---
license: mit
task_categories:
- robotics
---

# Dataset of Reactive Diffusion Policy

## Contents

- [Description](#description)
- [Structure](#structure)
- [Usage](#usage)
- [Tactile Dataset](#tactile-dataset)

## Description

This is the raw and postprocessed dataset used in the paper
[Reactive Diffusion Policy: Slow-Fast Visual-Tactile Policy Learning for Contact-Rich Manipulation](https://reactive-diffusion-policy.github.io).

- [Paper](https://arxiv.org/pdf/2503.02881)
- [Project Homepage](https://reactive-diffusion-policy.github.io)
- [GitHub Repository](https://github.com/xiaoxiaoxh/reactive_diffusion_policy)
- [Pretrained Models](https://huggingface.co/WendiChen/reactive_diffusion_policy_model)

## Structure

We offer two versions of the dataset:
one is the [full dataset](https://huggingface.co/datasets/WendiChen/reactive_diffusion_policy_dataset/tree/main/dataset_full) used to train the models in our paper,
and the other is a [mini dataset](https://huggingface.co/datasets/WendiChen/reactive_diffusion_policy_dataset/tree/main/dataset_mini) for easier examination.
Both versions include raw and postprocessed subsets of peeling, wiping and lifting.

Each raw subset is structured as follows:
```
subset_name
├── seq_01.pkl
├── seq_02.pkl
├── ...
```
> Note that we split the full raw lifting subset into 2 parts due to file size restrictions.

Each postprocessed subset is stored in [Zarr](https://zarr.dev) format, which is structured as follows:
```
 ├── action (25710, 4) float32
 ├── external_img (25710, 240, 320, 3) uint8
 ├── left_gripper1_img (25710, 240, 320, 3) uint8
 ├── left_gripper1_initial_marker (25710, 63, 2) float32
 ├── left_gripper1_marker_offset (25710, 63, 2) float32
 ├── left_gripper1_marker_offset_emb (25710, 15) float32
 ├── left_gripper2_img (25710, 240, 320, 3) uint8
 ├── left_gripper2_initial_marker (25710, 25, 2) float32
 ├── left_gripper2_marker_offset (25710, 25, 2) float32
 ├── left_gripper2_marker_offset_emb (25710, 15) float32
 ├── left_robot_gripper_force (25710, 1) float32
 ├── left_robot_gripper_width (25710, 1) float32
 ├── left_robot_tcp_pose (25710, 9) float32
 ├── left_robot_tcp_vel (25710, 6) float32
 ├── left_robot_tcp_wrench (25710, 6) float32
 ├── left_wrist_img (25710, 240, 320, 3) uint8
 ├── right_robot_gripper_force (25710, 1) float32
 ├── right_robot_gripper_width (25710, 1) float32
 ├── right_robot_tcp_pose (25710, 9) float32
 ├── right_robot_tcp_vel (25710, 6) float32
 ├── right_robot_tcp_wrench (25710, 6) float32
 ├── target (25710, 4) float32
 └── timestamp (25710,) float32
```

## Usage

Follow the [README in our GitHub repo](https://github.com/xiaoxiaoxh/reactive_diffusion_policy)
to [postprocess the raw data](https://github.com/xiaoxiaoxh/reactive_diffusion_policy#data-postprocessing)
and [train the model](https://github.com/xiaoxiaoxh/reactive_diffusion_policy#-training).

## Tactile Dataset
We also provide the raw videos of the [tactile dataset](https://huggingface.co/datasets/WendiChen/reactive_diffusion_policy_dataset/tree/main/dataset_tactile_embedding) used for generate the PCA embedding in our paper.