looking for project materials; the dated evidence does not confirm which artifacts are usable.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.