---
tags:
- robotics
- mujoco
- articulated-objects
- humanoid
---
# DragMesh-3 · artifacts

Assets and evidence for [DragMesh-3](https://github.com/AIGeeksGroup/DragMesh-3).
This model-type repository is an artifact hub, not a newly trained model release.

- **interaction/**: curated 200-rollout baseline, exact asset bundle, manifests,
  audits, selected review videos and frozen pipeline reference/runtime archives.
- **scene/**: seven complete room-scene archives, SHA-256 manifests and model-load
  results. Each archive contains MJCF and every mesh/texture referenced by that MJCF.

Code lives on GitHub: `main` for CoDA → UMR → MuJoCo interaction and controller
adaptation; `scene` for SceneSmith generation/export. SONIC is the current baseline;
SuperMimic training and adapter integration are planned.

## Download

```bash
hf download AIGeeksGroup/DragMesh-3 --include 'interaction/*' --local-dir assets/hub
hf download AIGeeksGroup/DragMesh-3 --include 'scene/*' --local-dir assets/hub
```

Inspect `interaction/source_manifest.json` and `scene/index.json`. Verify the archive
SHA-256 against its manifest before extraction. Extract a room archive into a local
data directory, then open `<room>/scene.xml` in MuJoCo. No absolute mesh paths are needed.

## Provenance and limits

Interaction source: [Neptune-T/DM3-rollouts](https://huggingface.co/datasets/Neptune-T/DM3-rollouts)
at `1ac15a86cfc43abeea33c6eb4c98b4a7475d7854`. Original diagnostic variants, server dumps
and redundant Git bundles remain at the source. The 200 rollouts are not 200 distinct
objects and do not establish general success on new geometry.

Scene source: local working snapshots originating from
[Neptune-T/DragMesh-3-Scene](https://huggingface.co/datasets/Neptune-T/DragMesh-3-Scene)
at `1796aebf7dc5837f88476010ea1da350ccc03241`. Local layouts are repackaged by actual
MJCF dependency, with per-file hashes and model-load records. They are not asserted
to be byte-identical copies of every historical HF file. Duplicate unreferenced meshes,
render frames, caches, API logs, environments and raw source trees are omitted.

Scene compilation and object-animation trajectories are not physical robot task
success. A rollout must pass its own independent contact/goal/retention/collision
audit on its exact scene. Keep failures separate from accepted training records.
All underlying assets/models retain their source terms; no blanket relicensing applies.