charlenechenn/XELA-LEAP

LEAP-XELA Tactile Stiffness and Texture Dataset

Synchronised tactile and proprioceptive recordings from a LEAP robotic hand covered in XELA uSkin magnetic tri-axial sensors. The hand performed six exploratory procedures (EPs) on silicone cubes spanning five stiffness classes and six surface textures. The same contacts are provided as 2D tactile images and as taxel point clouds, so property recovery can be compared across representation formats.

The dataset includes raw recordings, processed tactile representations, and predefined recording-level splits for texture and stiffness classification.

This dataset accompanies the paper Lost in Encoding: Comparing Property Retention in Tactile Representation Formats.

Dataset summary

Hand	LEAP hand, 16 actuated joints
Tactile sensors	XELA uSkin (uSPa 44 phalanges, uSPa 46 palm, uSCu ALHA fingertips), 368 taxels, 3-axis output
Objects	Silicone cubes of identical geometry: Dragon Skin 20/30 and Ecoflex 10/30/50
Labels	Stiffness (5 classes), texture (6 classes), exploratory procedure (6 classes)
Recordings	180 (5 stiffness × 6 texture × 6 EPs)
Frames per recording	750 (cropped or padded)
Total frames	135,000

Exploratory procedures: sliding (passive), sliding (active), pressing, contour following, enclosure, and enclosure + sliding.

Layout

.
├── raw/               # Source .h5 recordings and raw point cloud inputs
├── interim/           # Training-ready outputs: frame-level .parquet files and labelled point cloud .npz files
├── processed/         # Preprocessed datasets (baseline correction, normalisation, temporal differencing)
├── splits.json        # Recording-level train/val/test splits
└── flat_splits.json   # Five-fold cross-validation splits for flat recordings

Splits

File	Description
splits.json	Recording-level train/val/test split, stratified by texture and stiffness labels: 120 / 30 / 30 recordings.
flat_splits.json	Five-fold cross-validation over flat recordings, stratified by texture label: 24 train / 6 test recordings per fold.

All splits are defined at the recording level, so frames from one recording never appear in more than one partition.

Representations

Name	Description
cube	Source tactile recordings: frame-level 2D tactile images of shape (26, 31, 3), with 16-joint position, velocity, and effort measurements.
pointcloud	World-frame taxel point clouds computed by forward kinematics (FK) using both the tactile readings and the recorded joint positions. Stored as NumPy arrays, typically of shape (T, 368, 3): T frames, 368 taxel points, xyz coordinates.
pointcloud_nojoint	Taxel point clouds computed by FK without the recorded joint positions, capturing tactile deformation without the contribution of joint motion.
sensor_frame	Tactile readings in each sensor's own frame, with no hand-level FK transformation. As in the 2D tactile image, each taxel moves relative to its own reference position.

The 2D tactile image maps the 368 taxels onto a 26 × 31 grid that preserves the physical sensor layout; 806 grid cells in total, of which 368 hold physical taxels. Each cell has three channels, one per force axis (f_x, f_y, f_z).

Schemas

Parquet (frame-level, converted from .h5)

Frame-level 2D tactile images, joint measurements, timestamps, and labels.

Column	Type	Notes
recording_id	string	H5 file stem, used as the recording ID
frame_index	int	0 to 749
timestamp_sec	int	ROS image stamp, seconds
timestamp_nanosec	int	ROS image stamp, nanoseconds
texture_label	int	Texture label (0–5), or -1 if unknown
stiffness_label	int	Material/stiffness label (0–4), or -1 if unknown
procedure_label	int	EP/procedure label, or -1 if unknown
image_flat	list<float>	Raw tactile image, flattened 26 × 31 × 3, length 2418
image_delta_flat	list<float>	Optional image difference from the first frame (image - image[0]), length 2418
joint_position	list<float>	Joint positions, length 16
joint_velocity	list<float>	Joint velocities, length 16
joint_effort	list<float>	Joint efforts, length 16
contact_mask_flat	list<bool>	Contact mask over the 26 × 31 grid, length 806
contact_area	int	Number of active contact cells
valid_frame	bool	False for padded rows

Notes

• Sequences are cropped or padded to 750 frames. Filter on valid_frame to drop padding.
• image_delta_flat is computed as image - image[0].
• fixed_encoding is used for squeeze and enclosure data.

Point cloud NPZ (recording-level, converted from calibrated .npy)

Key	Shape	Notes
recording_id	scalar string	Input filename stem, with a trailing _point_cloud removed if present
taxel_positions	(T, N, 3)	Float32 taxel xyz positions; N is inherited from the input (typically 368)
texture_label	scalar	Texture label (0–5), or -1 if unknown
stiffness_label	scalar	Material/stiffness label (0–4), or -1 if unknown
procedure_label	scalar	EP/procedure label, or -1 if unknown

Notes

• Sequences are cropped or padded to 750 frames.

Label mapping

stiffness_label	Material
0	TODO
1	TODO
2	TODO
3	TODO
4	TODO

texture_label	Texture
0	Large dot
1	Small dot
2	Small hole
3	Small indent
4	Fine grain
5	Smooth
Note: From most to least rough

procedure_label	Exploratory procedure
0–5	TODO

Usage

from huggingface_hub import snapshot_download
import json
import numpy as np
import pandas as pd

root = snapshot_download(repo_id="TODO/your-dataset-name", repo_type="dataset")

# Frame-level tactile images and joint states
df = pd.read_parquet(f"{root}/interim/<file>.parquet")
df = df[df["valid_frame"]]
image = np.asarray(df.iloc[0]["image_flat"]).reshape(26, 31, 3)

# Recording-level point cloud
pc = np.load(f"{root}/interim/<file>.npz")
positions = pc["taxel_positions"]          # (T, N, 3)
stiffness = int(pc["stiffness_label"])

# Recording-level splits
with open(f"{root}/splits.json") as f:
    splits = json.load(f)

Replace <file> with an actual filename from the repository.

Collection

Recordings were captured through a ROS 2 pipeline that replays prerecorded EPs on the LEAP hand and records XELA tactile frames in sync with joint position, velocity, and effort. For point clouds, raw taxel readings are calibrated per taxel (unloaded baseline plus manually measured sensing range) into displacements and propagated through a MuJoCo model of the hand by forward kinematics. Packet-loss dropouts are recovered by neighbouring-taxel interpolation or baseline substitution.

Limitations

• A single object geometry (cube). Only stiffness and texture vary.
• 180 recordings is small for deep learning, even though the frame count is large.
• Data were collected offline under controlled conditions with a stationary hand mount.

Citation

@inproceedings{chen2026lostinencoding,
  title     = {Lost in Encoding: Comparing Property Retention in Tactile Representation Formats},
  author    = {Chen, Charlene and Yanc, Lara and Hamid, Mohammad and Cheng, Hanxue and He, Yuk Ching},
  booktitle = {TODO},
  year      = {2026}
}

Contact

University College London, Department of Computer Science. Supervised by Lorenzo Jamone.