On October 11, LDROBOT (1236.HK), a company listed on the Hong Kong Stock Exchange, issued an announcement declaring the signing of a cooperative agreement on flexible 3D tactile perception for embodied robots with Huazhong University of Science and Technology. The two parties will focus on the tactile needs of robots to develop flexible 3D tactile sensors and high-density sensor arrays, develop supporting signal acquisition hardware, 3D force decoupling and tactile imaging algorithms, and verify the application capabilities of this system on carriers such as dexterous hands and robotic arms. The cooperation relies on the National Key Laboratory of Intelligent Manufacturing Equipment and Technology at the School of Mechanical Science and Engineering, Huazhong University of Science and Technology, and is led by Professor Li Kan.
Founded in Shenzhen in 2017, LDROBOT focuses on the "eyes" of robots—DTOF LiDAR and visual perception modules. In 2025, the number of intelligent robots equipped with its visual perception technology exceeded 9 million units, and the shipment volume of DTOF LiDAR surpassed 4 million units, ranking first globally. Visual perception products contributed approximately CNY 606 million in revenue, accounting for 81.1% of total revenue. In May this year, LDROBOT was listed on the Hong Kong Stock Exchange.
A company that makes "eyes" is now venturing into "skin." While it may seem like a mere business extension, it hits a larger industry node: robots are transitioning from "being able to see" to "being able to touch."
1. With Eyes Only, Robots Cannot Perform Fine Tasks
Over the past two years, VLA (Vision-Language-Action) models have equipped robots with the ability to understand environments and generate actions—robots see the world through cameras, comprehend human instructions, and then generate actions. However, VLA has a fundamental flaw: vision can tell the robot "where the target is," but it cannot tell it "how much force to apply."
Robots without tactile senses "look like they can work," but when touching physical objects, they "don't know the weight or force." They cannot tell whether the object in their hand is about to slip or has already been crushed.
This is precisely why VTLA (Vision-Tactile-Language-Action) models are beginning to gain industry attention. It adds touch to vision, language, and action. Tactile perception is not auxiliary information for vision, but a perception and control channel that requires an independent learning path. Fei-Fei Li's team conducted an ablation study in the T-Rex research: simply splicing tactile signals as extra inputs into existing VLA models caused the task success rate to plummet from 17% to 6%; however, after designing an independent model architecture and learning path for touch, the success rate increased to 65%. This indicates that touch cannot be treated as a vassal of vision; it needs its own encoder and representation space.
The VTLA embodied tactile model exhibited by Xense Robotics at WAIC 2026 has enabled robots to perform adaptive operations without preset trajectories in flexible material processing. When cardboard is displaced or deformed by external forces, the robot perceives changes in force through touch and autonomously adjusts the applied force and action trajectory. Dameng Robot took the lead in proposing the VTLA model architecture, with the core logic of natively integrating touch into perception, decision-making, and action generation. CraftNet released by Sharpa Robotics at CES 2026 is also an end-to-end hierarchical VTLA model, driving the North robot to complete fine operations such as folding paper pinwheels for four consecutive days.
CraftNet Released by Sharpa Robotics
The common judgment of these products and researches is: for robots to perform fine tasks, touch must enter the stage independently.
2. Why Is Tactile Sensing Booming Now?
Beyond technological consensus, another key variable for the explosion of touch in 2026 is cost.
Before 2015, the selling price of a single fingertip for high-end tactile sensors exceeded CNY 100,000, which was completely unaffordable for robots. This year, domestic manufacturers have reduced the overall cost of tactile sensors by over 90%, with Paxini's tactile sensor prices dropping to the hundred-yuan level. For the first time, tactile data has the prerequisites for large-scale collection.
Capital is also following suit. According to incomplete statistics, in 2026, about 20 tactile sensor companies completed approximately 30 rounds of financing, with the disclosed financing amount reaching about CNY 4.45 billion. The monthly delivery volume of Tashan Technology's tactile sensors has reached tens of thousands of units, and the order volume in the first half of the year exceeded four times the total for the entire last year, maintaining a market share of over 80%. The tactile penetration rate of some dexterous hand manufacturers has increased from 20% in 2025 to over 60%.
The tactile track has moved from the stage of "whether to use it" to "how to use it."
3. Five Routes, One Latecomer
Currently, there are five main technical routes for tactile sensors: piezoresistive, capacitive, Hall effect, visual-tactile, and iontronic.
Huaweike focuses on the dual-modal fusion of "magnetic + piezoresistive" and has achieved automotive-grade mass production and delivery; Fulai New Materials chose the piezoresistive route, which features low cost and easy mass production, and has entered the stage of batch delivery; Tashan Technology focuses on capacitive 3D force measurement, with a force measurement resolution accuracy of 0.01N; Paxini has deeply cultivated Hall-effect multi-dimensional perception, building a full-stack ecosystem from sensors to embodied large models; Xense Robotics and Dameng Robot are betting on the visual-tactile route; the former's visual-tactile sensors have achieved mass production at the ten-thousand-piece level, while the latter proposed the VTLA architecture and launched a tactile-anchored world model.
According to the announcement, LDROBOT and HUST will carry out the R&D and scenario verification of flexible 3D tactile sensing devices and array systems: First, develop flexible 3D tactile sensors and high-density sensor arrays to enhance perception capabilities and long-term stability under complex working conditions; Second, develop supporting signal acquisition hardware, 3D force decoupling, and tactile imaging algorithms to form a comprehensive technical solution from devices to arrays; Third, verify the landing capabilities on carriers such as dexterous hands and robotic arms.
LDROBOT's advantage lies in its customer network. It covers seven of the top ten global home service robot companies and all of the top five commercial service robot companies. These customers are not only buyers of visual products but may also become the first trial users of tactile products. However, the challenges are equally obvious: the customer base for tactile perception does not completely overlap with that of visual customers, and the procurement logic and verification cycles of humanoid robot and dexterous hand manufacturers are completely different from those of robot vacuum cleaner manufacturers.
4. The Hurdle of Mass Production: No One Has Crossed It Yet
The landscape of tactile technical routes has initially taken shape, but there are still several engineering bottlenecks spanning the journey of flexible 3D tactile sensors from the laboratory to mass production.
Signal crosstalk is the first challenge. Under high-density arrays, signals between adjacent sensing units interfere with each other, leading to a sharp decline in force decoupling accuracy. Second is the consistency of batch manufacturing—laboratory-scale preparation faces challenges of high cost, small scale, and inconsistent quality. Large-area arrayization places extremely high demands on material batch stability and process uniformity. Environmental interference is equally non-negligible; high humidity, extreme temperature differences, and electromagnetic interference can all cause signal drift, while the real working conditions of robots are far more complex than those in the laboratory.
A more fundamental issue lies in the lack of standardization. There is still no unified standard for tactile perception and operation evaluation benchmarks, and the performance data of various products lack comparability, making it difficult for downstream customers to make reliable selection judgments. How fast and how far the cooperation between LDROBOT and HUST can go in the stage "from prototype to mass-produced products" remains to be tested by time.
5. Tactile Sensing Is Not the Endpoint; Fusion Is
When a leading visual perception company decides to extend into tactile sensing, the signal it conveys is: the perception layer of embodied intelligence is moving from "superposition of single-modal capabilities" to "deep fusion of multi-modalities."
Vision determines "what the robot knows," and touch determines "what the robot can do." But what ultimately determines "how well the robot does it" is whether the two can work collaboratively within the same perception framework. This is much more difficult and much more important than simply making a single sensor well.
This article is compiled based on publicly available network information and is for reference only, not constituting investment advice.