README.md

---
license: apache-2.0
language:
  - en
tags:
  - robotics
  - manipulation
  - locomotion
  - whole-body-control
  - dexterous-hands
  - humanoid
  - lerobot
  - groot
  - unitree-g1
  - brainco
  - imitation-learning
  - long-horizon
  - household
task_categories:
  - robotics
pretty_name: G1 BrainCo WBT GR00T Dataset
size_categories:
  - 100K<n<1M
configs:
  - config_name: G1_WBT_Brainco_Collect_Plates_Into_Dishwasher
    data_files: G1_WBT_Brainco_Collect_Plates_Into_Dishwasher/data/**/*.parquet
  - config_name: G1_WBT_Brainco_Make_The_Bed
    data_files: G1_WBT_Brainco_Make_The_Bed/data/**/*.parquet
  - config_name: G1_WBT_Brainco_Pick_Up_Medicine
    data_files: G1_WBT_Brainco_Pick_Up_Medicine/data/**/*.parquet
  - config_name: G1_WBT_Brainco_Pickup_Pillow
    data_files: G1_WBT_Brainco_Pickup_Pillow/data/**/*.parquet
---

# G1-BrainCo-GR00T-Dataset

Unitree G1 humanoid robot + BrainCo 6-DOF dexterous hands demonstration dataset,
collected as part of the [UnifoLM WBT Dataset](https://huggingface.co/collections/unitreerobotics/unifolm-wbt-dataset) (BrainCo subset).

Converted to **LeRobot v2.1** format and prepared for fine-tuning [NVIDIA GR00T N1.7](https://huggingface.co/nvidia/GR00T-N1.7-3B).

---

## Tasks

| Dataset | Episodes | Avg Duration | Task Description |
|---------|----------|-------------|-----------------|
| `G1_WBT_Brainco_Collect_Plates_Into_Dishwasher` | 300 | ~47s | Collect plates from table and load into dishwasher |
| `G1_WBT_Brainco_Make_The_Bed` | 300 | ~23s | Straighten and make a bed neatly |
| `G1_WBT_Brainco_Pick_Up_Medicine` | 300 | ~26s | Open drawer and pick up medicine bottle |
| `G1_WBT_Brainco_Pickup_Pillow` | 300 | ~20s | Pick up pillow and place on sofa |

**Total: 1200 episodes · 4 tasks · 30 FPS · 3 synchronized cameras**

---

## Robot Setup

| Component | Spec |
|-----------|------|
| Robot | Unitree G1 humanoid |
| Hands | BrainCo 6-DOF dexterous hands (left + right) |
| Cameras | Head stereo left (front) + left wrist + right wrist |
| Control rate | 30 Hz demos → 500 Hz onboard PD execution |

---

## Data Format

**LeRobot v2.1** (converted from original v3.0 via `convert_v3_to_v2.py`)

### Observations

| Key | Dims | Description |
|-----|------|-------------|
| `observation.images.head_stereo_left` | 480×640×3 | Front-facing stereo camera (left) |
| `observation.images.wrist_left` | 480×640×3 | Left wrist camera |
| `observation.images.wrist_right` | 480×640×3 | Right wrist camera |
| `observation.state.ee_state` | 12D | Bimanual EEF pose (left: pos+quat, right: pos+quat) |
| `observation.state.hand_state` | 12D | BrainCo finger positions (left: 6D, right: 6D), range [0, 0.8] |
| `observation.state.robot_q_current` | 36D | Full-body joint angles in radians |

### Actions

| Key | Dims | Representation | Description |
|-----|------|---------------|-------------|
| `action.ee_action` | 12D | RELATIVE delta | EEF position+orientation deltas (bimanual) |
| `action.hand_cmd` | 12D | ABSOLUTE | BrainCo finger commands (left: 6D, right: 6D), range [0, 0.8] |
| `action.robot_q_desired` | 36D | ABSOLUTE | Full-body joint position targets in radians |

### Notes on non-standard parquet layout

Unlike standard LeRobot datasets, state and action columns are stored as **separate named columns** (not concatenated into `observation.state` / `action`). A `meta/modality.json` file with `original_key` mappings is included so GR00T's dataloader resolves them correctly.

---

## Directory Structure

```
G1_WBT_Brainco_<Task>/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet
│       ├── episode_000001.parquet
│       └── ...
├── videos/
│   └── chunk-000/
│       ├── observation.images.head_stereo_left_episode_000000.mp4
│       ├── observation.images.wrist_left_episode_000000.mp4
│       └── ...
└── meta/
    ├── modality.json         ← GR00T-specific key mappings (manipulation)
    ├── modality_wbc.json     ← GR00T-specific key mappings (WBC / loco-manipulation)
    ├── episodes.jsonl
    ├── tasks.jsonl
    ├── stats.json
    └── relative_stats.json
```

---

## GR00T N1.7 Fine-tuning

This dataset was used to fine-tune [GR00T N1.7](https://github.com/NVIDIA/Isaac-GR00T) with three runs:

| Run | Checkpoint | Action Space | Purpose |
|-----|-----------|-------------|---------|
| Run 1 | `nvidia/GR00T-N1.7-DROID` | 24D (arms + hands) | Manipulation — DROID init |
| Run 2 | `nvidia/GR00T-N1.7-3B` | 60D (arms + hands + full body) | Loco-manipulation (WBC) |
| Run 3 | `nvidia/GR00T-N1.7-3B` | 24D (arms + hands) | Manipulation — base init (baseline) |

### Quick start (manipulation)

```bash
git clone https://github.com/NVIDIA/Isaac-GR00T
cd Isaac-GR00T
uv sync --all-extras

bash examples/G1_Brainco/finetune_g1_brainco.sh \
    --base-model-path nvidia/GR00T-N1.7-DROID \
    --dataset-path dataset_g1/brainco/G1_WBT_Brainco_Collect_Plates_Into_Dishwasher:... \
    --output-dir outputs/g1_brainco
```

### Memory requirements (RTX 5090 / 32 GB)

```bash
# Required flags for 32 GB VRAM:
--gradient_checkpointing
--optim paged_adamw_8bit
--global_batch_size 1
--gradient_accumulation_steps 64
# Peak VRAM: ~26.5 GB
```

> **Important:** Always pass `--action-horizon 16` at eval/deployment.
> The DROID config defaults to 40 steps; without this flag the server outputs wrong chunk sizes.

---

## Dataset Statistics

| Task | Min | Median | Max | Std% |
|------|-----|--------|-----|------|
| Collect Plates | 19.5s | 43.3s | 124.3s | 30% |
| Make the Bed | 12.6s | 21.7s | 46.1s | 29% |
| Pick Up Medicine | 13.3s | 24.2s | 56.1s | 30% |
| Pickup Pillow | 12.3s | 18.9s | 37.2s | 23% |

All episodes: zero NaNs, `hand_cmd` range [0.0, 0.8], `ee_action` within ±π.

---

## Related Dataset

For single-object pick & grasp tasks (8 tasks, 1598 episodes, direct joint-space actions):
👉 [JeffrinSam/G1-BrainCo-Pick-GR00T](https://huggingface.co/datasets/JeffrinSam/G1-BrainCo-Pick-GR00T)

---

## Citation

```bibtex
@dataset{unifolm_wbt_brainco_g1,
  title   = {G1-BrainCo-GR00T-Dataset},
  author  = {JeffrinSam},
  year    = {2026},
  note    = {Unitree G1 + BrainCo dexterous hands manipulation dataset in LeRobot v2.1 format, prepared for GR00T N1.7 fine-tuning},
  url     = {https://huggingface.co/datasets/JeffrinSam/G1-BrainCo-GR00T-Dataset}
}
```

---

## License

Data collected from [UnifoLM WBT Dataset](https://huggingface.co/collections/unitreerobotics/unifolm-wbt-dataset) by Unitree Robotics.
Please refer to the original dataset license before commercial use.