Dialogue with Li Zhiqiang, CEO of Yimu Technology: Touch Becomes the New Battlefield for Embodied Intelligence
```
Author | Huang Yu
The battlefield of embodied intelligence is expanding from the "brain" to more "senses".
Over the past two years, as VLA (Vision-Language-Action) models and world models continued to develop, the "brain" of embodied intelligence has become increasingly smarter. However, many companies have found that when robots enter real environments such as factories and homes, they still frequently encounter problems like "unstable grasping, inaccurate picking, or poor assembly."
How to enable robots to truly understand the physical world through more high-quality data has become a new subject for the entire industry.
Recently, during the 2026 World Artificial Intelligence Conference (WAIC), Li Zhiqiang, founder and CEO of Yimu Technology, said in a conversation with Wall Street that in the past half year, there has been a very obvious increase in demand for tactile data. "Previously, everyone mainly focused on subject data collection, but now both clients and academia are strongly demanding tactile data."
Tactile perception is considered the last piece of the puzzle in embodied intelligence. Why has demand for tactile data in the industry grown significantly in the past six months?
Li Zhiqiang believes that the first reason is that the pure vision route is gradually approaching its capability limits. The industry first needed to verify the feasibility of solutions like VLA. Although the models can now accomplish many demos, a huge gap is exposed when it comes to actual execution in the real world.
"When people learn, they often say, 'I understand by looking,' but when they actually try it, they can't do it." Li Zhiqiang says robots are the same: relying solely on visual data enables understanding of actions, but cannot comprehend physical information such as friction, hardness, or feedback during real operations. So once they enter real environments, their performance noticeably drops.
The second reason is that the overall capability to understand tactile data from hardware to data interpretation is improving. Li Zhiqiang states that the industry's recognition of the value of tactile perception is rapidly increasing.
In fact, Yimu Technology's demonstration of distinguishing real and fake peanuts at WAIC is the most direct embodiment of the value of tactile data.
Vision cannot differentiate two peanuts, but a robot can obtain physical feedback such as pressure and deformation through real contact, thereby making a judgement. For robots, the most valuable data is not internet images, but the Ground Truth (physical truth) continuously produced in the real world.
How much tactile data is needed to train a mature embodied intelligence model?
Li Zhiqiang points out that the industry previously believed that a mature robot model required at least 500,000 hours of data training. Many companies are now pushing toward 500,000 to 1 million hours, and internal discussions suggest that next year the data collected may need to reach 10 million hours.
"Our view is that data with tactile perception should be on a similar scale—certainly not less than an order of magnitude." Li Zhiqiang also notes that adding tactile modalities to existing collection systems has a very low cost, maybe less than 1%, so the return on investment is very high.
However, compared to future demand, the industry currently has very limited tactile data.
Li Zhiqiang reveals that currently, truly effective tactile data is only between tens of thousands and 100,000 hours. Among all embodied intelligence data, tactile interaction data is still the rarest piece and the main shortcoming the entire industry needs to address urgently.
This means that in the next few years, tactile data will become a new focus for basic infrastructure construction in the embodied intelligence industry.
Yimu Technology's core products are a full series of visual-tactile sensors (square, wedge, fingertip, etc. forms).
Regarding the difficulty of collecting tactile data, Li Zhiqiang believes that compared to 2025, the biggest challenge for the industry is no longer whether tactile hardware can be mass-produced and standardized. This year, the real breakthrough needed is large-scale deployment, including product consistency, stability, alignment between different data sources, and algorithmic capabilities to understand and analyze tactile data.
To meet the high demand for tactile data, Yimu Technology has built a complete infrastructure system.
At the base level, visual-tactile sensors continuously collect real interaction data. The intermediate layer, through the data collection system and the Tactile Transformer Encoder, transforms complex physical interactions into structured knowledge that models can learn. It then integrates with multi-modal data like vision and language, helping robots establish Grounded Transition Models (GTM) and Affordance abilities, letting robots understand "what object they're facing, how much force to apply, and what result will occur."
Li Zhiqiang says Yimu Technology's customers are mainly two types: large tech companies with modeling teams and data service providers.
With industry demand rising rapidly, Yimu Technology's performance has also clearly improved. "This half year, performance growth is even more pronounced," Li Zhiqiang said.
Notably, on July 18, Yimu Technology announced the completion of an E Series financing of over 1 billion yuan, with post-investment valuation exceeding 10 billion yuan. This round was jointly participated in by several top-tier RMB funds, leading USD funds, and industry capital. Funds will mainly be used for embodied intelligence tactile perception materials, chips, algorithms, and model R&D, as well as mass production and order delivery.
In the past two years, attention in the embodied intelligence industry was focused mainly on "subject hardware" and "large brain models," but now a competition around data is quietly beginning.
Risk warning and disclaimerThe market involves risk, investment requires caution. This article does not constitute personal investment advice and does not take into account the particular investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their specific circumstances. Any investment based on this is at one's own responsibility. ```