Dataset of DeformPAM

Contents

• Description
• Structure
• Usage

Description

This is the dataset used in the paper DeformPAM: Data-Efficient Learning for Long-horizon Deformable Object Manipulation via Preference-based Action Alignment.

• Paper
• Project Homepage
• Github Repository

Structure

We offer two versions of the dataset: one is the full dataset used to train the models in our paper, and the other is a mini dataset for easier examination. Both datasets include data for the supervised and finetuning stages of granular pile shaping, rope shaping, and T-shirt unfolding. Each subset is structured as follows:

├── annotations
│   ├── 0aa71092-06c1-4d3f-8f70-e0bf86eeaeab
│   │   └── metadata.yaml annotations and other detailed information
│   ├── ...
└── observations
    ├── 0aa71092-06c1-4d3f-8f70-e0bf86eeaeab
    │   ├── mask
    │   │   └── begin.png mask img used for segmenting the point cloud
    │   ├── metadata.yaml detailed information
    │   ├── pcd
    │   │   ├── processed_begin.npz segmented point cloud of the object; processed_begin["points"]: np.ndarray (N, 3) float16
    │   │   └── raw_begin.npz raw point cloud of the object; raw_begin["points"]: np.ndarray (N, 3) float16
    │   └── rgb
    │       └── begin.jpg RGB image of the object
    ├── ...

Usage

There are two ways to utilize the dataset for training:

• Install the tool according to the data management toolkit's installation guide, and then store the metadata to MongoDB.
• Or, you can modify the dataset to load data from local files.