Dataset Viewer
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): (None, {}), NamedSplit('test'): ('meshfolder', {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

DragMesh-3 Scene assets

MuJoCo: complete standalone object assets

Download the MuJoCo object ZIP · Usage and conversion limitations · Validation report · Checksums

14 objects, including 8 articulated objects with 49 moving joints. Each has fixed-base and free-base MJCF entry points, all referenced meshes and textures, collision geometry, link mass/inertia, joint limits and motor actuators. Install mujoco==3.11.0, extract, and run python run_asset.py --asset fixture_cabinet_b01 from the extracted directory. No USD conversion, private repository access or HF login is needed to run it.

All 28 entry points were loaded and simulated for 3,000 steps each. Native joint kinematics, representative motor actuation and bowl/mug cavities were checked. Convex collision approximation and MuJoCo tuning are documented; this is not PhysX-equivalent dynamics certification. Small initial overlaps are reported. This release contains standalone objects, not composed rooms or a G1 robot.

Web preview: portable three-room demo (2026-11-03)

The new agentic-scenes/v1 release is a self-contained web preview of three room layouts: everyday kitchen, coffee room and preparation room. It includes all web models, native joint metadata, interaction location candidates, three construction videos, viewer code, scene plans and per-asset licenses.

Download the demo ZIP · Usage, provenance and limitations · SHA-256 checksum

Download only that ZIP, rather than cloning the entire repository. Extract it and run:

python3 -m http.server 8780 --bind 127.0.0.1 --directory dragmesh-agentic-scenes-v1/web

Then open http://127.0.0.1:8780/ . No GPU, API key or private demo login is needed. Use the object selector and joint slider to inspect moving parts. The purple G1 marker is a reserved placement slot, not robot execution.

This is a web-preview release, not a complete simulator or native USD/SDF distribution. Construction videos replay deterministic saved plans; they do not show autonomous LLM generation or physical robot task success.

Legacy archive

Existing directories (including DragMesh-3, dragmesh3-core, scenesmith_full, trajectories, and other historical outputs) are preserved unchanged. They contain earlier code, assets and experiment outputs and are not required for the v1 web demo. Their presence does not imply that every experimental pipeline is turnkey.

Attribution and reuse

The new release uses NVIDIA SimReady assets pinned to revision a72e6a76d53501cbf5a719910a674eb2c1f6f161, with original CC BY 4.0 notices included. Converted geometry and web-texture modifications are described in the release README. Three.js retains its MIT notice. Legacy files retain their own original licenses; no single blanket license is asserted for this mixed archive.

Source and rebuild scripts: https://github.com/AIGeeksGroup/DragMesh-3/tree/agentic-scene/sites/agentic-scenes (repository access may be required; the downloaded demo runs independently).

Downloads last month
98