---
license: mit
task_categories:
- robotics
- reinforcement-learning
tags:
- metaworld
- robotics
- manipulation
- multi-task
- r3m
- vision-language
- imitation
size_categories:
- 1K<n<10K
language:
- en
pretty_name: Short-MetaWorld Dataset
dataset_info:
  features:
  - name: image
    dtype: image
  - name: state
    dtype: 
      sequence: float32
  - name: action
    dtype:
      sequence: float32
  - name: prompt
    dtype: string
  - name: task_name
    dtype: string
  splits:
  - name: train
    num_bytes: 1900000000
    num_examples: 40000
  download_size: 1900000000
  dataset_size: 1900000000
---

# Short-MetaWorld Dataset

## Overview

Short-MetaWorld is a curated dataset from Meta-World containing **Multi-Task 10 (MT10)** and **Meta-Learning 10 (ML10)** tasks with **100 successful trajectories per task** and **20 steps per trajectory**. This dataset is specifically designed for multi-task robot learning, imitation learning, and vision-language robotics research.

## 🚀 Quick Start

```python
from short_metaworld_loader import load_short_metaworld
from torch.utils.data import DataLoader

# Load the dataset
dataset = load_short_metaworld("./", image_size=224)

# Create a DataLoader
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

# Get a sample
sample = dataset[0]
print(f"Image shape: {sample['image'].shape}")
print(f"State: {sample['state']}")
print(f"Action: {sample['action']}")
print(f"Task: {sample['task_name']}")
print(f"Prompt: {sample['prompt']}")
```

## 📁 Dataset Structure

```
short-MetaWorld/
├── README.txt                     # Original dataset documentation
├── short-MetaWorld/
│   ├── img_only/                    # 224x224 RGB images
│   │   ├── button-press-topdown-v2/
│   │   │   ├── 0/                   # Trajectory 0
│   │   │   │   ├── 0.jpg           # Step 0
│   │   │   │   ├── 1.jpg           # Step 1
│   │   │   │   └── ...
│   │   │   ├── 1/                   # Trajectory 1
│   │   │   └── ...
│   │   ├── door-open-v2/
│   │   └── ...
│   └── r3m-processed/              # R3M processed features
│       └── r3m_MT10_20/
│           ├── button-press-topdown-v2.pkl
│           ├── door-open-v2.pkl
│           └── ...
└── r3m-processed/                  # Additional R3M data
    └── r3m_MT10_20/
├── mt50_task_prompts.json          # Task descriptions & prompts
├── short_metaworld_loader.py       # Dataset loader
└── requirements.txt
```

## 🎯 Tasks Included

### Multi-Task 10 (MT10)
- `button-press-topdown-v2` - Press button from above
- `door-open-v2` - Open door by pulling handle  
- `drawer-close-v2` - Close drawer
- `drawer-open-v2` - Open drawer
- `peg-insert-side-v2` - Insert peg into hole
- `pick-place-v2` - Pick up object and place on target

### Meta-Learning 10 (ML10)
Additional tasks for meta-learning evaluation.

## 📊 Data Format

- **Images**: 224×224 RGB images in JPEG format
- **States**: 7-dimensional robot state vectors (joint positions)
- **Actions**: 4-dimensional continuous control actions
- **Prompts**: Natural language task descriptions in 3 styles:
  - `simple`: Brief task description
  - `detailed`: Comprehensive task explanation  
  - `task_specific`: Context-specific variations
- **R3M Features**: Pre-processed visual representations using R3M model

## 💾 Loading the Dataset

The dataset comes with a comprehensive loader (`short_metaworld_loader.py`):

```python
# Load specific tasks
mt10_tasks = [
    "reach-v2", "push-v2", "pick-place-v2", "door-open-v2", 
    "drawer-open-v2", "drawer-close-v2", "button-press-topdown-v2",
    "button-press-v2", "button-press-wall-v2", "button-press-topdown-wall-v2"
]
dataset = load_short_metaworld("./", tasks=mt10_tasks)

# Load all available tasks
dataset = load_short_metaworld("./")

# Get dataset statistics
stats = dataset.get_dataset_stats()
print(f"Total steps: {stats['total_steps']}")
print(f"Tasks: {stats['tasks']}")

# Get task-specific prompts
task_info = dataset.get_task_info("pick-place-v2")
print(task_info['detailed'])  # Detailed task description
```

## 🔬 Research Applications

This dataset is designed for:

- **Multi-task Reinforcement Learning**: Train policies across multiple manipulation tasks
- **Imitation Learning**: Learn from demonstration trajectories
- **Vision-Language Robotics**: Connect visual observations with natural language instructions
- **Meta-Learning**: Adapt quickly to new manipulation tasks
- **Robot Policy Training**: End-to-end visuomotor control

## 📈 Dataset Statistics

- **Total trajectories**: 2,000 (100 per task × 20 tasks)
- **Total steps**: ~40,000 (20 steps per trajectory)
- **Image resolution**: 224×224 RGB
- **State dimension**: 7 (robot joint positions)
- **Action dimension**: 4 (continuous control)
- **Dataset size**: ~1.9GB

## 🛠️ Installation

```bash
pip install torch torchvision Pillow numpy
```

## 📖 Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{yu2020meta,
  title={Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning},
  author={Yu, Tianhe and Quillen, Deirdre and He, Zhanpeng and Julian, Ryan and Hausman, Karol and Finn, Chelsea and Levine, Sergey},
  booktitle={Conference on robot learning},
  pages={1094--1100},
  year={2020},
  organization={PMLR}
}

@inproceedings{nair2022r3m,
  title={R3M: A Universal Visual Representation for Robot Manipulation},
  author={Nair, Suraj and Rajeswaran, Aravind and Kumar, Vikash and Finn, Chelsea and Gupta, Abhinav},
  booktitle={Conference on Robot Learning},
  pages={892--902},
  year={2023},
  organization={PMLR}
}
```

## 📧 Contact

- Original dataset: liangzx@connect.hku.hk
- Questions about this upload: Open an issue in the dataset repository

## ⚖️ License

MIT License - See LICENSE file for details.