README.md

---
language:
- en
license: apache-2.0
task_categories:
- robotics
tags:
- RDT
- rdt
- RDT 2
- manipulation
- bimanual
- ur5e
- webdataset
- vision-language-action
arxiv: 2602.03310
---

# RDT2 Dataset

[Project page](https://rdt-robotics.github.io/rdt2/) | [Paper](https://huggingface.co/papers/2602.03310) | [GitHub](https://github.com/thu-ml/RDT2)

## Dataset Summary

This dataset provides shards in the **WebDataset** format for fine-tuning [RDT-2](https://rdt-robotics.github.io/rdt2/) or other policy models on **bimanual manipulation**.
Each sample packs:

* a **binocular RGB image** (left + right wrist cameras concatenated horizontally)
* a **relative action chunk** (continuous control, 0.8s, 30Hz)
* a **discrete action token sequence** (e.g., from an [Residual VQ action tokenizer](https://huggingface.co/robotics-diffusion-transformer/RVQActionTokenizer))
* a **metadata JSON** with an instruction key `sub_task_instruction_key` to index corresponding instruction from `instructions.json`

Data were collected on a **bimanual UR5e** setup.

---

## Supported Tasks

* **Instruction-conditioned bimanual manipulation**, including:
  - Pouring water: different water bottles and cups
  - Cleaning the desktop: different dustpans and paper balls
  - Folding towels: towels of different sizes and colors
  - Stacking cups: cups of different sizes and colors

---

## Data Structure

### Shard layout

Shards are named `shard-*.tar`. Inside each shard:

```
shard-000000.tar
├── 0.image.jpg          # binocular RGB, H=384, W=768, C=3, uint8
├── 0.action.npy         # relative actions, shape (24, 20), float32
├── 0.action_token.npy   # action tokens, shape (27,), int16 ∈ [0, 1024)
├── 0.meta.json          # metadata; includes "sub_task_instruction_key"
├── 1.image.jpg
├── 1.action.npy
├── 1.action_token.npy
├── 1.meta.json
└── ...
shard-000001.tar
shard-000002.tar
...
```

> **Image:** binocular wrist cameras concatenated horizontally → `np.ndarray` of shape `(384, 768, 3)` with `dtype=uint8` (stored as JPEG).
> 
> **Action (continuous):** `np.ndarray` of shape `(24, 20)`, `dtype=float32` (24-step chunk, 20-D control).
> 
> **Action tokens (discrete):** `np.ndarray` of shape `(27,)`, `dtype=int16`, values in `[0, 1024]`.
> 
> **Metadata:** `meta.json` contains at least `sub_task_instruction_key` pointing to an entry in top-level `instructions.json`.


---

## Example Data Instance

```json
{
  "image": "0.image.jpg",
  "action": "0.action.npy",
  "action_token": "0.action_token.npy",
  "meta": {
    "sub_task_instruction_key": "fold_cloth_step_3"
  }
}
```

---

## How to Use

### 1) Official Guidelines to fine-tune RDT 2 series

Use the official [scripts](https://github.com/thu-ml/RDT2) and [guidelines](https://github.com/thu-ml/RDT2) provided in the GitHub repository.

### 2) Minimal Loading example 

```python
import os
import glob
import json
import random

import webdataset as wds


def no_split(src):
    yield from src

def get_train_dataset(shards_dir):
    shards = sorted(glob.glob(os.path.join(shards_dir, "shard-*.tar")))
    random.shuffle(shards)
    
    num_workers = wds.utils.pytorch_worker_info()[-1]
    workersplitter = wds.split_by_worker if len(shards) > num_workers else no_split
    
    assert shards, f"No shards under {shards_dir}"
    dataset = (
        wds.WebDataset(
            shards,
            shardshuffle=False,
            nodesplitter=no_split,
            workersplitter=workersplitter,
            resampled=True,
        )
        .repeat()
        .shuffle(8192, initial=8192)
        .decode("pil")
        .map(
            lambda sample: {
                "image": sample["image.jpg"],
                "action_token": sample["action_token.npy"],
                "meta": sample["meta.json"],
            }
        )
        .with_epoch(nsamples=(2048 * 30 * 60 * 60))    # 2048 hours
    )
    
    return dataset

with open(os.path.join("<Dataset Directory>", "instructions.json")) as fp:
    instructions = json.load(fp)
dataset = get_train_dataset(os.path.join("<Dataset Directory>", "shards"))
```
---

## Ethical Considerations

* Contains robot teleoperation/automation data. No PII is present by design.
* Ensure safe deployment/testing on real robots; follow lab safety and manufacturer guidelines.

---

## Citation

```bibtex
@article{rdt2,
  title={RDT2: Exploring the Scaling Limit of UMI Data Towards Zero-Shot Cross-Embodiment Generalization},
  author={RDT Team},
  journal={arXiv preprint arXiv:2602.03310},
  year={2026}
}
```

---

## License

* **Dataset license:** Apache-2.0.
* Ensure compliance when redistributing derived data or models.