Visuotactile Sensing Is Hard to Copy, Chinese Robot Firms Say as Funding and Products Advance

In 3C production lines, for example, inserting a USB cable into a millimeter-scale port requires knowing whether the plug is aligned, whether it is stuck and how much force to apply. Vision answers what is seen; touch answers how contact feels. As vision-language-action models hit performance ceilings in real tasks, more embodied-intelligence teams have concluded that the absence of tactile information is a core limitation.

Visuotactile sensing is a leading branch of tactile technology. High-end dexterous hands often adopt it as standard. Capital has followed. Daimon Robotics completed a 100-million-yuan Series A and a strategic round in the hundreds of millions of yuan led by Ant Group within two months; Yimu Technology raised more than 1 billion yuan in Series E in July, with a post-money valuation above 10 billion yuan; Qianjue and Weitu also raised 100-million-yuan rounds this year, according to Leiphone.

Doubts include claims that visuotactile sensors are just cameras and silicone, that optical structures cannot be made thin or fast, that visual data create computing, storage, wiring and power burdens, and that soft surfaces cannot survive long industrial use. Leiphone reported that some of these concerns are outdated or conflate surface-material engineering with fundamental flaws. Cost, thickness, consistency, durability and integration of tactile data into model training remain real hurdles, the report said.

The operating principle is straightforward. A transparent elastomer covers a cavity containing light sources and a miniature camera. When an object presses the gel, the deformation creates patterns of light and shadow that the camera captures; algorithms reconstruct contact geometry, force distribution and sliding. The core trade is turning touch into images. The route originated at MIT. In 2009, Edward Adelson's team at MIT CSAIL proposed GelSight at CVPR for microscopic surface observation. In 2014, the team produced the first ultra-high-resolution fingertip GelSight sensor. Its first author, Li Rui, later founded Weitu Robotics. In 2020, Meta open-sourced the full hardware design of DIGIT, lowering research barriers; in 2021 Meta partnered with GelSight for commercialization, and optical tactile sensing spread through global labs, including variants such as OmniTact.

Critics say visuotactile sensing lacks its own algorithms. Leiphone reported that although photometric stereo comes from computer vision, GelSight combined it with elastomer reflective coatings and controllable multi-angle illumination to create a high-resolution geometric reconstruction paradigm. Later work on gel nonlinearity, shear-force decoupling and marker tracking was tailored to tactile physics rather than simply applying vision tools. On contamination, the optical cavity is closed and the camera looks inward, so oil or dirt does not enter or blur the image like a dirty camera lens. Contamination mainly affects the gel's mechanical and optical coupling, reducing local deformation, accuracy and life, a surface-materials issue shared with resistive and capacitive soft tactile approaches.

The value proposition is density. Traditional tactile sensors use discrete sensing units, so adding more points requires more physical devices. Visuotactile density is set by camera resolution. A 640-by-480 image has 300,000 pixels, each corresponding to a tactile point; on a 3-square-centimeter sensor, that can mean tens of thousands of points per square centimeter, Li Rui of Weitu Robotics said. Yimu Technology founder Li Zhiqiang compared the human fingertip's roughly 12,000 tactile points with Yimu's claimed 240,000. Traditional piezoresistive sensors offer only dozens of points per square centimeter, while capacitive and Hall-effect approaches are around 100. Visuotactile sensors can measure normal force, tangential force and sliding, handle flexible objects such as clothing, cables and food, and resist temperature, humidity and electromagnetic interference, Li Rui said. Several companies called it the most human-like and possibly the ultimate tactile route.

Players include GelSight abroad and Daimon Robotics, Yimu Technology, Weitu Robotics, Qianjue Robotics, Moliang Technology, Diedong Technology and Zhixing Embodied in China. Dexterous-hand makers such as Sharpa, Xinuo Future and Shiyue Technology are also developing their own. Backgrounds vary: GelSight and Weitu come from MIT; Daimon was incubated by Hong Kong University of Science and Technology's Robotics Institute; Qianjue founder Ma Daolin is an associate professor at Shanghai Jiao Tong University; Diedong's core team also comes from HKUST; Yimu moved from spectral chips; Moliang's founding team comes from City University of Hong Kong, Xiamen University and other schools.

The companies share common traits. Most are becoming system-level firms, extending from sensors into data collection, algorithm platforms, datasets and end effectors. Daimon, Yimu, Qianjue and Weitu have built data-collection devices, datasets and models; Moliang released a data-collection glove; even the youngest, Diedong, launched a gripper with integrated computing. Qianjue co-founder and CTO Zhao Haonan said sensors, data and models are interlinked, and interface areas are opportunities before better hardware or models arrive. Yimu's Li called it a transitional phase of first going full-stack and later narrowing. A Daimon executive said full-stack is not the goal; the company will cooperate where suitable but must own the tactile data-processing chain and tactile model.

Industrial capital has entered densely. Daimon's shareholders include Inovance, China Mobile, China Telecom, Ant Group, Lenovo and China Merchants Venture. Yimu is backed by Xiaomi-linked Shunwei Capital, TCL and Bosch's Boyuan Capital. Weitu's angel round was led by Xiaomi Strategic Investment, with Wethink Chip in Pre-A. Qianjue's angel round was led by AgiBot, followed by Li Auto and Jide Electric. Home-appliance and consumer-electronics investors are most visible. Daimon works on USB insertion and label application in 3C lines; Weitu works with Xiaomi and several Fortune 500 companies; Qianjue has served more than 300 leading customers, mass-produced gripper products last year, with orders in the thousands to tens of thousands of units.

Water and dust resistance and frame rates have advanced. Daimon's sensors are IP67-rated and tested for 5 million presses and 200,000 friction cycles. Weitu's gripper sensor is IP54 and upgradeable, operating from minus 25 to 80 degrees Celsius. Yimu's Sentra T0 series is IP65, with an industrial life of more than 5 million presses and no fear of temperature drift, according to the company. Qianjue is developing a waterproof sensor. The old claim that visuotactile frame rates are only 30 to 60 fps is outdated, Leiphone reported: Daimon's gripper version runs at 120 Hz and its dexterous-hand version at 180 Hz; Qianjue's sampling rate reaches 150 Hz; Moliang's industrial gripper reaches 120 Hz; Diedong's full line can exceed 400 Hz.

Product forms are also diversifying, including curved surfaces. Yimu offers square, wedge and fingertip shapes. Qianjue's Jingling is a bionic fingertip, Quark N1 and W1 are bionic curved surfaces, and Photon is a wedge-shaped flat surface. Homogenization claims do not hold up, Leiphone reported. Open-source designs lower entry barriers, but commercial products demand high-level gel formulations, optical calibration, batch consistency, replaceable designs and software stacks. Evaluation criteria include spatial resolution and depth accuracy, force sensing including normal, tangential and torque, frame rate and latency, size and integration, and durability and environmental adaptability. Parameter labeling is not standardized, making direct comparison difficult, but differences in emphasis reveal strategies.

Daimon emphasizes durability and consistency, with IP67 protection, resistance to water, dust and electromagnetic interference, detailed sensitivity, zero drift, output time drift and nonlinearity, and surfaces that withstand scratches, punctures and scrapes. This stems from multilayer material processes. Yimu emphasizes thinness, using a self-developed metasurface silicon-photonic photosensitive chip to compress the optical path to under 10 millimeters, with a second generation at 3 millimeters. Yimu said many core team members have Huawei backgrounds and strong engineering capability; it has won orders from several leading customers this year and has begun mass production of a bionic fingertip tactile sensor for dexterous hands. Qianjue stresses consistency and precision, with self-calibration algorithms across its line, cross-sensor consistency above 97 percent in whole-hand tactile systems, and detailed force and deformation-field accuracy. Weitu makes scene-specific trade-offs: its gripper version has a lower sampling rate but a wide operating temperature range, while its fingertip version has a higher frequency for 0 to 50 degrees Celsius; the GF225 has normal and tangential force ranges of 30 N, the highest among publicly labeled products cited in the report. Moliang combines visuotactile and piezoresistive sensing, using visuotactile for high-precision fingertips and piezoresistive for large-area palm coverage. Zhixing Embodied switched to event cameras, which only detect brightness changes and output when change occurs, avoiding frame-by-frame snapshots, reducing repeated image data, requiring little bandwidth and no external computing device. Diedong introduced MEMS processes and pushed full-line sampling above 400 Hz.

On ecosystem, Daimon has the most complete layout, from sensors to data-collection devices, datasets, evaluation benchmarks, VTLA models and world models. In 2024 it proposed the VTLA architecture, or vision-tactile-language-action, and others followed. It recently released Daimon-TWM, a tactile-anchored world model. Four companies have public datasets, the report said; Daimon's Daimon-Infinity is described as the world's largest tactile dataset and it open-sourced 10,000 hours. Qianjue's TacVerse 1k was also cited among the public datasets.