The First Cross-Sensor Unified Tactile Annotation Benchmark
Every public tactile dataset today is sensor-specific — GelSight data only has GelSight format, DIGIT only DIGIT, DMA only DMA. No one has ever annotated the same set of objects across different sensors in a unified format that enables direct comparison.
TLabel-Bench fills this gap. It is the first benchmark that provides:
- Cross-sensor annotations for the same objects across GelSight Mini, DIGIT, DMA, and Xense
- Unified TLabel format — switch sensors by switching adapters, labels stay identical
- Multi-level annotations — material labels, episode segmentation (contact → press → release), and quality scores
- Three export formats — HDF5, CSV, and JSON for maximum compatibility
tlabel-bench/
├── annotations/ ← TLabel annotation files (JSON/CSV)
│ ├── gelsight/ ← GelSight Mini annotations
│ ├── digit/ ← DIGIT annotations
│ ├── dma/ ← DMA (Daimon) annotations
│ └── xense/ ← Xense annotations
├── scripts/ ← Auto-download & merge scripts
├── evaluation/ ← Benchmark evaluation scripts
├── docs/ ← Documentation
└── examples/ ← Quickstart notebooks
Note: This repository only contains annotation files and scripts. Raw sensor data must be downloaded separately using the provided scripts (see Data Sources).
pip install tlabel>=0.4.2bash scripts/download_data.shpython scripts/merge_annotations.pypython evaluation/material_classification.pyEach annotation file follows the TLabel standard schema:
| Field | Type | Description |
|---|---|---|
sensor_type |
string | Sensor adapter name (e.g., "gelsight_mini") |
object_id |
string | Unique object identifier |
material_label |
string | Material category (rigid, soft, deformable) |
episode |
object | Episode segmentation: contact, press, release |
quality_score |
float | Data quality score (0.0–1.0) |
timestamp |
float | Relative timestamp in seconds |
See annotation_schema.md for the full specification.
| Task | Metric | Description |
|---|---|---|
| Material Classification | Accuracy / F1 | Classify objects by material using TLabel features |
| Episode Segmentation | IoU / F1 | Detect contact→press→release boundaries |
| Cross-Sensor Transfer | Accuracy drop | Train on sensor A, test on sensor B |
| Quality Assessment | Correlation | Predict human quality ratings |
This repository contains only annotation files (labels and metadata). Raw tactile data must be downloaded from their original sources. See data_sources.md for:
- Download links for each dataset
- License information
- Citation requirements
Why only annotations? Many tactile datasets do not have permissive licenses for redistribution. Our annotation files are original works under CC-BY-4.0, while raw data remains subject to their respective licenses.
- TLabel — Sensor-agnostic tactile annotation toolkit (the tool that generates these annotations)
- Touch and Go — GelSight tactile-visual dataset
- SSVTP — DIGIT-based tactile pose estimation
- ObjTac — Multi-object tactile dataset
If you use TLabel-Bench in your research, please cite:
@dataset{tlabel_bench_2026,
title={TLabel-Bench: The First Cross-Sensor Unified Tactile Annotation Benchmark},
author={Wu, Sheng and Luo, Xi},
year={2026},
publisher={GitHub},
url={https://github.com/liesliy/tlabel-bench}
}We welcome contributions! See contribute.md for guidelines.
Areas we especially need help with:
- New sensor adapters (SynTouch, XELA, etc.)
- Additional objects and annotations
- Evaluation metrics and baselines
Annotation files are licensed under CC-BY-4.0. Raw data is subject to their respective original licenses.
Built with 🦞 by NiuXu Technology