By
Editor NUPIAO
Shenzhen’s embodied AI scene just got another unicorn.
On June 30, Crosswise Intelligence announced it had closed a Series B funding round worth 1 billion yuan (around $140 million), pushing its post-money valuation past the 10-billion-yuan mark. The company says it will now kick off its IPO process.
One quick look at the investor list and you’ll spot a heavy state-capital flavor: Shenzhen Capital Group and the Guiyang Digital Economy Fund have now doubled down for two straight rounds. New players include Qianhai Mother Fund, Lens Technology, ICBC Capital, Hengjian Asset, and Zhurui Capital, while existing backers like Nanshan Strategic Emerging Industries Investment, Chengdu Science and Technology Innovation Investment, and Sichuan Academician Fund also joined in.
What’s particularly interesting is the appearance of Zhou Qunfei, chairwoman of Lens Technology, on the shareholder roster. This is the first time she’s personally invested in an embodied AI company. Before this, Lens Technology was already one of Crosswise Intelligence’s clients.
Just a day before the announcement, two other Shenzhen embodied AI companies—Independent Variable Robotics and Zhipingfang—were reported to have crossed the 20-billion-yuan valuation mark. Independent Variable Robotics closed four rounds in two months, drawing in ByteDance, Xiaomi, Meituan, and others. Zhipingfang, meanwhile, added nearly 5 billion yuan. Almost at the same time, companies like Jijia Vision and Qianxun Intelligence also piled up multiple rounds at a dizzying pace.
In just a few short months, the embodied AI track has been popping out unicorns with 10-billion-plus valuations like popcorn. According to ITjuzi, over 46 billion yuan poured into the embodied AI and robotics space in the first half of 2026 alone. But here’s the catch: 70% of that money went to the top 20 players. The Matthew Effect is already brutal.
Crosswise Intelligence has managed to squeeze into that top tier. Founded in 2021, the company is led by Jia Kui, a tenured professor at the Chinese University of Hong Kong, Shenzhen. Jia had been a pure researcher for years, focusing on AI applications in 3D space, and Crosswise was one of the earliest teams in China to apply AI to 3D environments. That background naturally shaped the company’s technical approach: train robots using synthetic data and physics simulation, then deploy those models into real-world scenarios to make money.
Back then, that wasn’t exactly the mainstream bet. The industry largely believed you could just pile up real-world teleoperation data and feed it to a general-purpose robot brain. But Jia points out that a single teleoperator can collect maybe 100 to 150 data points a day. At that rate, covering the semantic generalization you’d need for embodied AI would theoretically take 100,000 years.
Now that the efficiency bottleneck of real-world data is becoming painfully obvious, the simulation-and-synthetic-data route is getting a second look. And that’s a consensus that’s starting to build across the industry.
Crosswise Intelligence has already run through the entire pipeline—from generating synthetic data, to training the model, to deploying on real robots—all powered by its self-developed DexVerse™ embodied AI engine. Jia explains that this system can slash the development cycle for a new robotic scenario by 90%. For a new task in a general smart manufacturing setting, from data generation to model training to actually going live, the whole thing can be done in six to eight hours. For a humanoid robot commercial service task, from project kickoff to deployment, it usually takes just a few days.
One of the hottest topics in embodied AI this year, after VLA (Vision-Language-Action) models, is that nearly every major player is now telling a world-model story. Crosswise is no exception.
Jia’s take is that VLA is essentially a form of “shortcut learning.” The robot sees an image, receives a language command, and directly outputs an action, skipping any real understanding of the physical laws behind the scene. And even with mountains of real-world data, there’s no evidence that this kind of end-to-end modeling is efficient at generalization. Throw in an unfamiliar object, a new environment, or a bit of interference, and the model tends to fall apart.
But world models haven’t won everyone over, either. Guo Yandong, founder of Zhipingfang, has pointed out that the world models being widely discussed today aren’t actually driven by physical laws. They’re trained on massive datasets. When there’s enough data, the model knows a cup will fall—but that’s not a summary of physics, it’s just the result of big-data learning.
Crosswise Intelligence recently released Dexterity-BEV, its concrete implementation of a world model. It borrows the well-proven BEV (Bird’s-Eye View) approach from autonomous driving, compressing raw observation data from multiple cameras and sensors into a unified, callable physical coordinate system. This lets the system understand the world in 3D space—the exact step that allowed the autonomous driving industry to make its technological leap.
Crosswise wants to bring that same logic to robots: align visual inputs, robot joint states, and end-effector target actions into a single BEV coordinate system. That way, the data silos that naturally exist between different robot bodies and different operators can be broken down. And the company’s differentiator is that, from the architecture design stage, it embeds 3D spatial geometric constraints and motion laws directly into the model structure—not just relying on a pile of data.
A week after Dexterity-BEV was released, Alibaba unveiled its own world model, Qwen-RobotWorld, which follows a similar logic. As end-to-end VLA modeling hits a generalization wall in complex scenarios, the industry is now poking at different angles, trying to find representations that are closer to actual physical laws, and searching for some kind of consensus on how to make robots truly understand the world.
On the commercialization side, Crosswise Intelligence is walking on two legs. One is using its model capabilities to empower third-party robotic arms or robots, tackling tasks like sorting, assembly, and plugging/unplugging in smart manufacturing. That’s currently its most mature business. The other is using its own humanoid robot, the DexForce W1 series, to break into commercial service scenarios that need human-robot interaction, like coffee making, smart retail, and cultural tourism guiding. During this year’s May Day holiday, Crosswise’s robots recorded over 100,000 yuan in revenue per store across more than a dozen locations.
As of now, Crosswise Intelligence has landed in over 50 niche industries. In the first half of 2026, its humanoid robot shipments were in the low hundreds, which still puts it in the second tier in the industry. But Jia revealed that the company crossed the 100-million-yuan revenue mark in 2025, and for 2026, it expects that number to triple year-on-year.
On the revenue model front, Jia has floated an interesting idea: take a page from how large language models charge by token usage. In the future, instead of just selling the hardware, you’d charge based on the actual compute and data consumed by the robot. Right now, though, real-world examples following that logic are still rare. Hardware remains the primary revenue vehicle. But he also believes that once token consumption can create incremental value for customers, embodied AI companies can further benefit from that, and the upfront importance of hardware purchases will start to fade.
That might just be the next big horizon for embodied AI players.
