README.md

---
license: mit
task_categories:
- robotics
- reinforcement-learning
tags:
- robotics
- teleoperation
- humanoid
- manipulation
- GR1
- GR00T
- imitation-learning
pretty_name: GR1 Robot Teleoperation Dataset
size_categories:
- 10M<n<100M
---

# GR1 Robot Teleoperation Dataset

A large-scale humanoid robot teleoperation dataset collected using the Fourier GR-1 (GR1T1) robot, part of the [GR00T](https://developer.nvidia.com/gr00t) robotics initiative.

This is a true subset of [nvidia/PhysicalAI-Robotics-GR00T-Teleop-GR1](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-Teleop-GR1).

## Dataset Summary

- **Episodes**: 22,209
- **Total frames**: 6,362,293
- **Total tasks**: 46,669
- **Videos**: 22,209
- **Frequency**: 20 FPS
- **Robot**: GR1T1 (Fourier GR-1 humanoid)

The dataset contains egocentric video observations (1280x800) paired with full-body robot states and actions across 44 degrees of freedom, including dual arms, dual dexterous hands, waist, and neck.

## Repository Structure

```
GR1_robot/
├── meta/                   # Metadata (8 files)
│   ├── info.json           # Dataset summary & feature schema
│   ├── stats.json          # Per-feature statistics (mean, std, min, max, quantiles)
│   ├── metadata.json       # Full metadata with model configuration
│   ├── modality.json       # Modality-to-key mapping
│   ├── embodiment.json     # Robot embodiment info
│   ├── episodes.jsonl      # Per-episode metadata (22,209 lines)
│   ├── tasks.jsonl         # Per-task metadata (46,669 lines)
│   └── initial_actions.npz # Initial action array
├── data/                   # Trajectory data as Parquet files (2.9 GB)
│   └── chunk-{000..022}/episode_{000000..022208}.parquet
└── videos/                 # Egocentric videos as MP4 files (37.2 GB)
    └── chunk-{000..022}/observation.images.ego_view_freq20/episode_{000000..022208}.mp4
```

## Features

Each episode Parquet file contains per-timestep records:

| Feature | Dtype | Shape | Description |
|---------|-------|-------|-------------|
| `observation.state` | float64 | (44,) | Full-body joint positions |
| `action` | float64 | (44,) | Full-body joint actions |
| `timestamp` | float64 | (1,) | Timestep timestamp (s) |
| `next.reward` | float64 | (1,) | Reward |
| `next.done` | bool | (1,) | Episode termination flag |
| `task_index` | int64 | (1,) | Task identifier |
| `episode_index` | int64 | (1,) | Episode index |
| `index` | int64 | (1,) | Global frame index |
| `annotation.human.*` | int64 | (1,) | Human annotations (verb, object, location, hand, rating, etc.) |

### State / Action Space

Both `observation.state` and `action` are 44-dimensional vectors:

| Group | Start | End | Dims | Description |
|-------|-------|-----|------|-------------|
| left_arm | 0 | 7 | 7 | Left arm joints |
| left_hand | 7 | 13 | 6 | Left dexterous hand |
| left_leg | 13 | 19 | 6 | Left leg (passive) |
| neck | 19 | 22 | 3 | Neck joints |
| right_arm | 22 | 29 | 7 | Right arm joints |
| right_hand | 29 | 35 | 6 | Right dexterous hand |
| right_leg | 35 | 41 | 6 | Right leg (passive) |
| waist | 41 | 44 | 3 | Waist joints |

### Video

- **Resolution**: 1280x800
- **FPS**: 20
- **Codec**: H.264 (YUV420P)
- **View**: Egocentric (robot head camera)

## Splits

The dataset uses a single training split (`0:100`). Normalization statistics (mean, std, min, max, quantiles) for all features are provided in `meta/stats.json`.

## Usage

### With Hugging Face `datasets`

```python
from datasets import load_dataset

# Load the full dataset (streaming recommended for this 40GB+ dataset)
dataset = load_dataset("Physis-AI/GR1_robot", split="train", streaming=True)

for episode in dataset:
    states = episode["observation.state"]  # (T, 44)
    actions = episode["action"]            # (T, 44)
    break
```

To load only metadata without downloading videos/parquets:

```python
from datasets import load_dataset

dataset = load_dataset("Physis-AI/GR1_robot", split="train",
                       data_files="data/**/*.parquet", streaming=True)
```

### With Pandas (local files)

```python
import pandas as pd

ep = pd.read_parquet("data/chunk-000/episode_000000.parquet")
states = ep["observation.state"].values   # (T, 44)
actions = ep["action"].values             # (T, 44)
left_arm_state = states[:, 0:7]
```

### Loading Videos

```python
import cv2

cap = cv2.VideoCapture("videos/chunk-000/observation.images.ego_view_freq20/episode_000000.mp4")
while cap.isOpened():
    ret, frame = cap.read()
    if not ret:
        break
    # frame is (800, 1280, 3) BGR
```

## Attribution

This dataset is a true subset of [nvidia/PhysicalAI-Robotics-GR00T-Teleop-GR1](https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-GR00T-Teleop-GR1).

## License

MIT