TouchScale: 500 Hours of Human Vision and Touch for Visual-Tactile Learning
TouchScale pairs 500 hours of egocentric video with full-hand tactile signals from one consistent wearable setup.
The dataset covers about 2,000 predefined tasks across everyday activity and structured manipulation. In the paper’s tests, training on the full set raised zero-shot contact IoU on an unseen tactile sensor from 0.134 to 0.383. Used for robot mid-training, it lifted average real-world success across four contact-heavy manipulation tasks from 22.5% to 57.5%. The authors say the full synchronized recordings and reconstructed object models will be released. HF Daily Papers' note
The dataset covers about 2,000 predefined tasks across everyday activity and structured manipulation. In the paper’s tests, training on the full set raised zero-shot contact IoU on an unseen tactile sensor from 0.134 to 0.383. Used for robot mid-training, it lifted average real-world success across four contact-heavy manipulation tasks from 22.5% to 57.5%. The authors say the full synchronized recordings and reconstructed object models will be released. HF Daily Papers' note
score 6