Visual-tactile sensing addresses the pain point of "unstable gripping", and Inspire Robotics has completed its angel round of financing.
Robots can run, jump, and perform backflips, yet they may not be able to grasp a single strawberry steadily. As embodied intelligence shifts from motion capability to manipulation capability, a repeatedly verified bottleneck is emerging: vision allows robots to "see" the target, but cannot tell them how large the contact force is, whether the object is slipping, or what the surface texture is like. If the grip is too light, the object cannot be held; if too tight, it will be crushed. This gap in the tactile dimension is a necessary step that robots must cross to move from demonstration to real-world operation.
AGIMind Robotics (the industrial entity is AGIMind Robotics Technology (Shenzhen) Co., Ltd.) chooses to cut into this gap. Founded in October 2025, the company does not manufacture robot bodies nor develop general large models, and its R&D focus is concentrated on robot tactile perception and fine manipulation, aiming to push robots from "being able to see" to "being able to understand contact". In October 2026, AGIMind Robotics completed tens of millions of yuan in angel round financing, with investors including institutions such as Lihe Capital and Hefei Innovation Investment. The funds from this round are mainly used for production line construction, mass production capacity optimization and R&D of tactile intelligence systems. According to public information, the company is currently pushing forward subsequent financing.
Where is the "last centimeter" of tactile perception stuck in the process
The core difficulty of robot fine manipulation does not lie in whether the model "wants to do it", but in whether the underlying signals are sufficient. Visual sensors output images, while traditional force sensors output a single scalar force value, neither of which can directly answer "what is happening at every point on the contact surface". In assembly tasks involving transparent workpieces, reflective surfaces, fragile objects or requiring slip prediction, this lack of information will directly lead to grasping failure or workpiece damage.
A deeper problem lies at the electronic architecture level. Industry research points out that a considerable proportion of domestic robot underlying electronic systems still rely on overseas real-time operating systems and communication protocols. The non-uniform interface standards and scattered technical routes limit the access cost and collaboration efficiency of new perception modules such as tactile sensors. The reason why tactile perception has stayed in the laboratory stage for a long time is not only due to the difficulty of the sensing principle itself, but also the lack of a mass-producible and integrable engineering path from the sensor to the robot control link.
AGIMind Robotics's response is to start from the "meta-component" layer of sensors, and then extend to data acquisition, algorithm library and control middleware, forming a full-stack link covering "tactile perception - data construction - model training - control deployment". Its core product Lux series visuo-tactile sensors have achieved large-scale delivery, which are compatible with dexterous hands, robotic arms and electric grippers, and have been applied in scenarios such as data acquisition, industrial assembly, logistics and scientific research automation.
Dense force field route and 5mm fingertip: The engineering implication of a technical choice
The basic principle of visuo-tactile sensing is not new: a micro camera is used to capture the deformation image of the flexible contact surface when force is applied, and then the force distribution is inversely solved through algorithms. The three-color light scheme proposed by MIT in 2009 is a classic route in this field, which has been led by overseas teams for a long time thereafter. The differentiation of AGIMind Robotics is that it does not follow the three-color light scheme, but takes a "dense force field" path: one contact imaging can simultaneously solve the force, deformation, slip and posture. At the algorithm layer, pixel-level deformation tracking and force field inverse solution are adopted to realize the output of "force and deformation from the same origin, native multi-modality".
Judging from the publicly disclosed product specifications, the direct result of this route is two engineering parameters: the thickness of the sensor module is 5mm, the density of deformation tracking points reaches more than 100,000 points per square centimeter, and the force perception accuracy is 0.01N. The 5mm thickness means that the visuo-tactile system can be compressed to the scale of commercial dexterous fingertips, instead of only being externally mounted on the end of the robotic arm in the form of an independent module. The tracking density of 100,000 points per square centimeter points to higher spatial resolution. In precision assembly and flexible grasping, the detailed difference of force distribution on the contact surface is often the key to judging whether the object is "grasped stably".
It should be distinguished that these indicators come from the laboratory or product specification conditions disclosed by the company, not independent reproduction by third parties under arbitrary working conditions. However, the fact that the product has entered the stage of large-scale delivery itself provides another reference: the sensor is no longer just a principle prototype, but needs to be tested by customer scenarios in engineering dimensions such as production line consistency, long-term drift, and packaging reliability. The company's founding team has been engaged in visuo-tactile research since 2016, developed a 5mm thick visuo-tactile sensor in 2019, and applied visuo-tactile to slip prediction in dexterous operation. The relevant accumulation has constituted a first-mover advantage in time after entering the productization stage.
Delivered products and ongoing mass production propositions
To judge the position of an early-stage tactile company, the delivery status is more worthy of attention than the parameter sheet. AGIMind Robotics's Lux series sensors have achieved large-scale delivery, with customer scenarios concentrated in industrial assembly, logistics sorting, data acquisition and scientific research automation. The common point of these scenarios is that the force information during the contact process directly determines the success or failure of the task, while the supplementary role of vision in them is limited. The plugging force control in industrial assembly, the adjustment of grasping force for unknown objects in logistics scenarios, and the large-scale acquisition of tactile signals in the data acquisition link all correspond to the operation links that traditional visual solutions are difficult to cover.
However, "large-scale delivery" in early-stage hardware companies usually corresponds to the order status of small batches and multiple batches, and there is still a gap before large-scale production on production lines. The company clearly designates the first use of angel round funds for production line construction and mass production process optimization, which is exactly the hurdle between "being able to deliver" and "delivering stably and at low cost". An engineering feature of tactile sensors is that the batch consistency of flexible sensing materials, the yield of assembly processes, and the wear and drift during long-term contact will be multiplied in the process of mass production scaling. The gap between laboratory indicators and industrial customer acceptance standards is usually not an algorithm problem, but a manufacturing consistency problem. The current progress of AGIMind Robotics needs to be understood within this framework.
In addition to sensors, the company is extending its capabilities to electronic skin, 6-axis force sensors, data acquisition systems and tactile operation algorithm libraries. Among them, the significance of the supporting algorithm library and control middleware is to reduce the access cost for customers: after customers get the hardware, they do not need to develop the parsing and control logic of tactile signals from scratch, and can directly call the interfaces for grasping, force-controlled assembly and surface inspection. If this layer can mature synchronously with the hardware, the penetration speed of tactile perception will no longer solely depend on the unit price of sensors, but more on the integration efficiency on the scenario side.
Dual background of the team and the window period of the tactile track
Zhang Yazhan, founder and CEO of AGIMind Robotics, graduated from the University of Science and Technology of China with a bachelor's degree in precision instrumentation, and obtained his doctorate in robotics from the Hong Kong University of Science and Technology. He used to be a senior AI researcher at Tencent, leading the R&D of large-scale data pipelines and AI models, and later served as a researcher and doctoral supervisor at a national laboratory. The core team comes from universities including HKUST, USTC, SUSTech and EPFL, and includes industrial personnel from enterprises such as Tencent and Orbbec, covering the full-stack R&D links from sensing, data, models to control. The academic accumulation provides in-depth research in the cross field of visuo-tactile sensing and robot learning, while the industrial experience corresponds to the practical operation capabilities of data pipelines, hardware engineering and mass production promotion. For an early-stage project that needs to deal with sensor physical design, signal parsing algorithm and customer scenario integration at the same time, the combination of these two capabilities is a necessary rather than decorative condition.
The warming up of the tactile perception track has traceable external signals. According to GGII's prediction, the global demand for flexible tactile sensors or electronic skin in the field of humanoid robots will reach 1.525 million square meters in 2030, corresponding to a market size of about 27.4 billion yuan. A more operationally meaningful reference comes from experimental data: in specific gripper pick-and-place tasks, the model introduced with tactile signals increases the success rate from 75% to 96.9% compared with the pure vision benchmark. Whether this improvement can be maintained in more complex scenarios such as dexterous hands and assembly lines depends on the advancement of data scale and generalization verification, but it at least shows that the gain of tactile sense on operation capability is not a conceptual judgment, but can be measured under certain conditions.
Policies are also pointing out the direction. The 2025 *Government Work Report* included "embodied intelligence" in the future industry cultivation direction for the first time. The *Guiding Opinions on the Innovative Development of Humanoid Robots* issued by the Ministry of Industry and Information Technology proposes to initially establish an innovation system by 2025 and form a safe and reliable industrial chain and supply chain by 2027. As the "interface layer" of robot physical interaction, the localization and mass production capability of tactile sensors are directly related to whether robots can move from structured environments to unstructured scenarios. AGIMind Robotics is right in the transition link from "whether it exists" to "whether it is affordable and accessible".
For AGIMind Robotics, which is pushing forward subsequent financing, the nodes worthy of continuous tracking in the next stage are clear: the actual ramp-up pace of the mass production line, the integrated verification of the algorithm library and control middleware in customer scenarios, and the extension speed from industrial assembly and logistics to a wider range of fine operation scenarios. The competition in the tactile track is shifting from "who develops the sensor first" to "who can turn the sensor into a stably operating part on the customer's production line", and the completion degree of this transformation will determine whether early technical accumulation can be converted into a scalable product position.