In the realm of embodied artificial intelligence, the focus is shifting from the "brain" to encompass more of the "senses."

Over the past two years, with advancements in Vision-Language-Action (VLA) models and world models, the cognitive capabilities of embodied AI have become increasingly sophisticated. However, many companies are finding that when robots are deployed into real-world environments like factories and homes, they still frequently encounter problems such as unstable grasping, inaccurate handling, and poor assembly.

How to enable robots to truly understand the physical world through higher-quality data has become a new challenge for the entire industry.

During the recent 2026 World Artificial Intelligence Conference (WAIC), Yimu Technology's founder and CEO, Li Zhiqiang, noted in a discussion that demand for haptic data has seen significant growth over the past six months. "Previously, the focus was mainly on proprioceptive data collection. Now, both clients and academia are strongly requesting haptic data," he said.

Haptic sensing is considered the final piece of the puzzle for embodied AI. Why has the demand for haptic data surged in the industry over the last half-year?

Li Zhiqiang believes the first reason is that the purely visual approach is gradually reaching its limits. Initially, the industry needed to verify the feasibility of solutions like VLA. While models can now perform many demos, a significant gap is exposed when it comes to real-world execution.

"People often say 'I've seen how to do it' when learning, but when they actually try, they can't do it," Li illustrated. The same applies to robots. Relying solely on vision allows them to understand actions but not the physical information like friction, softness/hardness, or contact feedback during actual operations. Consequently, performance drops noticeably in real environments.

The second reason is the growing capability, from hardware to data interpretation, across the entire haptic data ecosystem. Li Zhiqiang stated that the industry's recognition of the value of tactile sensing is also rising rapidly.

In fact, Yimu Technology's demonstration at WAIC of a robot distinguishing between real and fake peanuts is a direct showcase of haptic data's value. Vision cannot differentiate the two peanuts, but the robot, through a single physical touch, can obtain feedback on pressure, deformation, and other physical properties to make a judgment. For robots, the most valuable data is not images from the internet, but the continuous stream of Ground Truth (physical reality) generated by the real world.

How much haptic data is needed to train a mature embodied AI model?

Li Zhiqiang pointed out that the industry previously generally believed a mature robot model required at least 500,000 hours of data training. Many companies are now pushing towards 500,000 to 1 million hours, with some within the industry even suggesting the volume of data collected next year may need to reach 10 million hours.

"Our view is that data incorporating haptics should be on a similar scale, at least not an order of magnitude lower," Li noted. He also added that the cost of adding a haptic modality to existing collection systems is very low, possibly less than 1%, making the return on investment very high.

However, compared to future demand, the industry's current stock of haptic data remains very limited.

Li Zhiqiang revealed that the amount of truly effective haptic data available might only be between tens of thousands to one hundred thousand hours. Among all embodied AI data, haptic interaction data is still the scarcest component, a shortcoming the entire industry needs to address promptly.

This indicates that haptic data infrastructure will become a key focus for the embodied AI industry in the coming years.

Yimu Technology's core products are a full series of visual-tactile sensors (in shapes like square, wedge, and fingertip).

Regarding the challenges in haptic data collection, Li Zhiqiang believes that compared to 2025, the industry's biggest challenge is no longer whether haptic hardware can be mass-produced and standardized. The real breakthrough needed this year is achieving large-scale deployment, including product consistency, stability, data alignment across different sources, and algorithmic capabilities for understanding and interpreting haptic data.

To address the high demand for haptic data, Yimu Technology has built a comprehensive infrastructure system. At the foundation, visual-tactile sensors continuously collect real interaction data. In the middle layer, a data collection system and a Tactile Transformer Encoder convert complex physical interactions into structured knowledge that models can learn. This is further fused with multimodal data like vision and language to help robots develop a Grounded Transition Model (GTM) and Affordance capabilities, enabling them to understand "what object is being faced, how much force to apply, and what the final outcome will be."

Li Zhiqiang stated that Yimu Technology's current clients are mainly divided into two categories: large tech companies with in-house model teams, and data service providers.

With rapid growth in industry demand, Yimu Technology's business performance has also shown significant improvement. "The growth in performance over these past six months will be even more pronounced," Li said.

It is worth noting that on July 18th, Yimu Technology announced the completion of an E-round financing exceeding 10 billion yuan, with a post-investment valuation surpassing 100 billion yuan. This funding round saw participation from multiple top-tier RMB funds, leading USD funds, and industrial capital. The funds will primarily be used for R&D in embodied AI haptic sensing materials, chips, algorithms, and models, as well as for scaling up mass production and order fulfillment.

Over the past two years, attention in the embodied AI industry has been concentrated on "hardware platforms" and "large brain models." Now, a quiet competition around data has begun.