yonghoon96/antmaze-giant-navigate-10m-v0

AntMaze-Giant-Navigate 10M Dataset

This dataset contains 10M transitions for the antmaze-giant-navigate task from OGBench.

Dataset Structure

• 10 files of ~1.1M transitions each (split for easier loading)
• Each file includes:
  • Training set: 500 episodes × 2,001 steps
  • Validation set: 50 episodes × 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-v0-009.npz / -009-val.npz (seed=9)

Usage

import numpy as np
from huggingface_hub import hf_hub_download

# Download a specific file
file_path = hf_hub_download(
    repo_id="your-username/antmaze-giant-navigate-10m-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

@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)