DexSkin: High-Coverage Conformable Robotic Skin for Learning Contact-Rich Manipulation (CoRL 2025)

Authors: Suzannah Wistreich*, Baiyu Shi*, Stephen Tian*, Samuel Clarke, Michael Nath, Chengyi Xu, Zhenan Bao, Jiajun Wu
Stanford University, UAB

[Paper] | [Project Page] | [Code]

This repository contains real-world robot demonstration datasets collected with DexSkin. We release the three task datasets described in our paper:

Dataset	Task	Tactile	Demos
berry_manip.hdf5	Berry transporting	✗	25
pen_reorientation.hdf5	In-hand pen reorientation	✓	50
box_packaging.hdf5	Box packaging with rubber band	✓	50

Note on berry_manip.hdf5: This dataset serves as the training data for the no-tactile base policy used as the starting point for the online RL experiment described in the paper. Therefore, this dataset does not contain tactile fields (tactile, baseline, std).

---

Data Structure

Each file is an HDF5 dataset organized as follows. All fields are time-indexed along the first dimension (i.e., shape (T, ...)). Fields used in training are indicated in bold.

/data/
├── demo_0/
│   ├── O_T_EE                      # (T, 16)   End-effector transform (flattened 4x4)
│   ├── color_image                 # (T, 480, 640, 3)
│   ├── commanded_joint_positions   # (T, 8)
│   ├── cos_q                       # (T, 7)
│   ├── sin_q                       # (T, 7)
│   ├── q                           # (T, 7)    Joint positions 
│   ├── dq                          # (T, 7)
│   ├── ee_pos_quat                 # (T, 7)
│   ├── eef_pos                     # (T, 3)
│   ├── eef_quat                    # (T, 4)
│   ├── gripper_current             # (T,)
│   ├── gripper_position            # (T,)
│   ├── gripper_width               # (T,)
│   ├── gripper_recent_commanded_*  # (T,)  current / position / speed
│   ├── joint_positions             # (T, 8)    arm + gripper
│   ├── joint_velocities            # (T, 8)
│   ├── time                        # (T,)
│   ├── tactile                     # (T, 120) 
│   ├── baseline                    # (10, 12)  
│   └── std                         # (10, 12)  
├── demo_1/
└── ...

We use a non-commercial end-effector to collect our demonstrations. Tactile signal noise in the provided data is minimal, but can be reduced further with a commercial end-effector.

---

BibTex

@inproceedings{wistreich2025dexskin,
    title={DexSkin: High-Coverage Conformable Robotic Skin for Learning Contact-Rich Manipulation},
    author={Suzannah Wistreich and Baiyu Shi and Stephen Tian and Samuel Clarke and Michael Nath and Chengyi Xu and Zhenan Bao and Jiajun Wu},
    booktitle={Conference on Robot Learning (CoRL)},
    year={2025},
    eprint={2509.18830},
    archivePrefix={arXiv},
    primaryClass={cs.RO},
    url={https://arxiv.org/abs/2509.18830},
}