Hafnium49's picture
Add first 50 episodes (GR00T N1.5 training data)
b42e4f7 verified

Paper Return - First 50 Episodes (GR00T N1.5 Training Data)

Overview

This dataset contains the first 50 episodes (episodes 0-49) from the paper_return dataset.

This subset represents the exact training data used to fine-tune the phospho-app/gr00t-paper_return-7w9itxzsox GR00T N1.5 model.

Dataset Contents

  • Episodes: 50 (episodes 000000-000049)
  • Task: Paper-in-square manipulation task using SO-101 robot
  • Data format: LeRobot v2.1 dataset format
  • Total frames: ~69,051 frames across 50 episodes

Files Structure

paper_return_first50/
├── data/
│   └── chunk-000/
│       ├── episode_000000.parquet  # Action and state data
│       ├── ...
│       └── episode_000049.parquet
├── videos/
│   └── chunk-000/
│       ├── observation.images.main/
│       │   ├── episode_000000.mp4  # Top-view camera
│       │   ├── ...
│       │   └── episode_000049.mp4
│       └── observation.images.secondary_0/
│           ├── episode_000000.mp4  # Wrist camera
│           ├── ...
│           └── episode_000049.mp4
└── meta/
    ├── info.json           # Dataset metadata
    ├── stats.json          # Statistics for normalization
    ├── episodes.jsonl      # Episode information
    └── tasks.jsonl         # Task definitions

Training Details

This dataset was used to train the GR00T N1.5 model with the following configuration:

  • Model: GR00T N1.5 (3B parameters)
  • Training time: ~3 hours on A100/H100 GPUs
  • Epochs: 10
  • Batch size: 49
  • Learning rate: 0.0001
  • Platform: phosphobot

Related Resources

Usage

Load with LeRobot:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("Hafnium49/paper_return_first50")
print(f"Episodes: {dataset.num_episodes}")
print(f"Frames: {len(dataset)}")

Citation

If you use this dataset, please cite:

  • Original dataset: Hafnium49/paper_return
  • GR00T model: phospho-app/gr00t-paper_return-7w9itxzsox
  • GR00T foundation model: NVIDIA GR00T N1.5

License

Same license as the original paper_return dataset.

Dataset Creation

Created from episodes 0-49 of the paper_return dataset to represent the exact training data used for the GR00T N1.5 fine-tuning conducted on the phospho platform.