yonghoon96/antmaze-giant-navigate-20m-v0

---
license: mit
task_categories:
- reinforcement-learning
tags:
- offline-rl
- goal-conditioned-rl
- mujoco
- antmaze
size_categories:
- 10M<n<100M
---

# AntMaze-Giant-Navigate 20M Dataset

This dataset contains 20M transitions for the `antmaze-giant-navigate` task from [OGBench](https://github.com/seohongpark/ogbench).

## Dataset Structure

- **20 files** of ~1.1M transitions each (split for easier loading)
- Each file includes:
- Training set: 500 episodes x 2,001 steps
- Validation set: 50 episodes x 2,001 steps

## Files

```
antmaze-giant-navigate-v0-000.npz / -000-val.npz (seed=0)
antmaze-giant-navigate-v0-001.npz / -001-val.npz (seed=1)
...
antmaze-giant-navigate-v0-019.npz / -019-val.npz (seed=19)
```

## Usage

```python
import numpy as np
from huggingface_hub import hf_hub_download

# Download a specific file
file_path = hf_hub_download(
repo_id="yonghoon96/antmaze-giant-navigate-20m-v0",
filename="antmaze-giant-navigate-v0-000.npz",
repo_type="dataset"
)

# Load dataset
data = np.load(file_path)
print(data.files) # ['observations', 'actions', 'terminals', 'qpos', 'qvel']
```

## Generation Details

- **Environment**: `antmaze-giant-v0`
- **Dataset Type**: navigate
- **Expert Policy**: SAC agent trained for 400K steps
- **Action Noise**: Gaussian noise (sigma=0.2)
- **Generation Script**: OGBench v1.2.1 `generate_locomaze.py`

## Citation

```bibtex
@inproceedings{ogbench_park2025,
title={OGBench: Benchmarking Offline Goal-Conditioned RL},
author={Park, Seohong and Frans, Kevin and Eysenbach, Benjamin and Levine, Sergey},
booktitle={International Conference on Learning Representations (ICLR)},
year={2025},
}
```

## License

MIT License (same as OGBench)