taf-dataset
收藏资源简介:
TaF-Dataset是一个大规模、同步的多模态触觉-力数据集,旨在用于学习力感知的触觉表征和研究接触密集的机器人操作。数据采集时,使用ATI六轴力/力矩传感器记录交互力/力矩(Fx, Fy, Fz, Tx, Ty, Tz),压电传感阵列提供识别接触区域的空间压力分布图,视觉触觉传感器记录相应的触觉图像。这些信号在时间上同步,使得模型能够将局部触觉形变和接触位置与物理交互力关联起来。数据集规模庞大,包含10,053,265个同步帧、3,592个episode和408个物体/传感器序列。数据涵盖六种不同的触觉传感器配置:带标记的GelSight Mini、无标记的GelSight Mini、无标记的自定义设计传感器、带6x6标记的自定义设计传感器、带7x7标记的自定义设计传感器以及带8x8标记的自定义设计传感器,分别对应124、84、84、74、10和32个记录序列。每个同步数据帧包含以下模态:分辨率为240x320x3的RGB触觉图像(observation.image)、表示接触区域和空间压力分布的12x12压电压力矩阵(observation.pressure_matrix)、按Fx, Fy, Fz, Tx, Ty, Tz顺序排列的6维ATI力/力矩测量值(observation.force_torque),以及时间戳、帧索引和episode索引。其中法向力通道Fz最能反映主接触方向的负载,完整的六维力/力矩数据则用于捕捉切向力和旋转交互动力学。数据集遵循LeRobot v3布局进行组织,主要分为`gs_mini/`和`custom_designed/`两大目录,分别包含208和200个记录序列。每个序列目录使用统一的内部结构,包含数据块、视频和元数据文件。序列目录的命名遵循特定模式,并附有详细的ID范围说明。该数据集适用于机器人操作、特征提取等任务,特别是需要结合触觉感知与力反馈的研究与应用。
The TaF-Dataset is a large-scale, synchronized multimodal tactile-force dataset designed for learning force-aware tactile representations and studying contact-intensive robotic manipulation. It was proposed in the paper "TaF-VLA: Tactile-Force Alignment in Vision-Language-Action Models for Force-aware Manipulation." During data collection, an ATI six-axis force/torque sensor records interaction forces/torques (Fx, Fy, Fz, Tx, Ty, Tz), a piezoelectric sensing array provides spatial pressure maps identifying contact areas, and a vision-based tactile sensor records corresponding tactile images. These signals are synchronized in time, enabling models to associate local tactile deformations and contact locations with physical interaction forces. The dataset is extensive, containing 10,053,265 synchronized frames, 3,592 episodes, and 408 object/sensor sequences. It covers six different tactile sensor configurations: marked GelSight Mini, unmarked GelSight Mini, unmarked custom-designed sensor, custom-designed sensor with 6x6 markers, custom-designed sensor with 7x7 markers, and custom-designed sensor with 8x8 markers, corresponding to 124, 84, 84, 74, 10, and 32 recorded sequences, respectively. Each synchronized data frame includes the following modalities: an RGB tactile image with a resolution of 240x320x3 (observation.image), a 12x12 piezoelectric pressure matrix representing contact areas and spatial pressure distribution (observation.pressure_matrix), a 6-dimensional ATI force/torque measurement in the order Fx, Fy, Fz, Tx, Ty, Tz (observation.force_torque), as well as timestamps, frame indices, and episode indices. Among these, the normal force channel Fz best reflects the load in the main contact direction, while the full six-dimensional force/torque data captures tangential forces and rotational interaction dynamics. The dataset is organized following the LeRobot v3 layout, primarily divided into two directories: `gs_mini/` and `custom_designed/`, containing 208 and 200 recorded sequences, respectively. Each sequence directory uses a unified internal structure, including data chunks, videos, and metadata files. Sequence directory names follow a specific pattern, accompanied by detailed ID range descriptions. This dataset is suitable for tasks such as robotic manipulation and feature extraction, particularly in research and applications requiring the integration of tactile perception and force feedback.
TaF-Dataset 数据集详情
基本信息
- 许可证: MIT
- 标签: 机器人学、触觉感知、力/力矩、多模态
- 任务类型: 机器人学、特征提取、其他
- 来源论文: TaF-VLA: Tactile-Force Alignment in Vision-Language-Action Models for Force-aware Manipulation
数据集概述
TaF-Dataset 是一个大规模、同步的触觉-力数据集,专为力感知触觉表示学习和接触丰富的机器人操作研究而设计。数据采集过程中使用了 ATI 六轴力/力矩传感器记录交互力矩,压电传感阵列提供空间压力分布图,视觉触觉传感器记录相应触觉图像,所有信号在时间上同步。
规模统计
- 同步帧数: 10,053,265 帧
- 片段数: 3,592 个
- 物体/传感器序列数: 408 个
触觉传感器配置
数据集包含六种触觉传感器配置:
| 配置 | 父目录 | 序列数 |
|---|---|---|
| GelSight Mini(带标记) | taf_dataset/gs_mini/ |
124 |
| GelSight Mini(无标记) | taf_dataset/gs_mini/ |
84 |
| 定制传感器(无标记) | taf_dataset/custom_designed/ |
84 |
| 定制传感器(6×6 标记) | taf_dataset/custom_designed/ |
74 |
| 定制传感器(7×7 标记) | taf_dataset/custom_designed/ |
10 |
| 定制传感器(8×8 标记) | taf_dataset/custom_designed/ |
32 |
数据模态
每一帧包含以下同步观测数据:
| 字段 | 形状 | 描述 |
|---|---|---|
observation.image |
240×320×3 | RGB 触觉图像 |
observation.pressure_matrix |
12×12 | 压电压力分布图 |
observation.force_torque |
6 | ATI 力/力矩测量值,顺序为 Fx, Fy, Fz, Tx, Ty, Tz |
timestamp |
1 | 帧时间戳(秒) |
frame_index |
1 | 片段内帧索引 |
episode_index |
1 | 片段标识符 |
法向力通道 Fz 是主要接触方向上的载荷指标,而完整的六维力矩保留了切向力和旋转交互动态信息。
数据组织结构
数据集遵循 LeRobot v3 格式,仓库结构如下:
TaF-Dataset/ ├── README.md └── taf_dataset/ ├── gs_mini/ (含 208 个序列) └── custom_designed/ (含 200 个序列)
每个序列采用相同的 LeRobot 布局:
<sensor_configuration>_obj<ID>/ ├── data/chunk-000/file-000.parquet ├── videos/observation.image/chunk-000/file-000.mp4 └── meta/ ├── episodes/chunk-000/file-000.parquet ├── collection_config.json ├── info.json ├── stats.json └── tasks.parquet
ID 范围说明
各传感器配置使用独立的数字 ID 编号:
| 相对路径模式 | ID 范围 |
|---|---|
taf_dataset/gs_mini/gs_mini_obj<ID> |
1–124 |
taf_dataset/gs_mini/gs_mini_wo_marker_obj<ID> |
1–84 |
taf_dataset/custom_designed/custom_designed_no_mark_obj<ID> |
1–84 |
taf_dataset/custom_designed/custom_designed_6*6_mark_obj<ID> |
1–74 |
taf_dataset/custom_designed/custom_designed_7*7_mark_obj<ID> |
1–10 |
taf_dataset/custom_designed/custom_designed_8*8_mark_obj<ID> |
1–32 |
注意: 数字 ID 仅在各传感器配置内局部有效,不同前缀但相同 ID 的目录可能代表不同的序列,不能视为包含相同物理对象。
引用信息
bibtex @article{huang2026tafvla, title = {TaF-VLA: Tactile-Force Alignment in Vision-Language-Action Models for Force-aware Manipulation}, author = {Huang, Yuzhe and Lin, Pei and Li, Wanlin and Li, Daohan and Li, Jiajun and Jiang, Jiaming and Xiao, Chenxi and Jiao, Ziyuan}, journal = {arXiv preprint arXiv:2601.20321}, year = {2026} }




