Tsinghua University associate professor Ding Wenbo founded Xspark AI to develop tactile intelligence for robots, aiming to enhance reliable manipulation in industrial settings through multimodal tactile models. The company introduced a three-layer collaborative architecture and is building tactile-visual datasets to scale embodied AI solutions.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
As robots begin to truly 'touch' the physical world, how far are they from reliably entering factories, homes, and daily human life? Xspark AI founder Ding Wenbo believes that touch is the final link in physical interaction and the key to robots moving from 'sensing' to 'knowing.'
When he graduated with his PhD in 2016, Ding Wenbo had postdoctoral opportunities at MIT and Georgia Tech. The former continued communications research, while the latter pointed to flexible electronics and electronic skin. He observed that communications research was shifting from academia to industry, while flexible electronics still had many fundamental problems. He ultimately chose Georgia Tech, starting from scratch under Academician Wang Zhonglin.
From 2018 to 2019, Ding Wenbo truly began to focus on machine touch as a long-term research direction. At that time, there were perhaps only 5-10 teams in China working on machine touch long-term, and many fundamental problems remained unsolved. His judgment was that for robots to enter the real world to complete grasping, manipulation, and assembly, vision alone was insufficient; information such as force, slip, material, and temperature after contact ultimately required touch.
Early research focused on sensors. The team developed the multimodal tactile sensor SuperTac, which integrates multispectral imaging and triboelectric sensing to simultaneously perceive force, temperature, material, texture, slip, proximity, and more. But by 2022-2023, he realized that simply making sensors stronger was not enough; the problem shifted to data and models.
Large-scale tactile data collection is more difficult than vision. First is the sensor consistency problem: even with the same design, different devices vary, and wear and drift over time require continuous calibration. Second, modality definitions are not unified; there is no consensus on the scope and key variables of touch. Third, representation: there is no standard answer for how tactile signals enter models, and conversion processes may lose important information.
Cost is also a practical issue: visual data can be passively collected, but tactile data must be generated through real contact. Each data point is a physical interaction involving hardware, calibration, and wear. Different sensors have different principles and output formats, making data less naturally interoperable than images. Simulation and generated data can supplement scale, but touch involves complex contact physics such as friction and deformation, making sim-to-real significantly harder than vision, and real data cannot be fully replaced.
HumanTouch is an engineering attempt: continuously calibrating devices, judging real contact using hand pose and time history, and leveraging natural human manipulation to synchronously record tactile, action, and visual data. The dataset has accumulated over 1,000 hours covering more than 100 tasks, with the first public subset of about 100 hours. Ding Wenbo firmly believes there is a scaling law for touch, but simply increasing data volume may not be key; a more realistic approach is to expand the scale of high-quality data and study generalization.
This article is based on DeepTech's interview with Ding Wenbo. All facts are from the interviewee's own statements and have not been independently verified by third parties. Descriptions of company products, data, and architecture are at the source_claim level.
Primary report
Primary source