TouchWorld: Giving Robots a Sense of Touch

量子位 · wechat · 2026-07-12

This article introduces the tactile approach of Yang Shuo's team at Harbin Institute of Technology (Shenzhen) and their startup PHANESAI: from collecting visual-tactile data via **EgoTouch**, to recovering touch from first-person videos using **TouchAnything**, to integrating tactile feedback into world models and manipulation policies with **TouchWorld**. ### What TouchWorld Does - Predicts not only visual changes but also **how contact occurs** and what the target contact state should be. - Introduces high-frequency tactile feedback during execution for real-time action correction, making it particularly effective for contact-intensive tasks like grasping, plugging, pressing, and wiping. ### Experimental Results - Tested on 6 real-world robot tasks, including watering plants, clearing tables, plugging in power cords, inserting cups, wiping pots, and pulling tissues. - Achieved an average success rate of **65.0%** in clean settings and **57.2%** under human disturbance scenarios. - Outperformed baseline methods by **15.7** and **16.0** percentage points in the two settings, respectively. ### Core Insights - Tactile sensing is not just "another sensor"; it is a critical infrastructure for embodied AI to enter the real world. - Vision and language alone are insufficient; robots need to know if they have made contact, grasped firmly, slipped, or applied appropriate force, and how to quickly correct themselves during contact. - The team is also advancing data collection, evaluation, and control loops for whole-body mobile dexterous manipulation in humanoid robots.

Related event: TouchWorld: A Tactile World Model for Robot Error Correction(2 posts)→

Original post →

More from Embodied

Embodied channel →