Voxel51/TaF-Tactile-Force

TaF-Dataset → FiftyOne (Native Multimodal MCAP)

jiamig/taf-dataset converted to native multimodal MCAP episodes.

Every frame pairs what a tactile sensor sees with what a force sensor measures. A vision-based tactile sensor records the gel deforming, a 12x12 piezoelectric array records where the contact lands, and an ATI six-axis sensor records the wrench that caused it, all on one 30 fps clock. Contacts reach 90.8 N and 1.58 Nm.

3,594 episodes across 408 sequences, 10,057,742 frames, 93.1 hours.

Installation

pip install fiftyone

Usage

import fiftyone as fo
import fiftyone.utils.huggingface as fouh

dataset = fouh.load_from_hub(
    "Voxel51/TaF-Tactile-Force",
    name="TaF-Tactile-Force",
    persistent=True,
)
fo.launch_app(dataset)

The firmest contacts:

view = dataset.match({"peak_force": {"$gt": 50}})

What you get

Each episode carries:

• /tactile-camera, the tactile image, as foxglove.CompressedVideo
• /pressure-map, the 12x12 piezoelectric array as a heatmap, as foxglove.CompressedImage
• /force-torque, the six-axis wrench, with /force-torque.plot carrying fx, fy, fz, tx, ty and tz as a timeline
• /pressure.plot carrying the summed pressure, the peak cell, and how many cells are in contact
• /instruction, naming the sensor configuration

Episodes carry the fields sequence, sensor_family, episode_index, num_frames, duration, fps, peak_force, peak_fz, peak_torque and peak_pressure, so a view can select on how hard the contact got.

Six sensor configurations are represented:

Configuration	Episodes
GelSight Mini with markers	1,103
GelSight Mini without markers	810
Custom sensor without markers	657
Custom sensor, 6*6 markers	650
Custom sensor, 8*8 markers	285
Custom sensor, 7*7 markers	89

Peak force runs to a median of 25.0 N and a maximum of 90.8 N; peak pressure to a median of 884 and a maximum of 4052.

Notes on the conversion

The tactile video is re-encoded to Annex-B H.264 without B-frames. The source is AV1, which the viewer does not decode.

The pressure array is 12x12, written as an image and upscaled with nearest-neighbour so each cell stays a square. Colour is scaled per episode against that episode's 99.5th percentile, so a light touch and a hard press are both readable. The /pressure.plot channel carries the unscaled values.

Three of the six sensor configurations spell their marker grid with an asterisk, as in custom_designed_6*6_mark_obj1. That character cannot appear in a Windows path, so directories on disk carry 6x6, 7x7 and 8x8 while the sequence field keeps the name as the release writes it.

Tactile image resolution varies by sequence.

This conversion holds 3,594 episodes and 10,057,742 frames against the 3,592 and 10,053,265 the source card states, across the same 408 sequences.

License & attribution

TaF-Dataset is released under the MIT license, and this conversion is distributed under the same license.

Changes from the source: conversion to the FiftyOne MCAP flavor, re-encoding of the tactile video to H.264, and rendering of the pressure array as an image stream alongside its scalar channels.

The dataset accompanies TaF-VLA: Tactile-Force Alignment in Vision-Language-Action Models for Force-aware Manipulation.