TLabel

A Unified Annotation Framework for Cross-Sensor Tactile Manipulation Data

TLabel is the first cross-sensor tactile annotation schema with capability declarations and Compliance Level stratification. It enables heterogeneous tactile sensors — regardless of operating principle — to produce compatible 14-dimensional semantic annotations while preserving their unique strengths.

TL;DR — The Unicode for tactile data: one standard schema, every sensor.

Why TLabel?

Tactile datasets today ship as raw sensor signals without semantic annotations. Each sensor type demands its own ad-hoc processing, and results from different sensors cannot be compared or fused. TLabel addresses this by:

• Standardizing annotations — 14 dimensions covering spatial, mechanical, surface, dynamic, and meta perceptions
• Declaring capabilities — each adapter explicitly states which dimensions it can and cannot annotate
• Stratifying compliance — Compliance Level (L1–L4) ensures every sensor participates at its appropriate information density
• Enabling cross-sensor comparison through a shared output format

Quick Start

Install

pip install tlabel

Load and explore data

import tlabel

# Load tactile data (auto-detects sensor format)
data = tlabel.load("path/to/sensor_data.pkl")

# Or try the built-in demo — no files needed
data = tlabel.demo("gelsight")

# Inspect annotation metadata
print(data.describe())

Interactive annotation (Jupyter)

data.review()  # Bilingual annotation panel (Chinese / English)

Export to training formats

data.export("output.json")              # JSON / CSV
data.export_ftp1("output.zarr")         # FTP-1 Zarr for foundation models

from tlabel.converters import tlabel_to_lerobot
tlabel_to_lerobot("annotations.json", "lerobot_episode/")  # LeRobot

CLI

tlabel list                       # List all registered adapters
tlabel info gelsight              # Adapter details & compliance level
tlabel validate data.json         # Schema compliance check

Optional dependencies

pip install tlabel[gelsight]      # GelSight / DIGIT (.pkl)
pip install tlabel[paxini]        # PaXini PXCap (.h5)
pip install tlabel[daimon]        # Daimon DM-TacClaw (.parquet)
pip install tlabel[ftp1]          # FTP-1 export (zarr)
pip install tlabel[all]           # Everything

Schema — 14 Dimensions, 4 Compliance Levels

TLabel defines 14 semantic dimensions with Compliance Levels (L1–L4) indicating annotation completeness:

The 14 dimensions span: contact, contact_centroid, force_magnitude, slip_event, confidence, compliance_level, contact_region, force_vector, torque_vector, slip_velocity, manipulation_phase, texture_class, object_deformation, temperature.

Full dimension spec → docs/tlabel-format.md

Supported Sensors

Dataset Adapters (offline): GelSight/DIGIT (L3) · Daimon DM-TacClaw (L3) · PaXini PXCap (L2) · UniVTAC (L3, legacy *_gsmini and new *_tactile HDF5 layouts) · TacQuad/AnyTouch (L1) · VTouch (L2) · ToucHD-Force/AnyTouch 2 (L3) · YCB-Slide (L2) · SynTouch BioTac (L2) · Tashan TS-F-A (L3) · XELA uSkin/UniTac-NV (L1)

Real-time Adapters (hardware): PaXini GEN3 (L2) · Daimon DM-Tac (L2)

Adding a new sensor takes ~30 min — fork contrib/adapter-template/

Architecture

┌─────────────────────────────────────────────────┐
│  Layer 1: Schema                                │
│  14 semantic dimensions + Compliance Level L1-L4│
├─────────────────────────────────────────────────┤
│  Layer 2: Adapters                              │
│  DataAdapterBase │ SensorAdapterBase             │
│  (7 built-in + community-extensible)            │
├─────────────────────────────────────────────────┤
│  Layer 3: Downstream                            │
│  Feature derivation · Augmentation · Export      │
│  PredictEngine · FTP-1 · LeRobot · RLDS · ROS2 │
└─────────────────────────────────────────────────┘

LeRobot Integration

TLabel provides bidirectional converters for LeRobot format, enabling seamless integration with the LeRobot ecosystem.

LeRobot → TLabel

Convert LeRobot parquet data to TLabel for annotation and quality assessment:

from tlabel.converters import lerobot_to_tlabel

# Load LeRobot episode and convert to TLabel
data = lerobot_to_tlabel("path/to/lerobot_episode/")

# Review and annotate
data.review()  # Interactive bilingual annotation panel

# Export annotated data
data.export("annotated_tlabel.json")

TLabel → LeRobot

Convert TLabel annotations back to LeRobot format for training:

from tlabel.converters import tlabel_to_lerobot

# Convert TLabel annotations to LeRobot parquet format
tlabel_to_lerobot(
    "annotated_tlabel.json",
    "path/to/lerobot_episode/",
    tactile_field="observation.tactile",
)

Complete Example

import tlabel
from tlabel.converters import lerobot_to_tlabel, tlabel_to_lerobot

# 1. Load LeRobot data
lerobot_data = lerobot_to_tlabel("my_robot_episode/")

# 2. Review and annotate (optional)
lerobot_data.review()

# 3. Export to TLabel format for inspection
lerobot_data.export("tlabel_annotations.json")

# 4. Convert back to LeRobot format (with annotations)
tlabel_to_lerobot("tlabel_annotations.json", "my_robot_episode/")

Requirements

pip install tlabel[lerobot]  # or: pip install pyarrow

Paper

TLabel: A Unified Annotation Framework for Cross-Sensor Tactile Manipulation Data

Xi Luo, Sheng Wu (Niuxiu Tech)

Submitted to SoftwareX, 2026. Manuscript: SOFTX-S-26-01665

[PDF] · LaTeX source: paper/

Citation

@software{tlabel2026,
  title  = {TLabel: A Sensor-Agnostic Tactile Data Annotation Toolkit and Format Standard},
  author = {Wu, Sheng and Luo, Xi},
  year   = {2026},
  url    = {https://github.com/liesliy/tlabel}
}

Documentation

Contributing

TLabel is designed to be extensible. Add your sensor in ~30 minutes:

1. Fork contrib/adapter-template/
2. Subclass DataAdapterBase or SensorAdapterBase
3. Submit a PR or publish as a standalone package

See CONTRIBUTING.md for details.

License

MIT © 2026 Niuxiu Tech

TouchLabel AI — Tactile Data Annotation Infrastructure
GitHub · PyPI · Discord
Niuxiu Tech · Hangzhou, China