Multi-Sensor Tactile Dataset

TacVerse

A Multi-Sensor Dataset and Benchmark for Cross-Sensor Vision-Based Tactile Perception

Seven sensors. Three tasks. One controlled benchmark for sensor shift.

TacVerse provides 106,800 tactile images and standardised protocols for within-sensor learning, zero-shot cross-sensor transfer, and few-shot target adaptation.

Paper Dataset Code

Lan Wei · Gurmeher Khurana · Sirine Bhouri · Wenhao Hong · Zeyuan Xin · Qingzheng Cong · Wen Fan · Yanzheng Xiang · Dandan Zhang

Imperial College London · Queen Mary University of London · King's College London

Seven vision-based tactile sensors and corresponding tactile images.

106,800 Tactile Images

7 VBTS Platforms

3 Perception Tasks

3 Evaluation Protocols

Sensor Shift

Why TacVerse?

Vision-based tactile sensors encode contact through different optical, mechanical, and marker configurations. As a result, a model that performs well on one sensor may not transfer reliably to another. TacVerse provides task-aligned data and controlled protocols to make this sensor-shift gap measurable and reproducible.

TacVerse tactile sensors and tactile observations.

GelSightNoMarker GelSightMarker MagicGripper MagicTac TacTip ViTac ViTacTip

Covering IMM, MDM, MFM, and hybrid sensing configurations.

Dataset & Tasks

Three Complementary Tactile Tasks

Shape Classification

9 classes · 30,094 images

Recognising local contact geometry across different tactile sensors.

Grating Classification

30 classes · 40,509 images

Distinguishing fine-grained line and dot gratings with different spatial scales.

Force Regression

36,197 images · Fx, Fy, Fz

Estimating normal and shear forces directly from tactile observations.

Evaluation Protocols

A Controlled Cross-Sensor Benchmark

Within-Sensor

Train and evaluate on the same tactile sensor.

Zero-Shot Cross-Sensor

Train on a source sensor and evaluate directly on a target sensor without target labels.

Few-Shot Adaptation

Fine-tune the source-trained model using a small labelled subset from the target-sensor training split. Evaluated for force regression.

Results Summary

Key Findings

Cross-sensor transfer heatmap for TacVerse. — Cross-sensor transfer pattern

Few-shot adaptation curve for TacVerse. — Few-shot adaptation trend

Strong within-sensor performance does not guarantee cross-sensor transfer.
Shape classification is comparatively robust to sensor shift.
Grating classification and force regression are more sensor-dependent.
Few-shot adaptation narrows the transfer gap, while tactile MAE pretraining provides the most consistent gains across tasks.

MAE pretraining delivers the most consistent representation across the evaluated tasks and sensors.

Resources & Citation

Resources

Dataset

Tactile images, annotations, metadata, and fixed data splits.

View Dataset

Code

Training, evaluation, and experiment configuration files.

View Code

Paper

Full benchmark design, experimental settings, and results.

Read Paper

Citation

If you find TacVerse useful, please cite:

[update later]