TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning
500小时人类视觉-触觉数据，到底能让机器人学到什么? | GAIR Paper 140
Researchers from Texas A&M, Google DeepMind, Carnegie Mellon University, Stanford University, NVIDIA, and others released TouchScale, a unified dataset comprising approximately 500 hours of human vision and touch data collected using consistent hardware. The study demonstrates that training models on this dataset improves cross-sensor contact prediction accuracy and increases robotic manipulation success rates by an average of 35 percentage points across four tasks, including sorting objects and removing bottle caps, without requiring human action labels or motion redirection.
Source: Chinese robotics — Dataset and collection discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
Reported duration
500 hours
Robot trajectories
≈87,000 trajectories
Trajectory count
Reported duration
16.2 hours
View original evidence
研究团队分别使用EgoTouch的16.2小时训练集和TouchScale约16小时子集训练相同的TouchAnything模型，再在独立的EgoTactile数据集上进行零样本评测。
Open source S17Reported duration
≈16 hours
View original evidence
研究团队分别使用EgoTouch的16.2小时训练集和TouchScale约16小时子集训练相同的TouchAnything模型，再在独立的EgoTactile数据集上进行零样本评测。
Open source S17Reported accuracy
24.45%
View original evidence
结果显示，TouchScale在三个基准、两种评估设置下均取得了对照组中的最高准确率，平均线性探测准确率达到24.45%，平均微调准确率达到49.62%。
Open source S20Reported accuracy
49.62%
View original evidence
结果显示，TouchScale在三个基准、两种评估设置下均取得了对照组中的最高准确率，平均线性探测准确率达到24.45%，平均微调准确率达到49.62%。
Open source S20
- paper: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Dataset licensing and commercial-use rights are not established here.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
研究团队构建了总时长500小时的人类视觉-触觉数据集，并从跨传感器触觉预测、视觉表征学习和真实机器人操作三个方向开展验证。
Open source S9
最终，TouchScale积累了约8.7万条轨迹、涉及约2000种任务描述和1500余件物体。
Open source S15
从跨传感器触觉预测，到机器人操作成功率提升35个百分点，TouchScale开始系统检验大规模人类触觉数据的学习价值。
Open source S2
研究团队分别使用EgoTouch的16.2小时训练集和TouchScale约16小时子集训练相同的TouchAnything模型，再在独立的EgoTactile数据集上进行零样本评测。
Open source S17
结果显示，TouchScale在三个基准、两种评估设置下均取得了对照组中的最高准确率，平均线性探测准确率达到24.45%，平均微调准确率达到49.62%。
Open source S20
作者丨TachinTech 编辑丨陈淑瑜 使用恰到好处的力道拿起一颗鸡蛋，对人类而言，是再简单不过的事情。
Open source S3
Implications for data suppliers
RoboSignal interpretation and collection questions, not statements of buyer demand.
- Confirm the required data type and collection setting with the buyer; this source does not establish a complete collection specification.
- Compare the reported units and scope before using these quantities in a budget. Recording hours, sensor-hours and trajectories are different measures.
- Validate demand and acceptance criteria with a buyer before scaling. Publication, popularity and a research result do not establish a purchase commitment.
Source：Chinese robotics — Dataset and collection discovery · msn.cn