Hundreds of billions of capital has flooded into the haptics track in half a year: a latecomer trying to take a shortcut, and the unavoidable hardware lesson it cannot evade.

A company named Gewu Tongzhi based in Hangzhou was founded in August this year, and secured nearly 100 million yuan in its angel round in October. Its investors include Peak Investment, Hongfu Assets, Hangzhou Capital, Hangzhou Talent Group, as well as a provincial-level embodied industry investment institution and multiple listed companies.

What it does is to endow robots with "tactile sensation".

This direction is not new. What is really worth noticing is its path selection: at the current stage, what it can present is a set of methods to "infer tactile sensation from videos", and it does not have self-developed sensors or robot bodies yet. But in the specified usage of this round of financing, "hardware and body construction" is clearly written - it knows exactly what it lacks, and is using this fund to make up for the shortcomings.

And this track is already extremely crowded now.

I. Tactile Sensation: The Last Centimeter of Embodied Intelligence

In the past two years, the fastest progress of embodied intelligence has been in the "vision" dimension. Robots can identify objects, plan paths, and reproduce complex movements. But vision has a ceiling: it can only observe the surface of things.

Once put into the actual factory scenario, problems will emerge. Fabrics will slip, parts will get stuck, and insertion operations will encounter resistance. These problems cannot be detected only by vision - you cannot tell the friction coefficient of the fabric, whether the part is stuck due to size deviation or angle distortion, or how much force should be applied during the insertion process.

When human hands perform these tasks, they do not only rely on eyes. Fingers can perceive "slippage", "jamming" and "reaching the limit". The magnitude, direction of these forces and the contact state cannot be provided by vision.

Without tactile sensation, robots will never go beyond the state of "looking like they can work".

This is why tactile sensation is called the "last centimeter" of embodied intelligence. It was ignored for decades: data gloves in the 1980s once sparked the first wave of imagination, and tactile companies emerged overseas around 2010, but most of them were trapped in laboratories. Until the humanoid robot industry gained momentum, this last centimeter became a must-win battlefield.

But the judgment that "tactile is a rigid demand" itself needs to answer a question first: if vision plus force control is already sufficient, is tactile sensation just a bonus item rather than a necessity? This doubt cannot be underestimated, because the valuation of the entire track is built on this premise.

Interestingly, what pushed tactile sensation from a "bonus item" to a "necessity" may not be technology, but price. Before 2015, a single fingertip of high-end tactile sensor was sold for more than 100,000 yuan, which was unaffordable for the robot industry; this year, domestic manufacturers have reduced the cost by more than 90%, and some products are priced from 199 yuan. Once the price drops, the question of "whether to install it" turns into "it is no loss to install it" - the gate of demand is opened in this way.

How fierce is the competition? The data is straightforward: in 2025, the domestic tactile sensor track completed about 20 financing events, with the total disclosed amount close to 18 billion yuan; in the first seven months of 2026, the humanoid robot sensor sector completed 15 financing events, totaling 31 billion yuan. If we add the upstream and downstream tactile-related sectors such as dexterous hands and sensors, the capital poured into this field in the past six months has reached the level of 100 billion yuan according to industry estimates - Pacini alone has raised a total of more than 40 billion yuan since its establishment.

The global tactile sensor market reached a scale of about 60 billion yuan in 2026, and is expected to hit 290 billion yuan by 2030.

A market with a scale of 60 billion yuan has accommodated a number of companies with a valuation of 100 billion yuan.

II. This Track Is Already Full of Competitors

By sorting out the major players, we can figure out the position of Gewu Tongzhi.

There are two most influential categories of players.

One category is sensor manufacturers. Pacini adopts the Hall array route, reduces the sensor price to 199 yuan and above, and has built a 12,000-square-meter embodied data factory in Tianjin, using its own sensors to collect data at low cost to feed back the model. It has completed four rounds of financing this year, with a total amount of over 40 billion yuan since its establishment. BYD is its largest external shareholder, and it has completed the counseling filing with the Beijing Securities Regulatory Bureau in September, sprinting for the title of "the first stock of embodied tactile".

Tashan Technology takes the route of capacitance and AI tactile chips. According to the company, its shipment of tactile sensors accounts for more than 80% of the domestic humanoid robot market; its revenue exceeded 100 million yuan in 2024, and the robot business is expected to account for about two-thirds of its total revenue in 2026. It has built two major tactile data collection centers, is listed by NVIDIA as the world's first tactile simulation partner for Isaac Sim, and has co-built the "Robot Kindergarten" with Turing Award winner Sutton.

Qianjue Robotics takes the capacitance route, claiming a measurement density of 50,000 points per square centimeter, and provides the tactile simulation tool Xense_Sim, and has reached cooperation with Agibot, Zibianliang, Galaxy Universal and Qwant.

Daimeng takes the visual-tactile route - using cameras to capture the deformation of the elastic body under pressure to infer the contact force, which is a technology homologous to GelSight, with 40,000 sensing units integrated per square centimeter. It not only develops sensors, but also builds a vision-tactile-language operation model and a tactile evaluation benchmark, forming a closed loop of "sensor-data-model".

There are also Weitai (visual-tactile fingertip sensors that have been applied in Xiaomi's mobile phone production lines), Yimu Technology (the thinnest 3mm visual-tactile sensor, with over 1 billion yuan in Series E financing and a valuation of over 10 billion yuan), Xinuo Future (fingertip tactile, with a total of about 1.5 billion yuan in three rounds of financing in four months), and Moxian Technology (flexible thin film piezoresistive).

The other category focuses on model development. Lingxin Qiaoshou independently develops tactile arrays and also manufactures dexterous hands. It has obtained seven rounds of financing within one year, with a valuation of about 3 billion US dollars after the B+ round, and occupies 80% of the global high-degree-of-freedom dexterous hand market. Oak Nut adopts the "instinct-driven" route, advocates zero-data cold start, and has obtained two rounds of large-scale financing in four months.

This list illustrates one thing: almost all players on this track are "software and hardware integrated" - they develop their own sensors, build their own data factories or collection centers, and develop their own models. But almost all of their data comes from real sensors, with high collection cost and slow expansion speed.

Gewu Tongzhi is the only company on this list that, according to public information, does not have available sensors and robot bodies yet. This does not mean it does not plan to develop them - "hardware and body construction" is clearly stated in the financing purpose - but its hardware products have not been produced yet, while other players' products have already been shipped.

III. The Path It Chose Has Already Been Tried by Others, and Relevant Error Data Has Been Obtained

Gewu Tongzhi's solution is called Real to Sim: "infer" tactile signals from Internet videos.

The idea is: there are massive scenes of "hands twisting, inserting, pressing" in videos, the movements are visible but the force is invisible; so we can infer the invisible contact, force and deformation from the visible images, and add physical labels to the movement data for pre-training.

Mathematically, this idea is called ill-posed inverse problem - the same result may correspond to countless causes. In the image, the hand presses down and the fabric dents, but how many newtons of force are applied, what the contact area is, what the material stiffness is, and what the friction coefficient is, all are unknown. The number of unknown parameters is more than that of observation equations, so the solution is not unique. There is no direct force signal in the image, and only physical prior knowledge can be supplemented - but the accuracy of the prior knowledge ultimately has to be calibrated by real force and tactile data. Even if ten times more videos are input, the information that does not exist in the original images cannot be supplemented.

What is more noteworthy is that Gewu Tongzhi is not the first to take this path.

PoXiao Intelligence - founded by Yang Shuo, a tenured professor at Harbin Institute of Technology (Shenzhen) born in 1998 - is doing research in the same direction. Its TouchAnything model is trained to restore the pressure distribution of both hands from ordinary first-person videos. And it has released the accuracy data: the estimation error of contact force map on real objects is about 0.56 newtons, while the previous best method had an error of 2.1 newtons, representing an improvement of nearly four times. It should be noted that this is a public research result, not a mass-produced product indicator, but it at least proves that this path is feasible in terms of accuracy.

In the same direction, ZhiZai WuJie - founded by Lu Zongqing, a tenured associate professor at Peking University - adopts a different method: using 500,000 hours of human videos to pre-train the "implicit tactile world action model", and using the TactoHand module to infer whether contact occurs based on the 3D geometric relationship between hands and objects.

However, there is a conflicting view in the industry. Ma Daolin from Qianjue Robotics said: Tactile is the only modality that cannot "get data for free" from the Internet at present.

This sentence is almost directly targeted at the idea of "inferring tactile sensation from videos". If tactile data really cannot be obtained for free from Internet videos, then Real to Sim is doing something that top industry researchers consider impossible.

Therefore, it is not facing a path that no one has ever walked. On the contrary, some people have already achieved results on this path, and the outcome is still undecided. Gewu Tongzhi stands on one side of the debate, but it is not the first in this camp.

IV. The Unavoidable Hardware Course and Its Real Opportunity

Looking at its own capability structure, the problem becomes clearer.

To realize the landing of tactile data, it must eventually be implemented on sensors. But Gewu Tongzhi does not have its own sensors at present. There are several incompatible routes for tactile sensors - piezoresistive, capacitive, piezoelectric, Hall, optical, and each route has different data formats, ranges, accuracy and noise characteristics. If the type of sensor changes, the model may need to be recalibrated.

This means its "versatility" is restricted by the fragmentation of upstream hardware. And this variable is out of its control.

What is more delicate is that upstream players are already expanding downstream. Tashan has built two major tactile data collection centers, Pacini has built a 12,000-square-meter data factory, and Daimeng has developed the VTLA model and evaluation benchmarks - sensor manufacturers are also developing data and models, and they have hardware, production lines, customers and sufficient capital.

Downstream players are also developing their own related technologies. Agibot not only invested in Qianjie Technology, which develops a unified tactile action model, but also adopted the tactile solution from Qianjue - robot body manufacturers do not put all their eggs in one basket.

Therefore, Gewu Tongzhi is in an awkward position: upstream are sensor manufacturers that already have production lines and customers, and downstream are body manufacturers that not only invest in model developers but also select their own sensor solutions. The layer it wants to develop is exactly sandwiched in the middle.

There is another account to calculate. One of the endorsements for its founder Cai Jia is "a systematic methodology for ten million-hour level data governance and nearly ten-thousand-card cluster training". This description refers to his experience at Huawei, not the resources of this company. The computing power hardware of a ten-thousand-card cluster alone costs from billions to tens of billions of yuan, while it only raised nearly 100 million yuan in this round - which is not even enough to buy 1,000 cards. And "large-scale pre-training of physical inversion world model" is exactly a work that consumes a huge amount of computing power.

The more expensive cost is time. Tashan was founded in December 2017, Pacini started its business in June 2021, and Daimeng began to research visual-tactile technology in 2016. The head companies on this track have polished their hardware products for 5 to 10 years. But Gewu Tongzhi is just starting its "hardware and body construction" now.

This means it has to make up lessons from scratch in a field where others have already built production lines, optimized yield rates and acquired customers. Making up for lessons is not impossible: hardware can be purchased, OEM services can be adopted, and talents can be recruited. But the gap of "5 to 10 years" cannot be filled at once with a fund of nearly 100 million yuan.

Therefore, its real bet is to make the data layer a non-negligible part for others before completing the hardware layout - whether hardware is important or not is another matter. This is a race against time: either the moat of data is built first, or the hardware course drags it down first.

But focusing on data first and then supplementing hardware does not mean it has no cards to play.

If the Real to Sim technology is proved feasible, the most valuable asset of this company will not be the "embodied brain", but the production line that "converts videos into data with tactile labels". Because the track for developing embodied brains is already overcrowded, if the technology of "batch converting Internet videos into tactile data" is realized, it can be sold as data infrastructure to all companies engaged in dexterous operation - this business model is lighter and has stronger scale effect.

But the story it is telling now is: to build an embodied brain, enter industrial scenarios, and deliver projects one by one.

This is a project-based business: each customer requires customization, on-site support, debugging and acceptance, with linear revenue and high labor intensity. The valuation logics of the two paths are completely different, the former is valued as a platform, while the latter is valued based on orders.

Its cooperation with Peak is the most typical case - Peak Investment is an investor in this round, and it also signed a strategic cooperation agreement with Gewu Tongzhi, focusing on the "double flexibility and one high precision" (flexible manufacturing, flexible materials, sub-millimeter level precision) of shoe manufacturing to "obtain large-scale orders" in the early stage of the company. This structure is the same as that of car manufacturers who act as both customers and shareholders: it is difficult to distinguish in the early stage whether the order is the market recognition of the technology, or the care from the shareholder to the early-stage project.

Cai Jia said that the most important indicator the company values is not the precision success rate, but the customer repurchase rate. This judgment is very sober. But a company founded in August whose first PoC will not be delivered until the end of the year has not even completed its first order, so talking about repurchase is talking about a number that does not exist yet.

As for the reason why it chose Hangzhou - it has obtained investment from Runmiao Fund, Disruptive Achievement Transformation Fund and Talent Fund in the first round, and it is also an ecological partner of the National Embodied Intelligence Application Pilot Base and a member of the Hangchuang Embodied Intelligence Camp. Hangzhou already has Unitree and Deep Robotics in the embodied intelligence sector, both of which focus on "legs"; while Gewu Tongzhi focuses on "hands". With legs available and hands missing, the company was recruited to fill this gap.

This also explains the speed of "obtaining nearly 100 million yuan two months after its establishment": part of this fund is the capital for industrial investment attraction, not purely based on financial judgment. For Hangzhou, whether Gewu Tongzhi can become a great company is secondary, the important thing is that the "hand" sector needs to be filled.

Conclusion

Back to this company.

The direction it chooses is correct - tactile sensation is changing from a "bonus item" to a "necessity", which is the last centimeter of embodied intelligence. But on this track, there are already sensor companies with a valuation of 10 billion yuan sprinting for IPO, chip companies with 80% shipment share building collection centers, dexterous hand companies with a valuation of 3 billion US dollars, and young professor teams that have achieved an error of 0.56 newtons in inferring tactile sensation from videos.

As for Gewu Tongzhi, it was founded two months ago, raised nearly 100 million yuan, and has no self-developed sensors, robot bodies or data factories yet. Its founder has an impressive resume and an unproven method. It plans to use this fund to develop hardware, but other players have been working on hardware for many years.

But on the other hand, focusing