BallNet-100K

This dataset contains 100K samples of data for BallNet, generated by finite element simulations.

Dataset Schema

The dataset is stored in Parquet format and organized as follows:

data/
├── train/
│ ├── data_00000.parquet
│ ├── data_00001.parquet
│ └── ...
└── test/
├── data_00000.parquet
└── ...

Each sample in this dataset contains three components:

Field Name	Type	Shape	Description
motion	List[float64]	[6]	The 6D motion of the finger, including translation (dx, dy, dz) and rotation (rx, ry, rz) in mm and rad.
force	List[float64]	[6]	The 6D force and torque on the bottom surface of the finger, corresponding to (fx, fy, fz, tx, ty, tz) in N and Nmm.
nodes	List[float64]	[3N]	The 3D displacement of N surface nodes of the finger, where each node is represented as [dx, dy, dz] in mm.

Usage

from datasets import load_dataset

dataset = load_dataset("han-xudong/ballnet-100k")

# Access the 'train' subset
for sample in dataset['train']:
    motion = sample['motion']
    force = sample['force']
    nodes = sample['nodes']

# Access the 'test' subset
for sample in dataset['test']:
    motion = sample['motion']
    force = sample['force']
    nodes = sample['nodes']

Citation

If you use this model in your research, please cite:

@article{han2025anchoring,
    title={Anchoring Morphological Representations Unlocks Latent Proprioception in Soft Robots},
    author={Han, Xudong and Guo, Ning and Xu, Ronghan and Wan, Fang and Song, Chaoyang},
    journal={Advanced Intelligent Systems},
    volume={7},
    pages={e202500444},
    year={2025}
}