---
license: apache-2.0
tags:
  - robotics
  - unitree-g1
  - pointcloud
  - navigation
size_categories:
  - n<1K
---

# G1 Physical AI Maps

3D point cloud maps and navigation goals collected with a Unitree G1.

| map | points | size (m) | tasks |
|---|---|---|---|
| `lab` | 19,124 | 12.8 x 12.4 x 3.4 | goal_1 ~ goal_4 |

## Layout

```
maps/g1/lab/
  map.pcd          point cloud (PCD binary; x y z intensity normal_x normal_y normal_z curvature)
  tasklist.json    goal poses (position + quaternion in the map frame; array order is registration order)
  metadata.json    bounds, md5
```

## Usage

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "mncai/G1_physical_ai_maps",
    "maps/g1/lab/map.pcd",
    repo_type="dataset",
    revision="map-g1-lab-v1",   # omit for latest
)
```

- Map versions are pinned with git tags.
- Verify downloads against the `md5` values in `metadata.json`.
- Coordinates are in the `map` frame, gravity aligned and z-up, so no extra transform is needed.
- The HF dataset viewer cannot render `.pcd`, so 3D inspection is done in a separate Space.