Chinese embodied intelligence going global: from selling robots to open-source models entering overseas toolchains

In April 2025, on Tesla's first-quarter earnings call, Morgan Stanley analyst Adam Jonas asked Musk: in physical AI fields such as humanoid robots, who is more advanced, China or the United States?

Musk was confident that Tesla would take first place, but his next sentence shifted the topic to China. He believed that China's story of leapfrogging in new energy vehicles could repeat itself in embodied intelligence. "I'm a bit worried that on the (robot) leaderboard, the companies ranked second through tenth will be Chinese companies."

More than a year later, it is still difficult to rank global robotics companies. But the trend Musk predicted—"Chinese embodied intelligence companies going overseas"—is becoming a reality.

Data from EO Intelligence shows that among the 182 funded embodied intelligence companies it counted, 73 have already expanded overseas and 21 are preparing to do so, together accounting for about 51.6%.

Unitree founder Wang Xingxing once said that overseas business accounted for about 50% of the company's annual performance in the past. A group of embodied intelligence companies such as AgiBot and UBTech have also, in their strategic plans and launch events this year, tacitly placed globalization and overseas expansion at the center of the stage, and demand for overseas embodied intelligence talent recruitment has risen accordingly.

In addition to selling robots overseas, Chinese embodied intelligence companies are also continuing to push forward at the "brain layer," participating in the co-building of the global technology ecosystem by adhering to open-source embodied intelligence general large models. Take Ant Lingbo as an example: the LingBot series models have accumulated more than 33,000 stars on GitHub, and downloads have exceeded 2.2 million.

This is a structural turning point. Going overseas is becoming another incremental growth curve for China's embodied intelligence. From winning overseas attention through body manufacturing and motion control, to general model capabilities beginning to enter the global competitive landscape, China's embodied intelligence is standing on the world stage in different ways.

Overseas investment banks including Morgan Stanley have repeatedly raised their forecasts for Chinese humanoid robot shipments, and one of the core reasons is precisely overseas opportunities. According to its data, of the 13,000 to 16,000 humanoid robots shipped globally in 2025, Chinese manufacturers accounted for about 90%—on average, 9 out of every 10 sold came from China.

Morgan Stanley analysts even believe that the humanoid robot and robotics industry will become the next key driver of China's export machine in the next 5 to 10 years. "Similar to the EV industry ten years ago."

Source: CEIC, Haver, Morgan Stanley Research

The first layer is in the body—cerebellum: China has a complete supply chain, component supporting facilities, and advantages in cost and delivery speed. At the same time, many embodied intelligence companies that have long adhered to a "cerebellum-first" approach have accumulated leading cerebellum motion control capabilities, accelerating the integration of software and hardware.

The second layer is in scenarios—data: China's rich manufacturing and service scenarios provide conditions for robots to carry out real-task validation, collect interaction data, and continuously iterate. Experience accumulated in real scenarios helps improve products and model capabilities, laying the foundation for expanding overseas applications.

The third layer is in open source—brain: through open source, efforts are being made to turn the local innovation experience of China's embodied general models into a key force in the global embodied technology infrastructure. In global developer collaboration and the co-building of technology and industry ecosystems, the technological influence and ecosystem discourse power of China's embodied intelligence continue to expand.

It can be seen that the advantages of the three layers support one another and jointly drive today's Chinese embodied intelligence going overseas, which is extending from hardware product exports to model capability openness and global ecosystem collaboration. Supporting this key process are manufacturing and supply chain advantages, data accumulation and capability iteration in real scenarios, and global developer participation brought by open source.

02

China's embodied brain refuses "cross-server communication"

If what the body and cerebellum solve is "can be built and can move steadily," then China's advantages in embodied intelligence are self-evident. On the contrary, China's embodied brain power—responsible for robot perception, decision-making, and driving the body to complete tasks—has to some extent been underestimated.

The reason for being underestimated is simple, stemming from an established consensus in the industry: China's embodied intelligence is strong in hardware and bodies, while the United States is strong in brains and models. With the emergence of GPT-6 Astra, discussions about whether "large language models will eat embodied models" have arisen again.

Image | U.S. Wired magazine reports on Chinese robots

But as many industry insiders have said, the value of this discussion does not lie in determining the winner of technological routes, but in prompting the industry to return to pragmatism: return to real tasks and the physical world, and re-examine what capabilities robots actually need.

From "can understand" to "can work," after understanding tasks and making plans, robots also need to predict how actions will change the environment, continuously correct based on execution feedback, and meet real-time response requirements. These specific needs not only pose new questions for model training objectives, architecture, and reasoning methods, but also leave room for Chinese embodied teams to contribute foundational innovation.

Even before the industry debated whether "large models can do everything," many Chinese embodied teams had already begun to give their own answers. Ant Lingbo's "embodied native" exploration is one of them.

In LingBot-VA 2.0 released in July this year, the Ant Lingbo team chose to pretrain a causal video-action model from scratch, allowing the model to learn the connection between environmental changes and actions, and to use the latest observations to correct predictions during execution.

The so-called "embodied native" is not hard to understand: from the very beginning of design, the model revolves around the perception, prediction, and action required for robots to complete tasks in the real physical world. More importantly, this also means that Chinese embodied brain companies have begun to independently build a native technology route.

Does such a technical route have value? Besides Lingbo's own tests, it also depends on which parts external teams choose after the model is open-sourced, and what they build with them.

The first type of use is taking the work of a Chinese team and continuing to do research with it.

Researchers at Stanford University's Vision and Learning Lab wanted to know: with so many world action models today, but each using different model backbones, training methods, and data, how can one determine which design actually brings the capability improvement?

The team therefore wanted to build a composable research framework so that different video-action interaction methods could be compared on a single "unified exam," and this was the starting point of the OpenWAM project released in October. At the implementation level, they directly reused LingBot-VA's Transformer architecture, text encoder, and tokenizer as the reference implementation for the entire framework.

Researchers from the National University of Singapore targeted another difficult problem in embodied intelligence: when a world model generates long-horizon interactions, how can it both "remember" previously generated scenes and avoid running slower and slower as historical information accumulates?

To this end, they proposed a spatial memory method called Velox that makes world models run faster and faster. The core idea is not to delete world memory, but to let the current frame read only the truly relevant historical regions. The key to achieving this is not training a new world model, but using LingBot-Fast-Cam as the base, leaving the pretrained generator unchanged, and adding a spatial information filtering mechanism only at the inference stage, so that the model reads historical content truly relevant to the current frame based on camera viewpoint and depth information.

Ultimately, across 158 WBench navigation cases, its API-free average score was only 0.15 points lower than the same LingBot-Fast-Cam baseline, which shows that Velox, while basically maintaining generation quality, alleviates the problem of long-sequence generation becoming slower and slower as it goes on.

The second type of use is turning the model into a tool that more people can operate.

For ordinary developers, no matter how advanced a model's capabilities are, if it remains only in papers or top-tier equipment, it is still hard to actually use. In the end, the value of a model must ultimately be tested by the countless developers who use and modify it.

Developer RealRebelAI integrated LingBot-World v2 into ComfyUI, which creators are familiar with. Users can drag in an image and arrange a camera movement path to generate subsequent frames. To make this workflow run on a computer, he quantized the model and loaded it layer by layer, and ran tests on an RTX 3070 8GB; the world model began to enter ordinary developers' workflows from the lab.

Developer Don Allen III built a Mac desktop tool based on LingBot-Map. Users drag a video into the window and can start reconstruction locally on Apple Silicon and view progress. After completion, they can also rotate and look down on the 3D point cloud, review the camera trajectory, and save the results. The spatial understanding capability originally hidden in code and weights has become a desktop application that can be operated intuitively and is fully visualized.

The practices of frontline community developers confirm that what Chinese embodied model companies have lowered is not only the threshold for using models, but also the cost for latercomers to continue innovating. The code, experience, and workflows left by ordinary developers will be reused and improved by the next developer, and then passed on to more people. Individual attempts thus form a diffusion effect, and the open-source ecosystem is activated through relay after relay.

The third type of use is connecting the model into engineering teams' existing toolchains.

Whether an open-source embodied brain can create practical value also depends on whether it can enter the toolchains of the industrial ecosystem.

NVIDIA added support for LingBot-World in its open-source inference library FlashDreams, so developers can use this toolkit to run a world model with controllable camera motion. AWS Labs, in its own open-source world model deployment project, set LingBot-World-Fast as the default EC2 deployment path and flagship entry point. Following the steps given by the project, developers can configure compute in the cloud, load weights, and call the video generation service.

Some influence has also remained in papers.

NVIDIA's SANA-WM team used LingBot-World as a comparison for generated image quality. Generalist, when discussing how robots can quickly learn new tasks through one or a few demonstrations, included LingBot-VA 2.0 as a relevant research reference; Dyna Robotics also cited LingBot-VA when comparing world action models with VLA.

Why a paper is worth comparing, why a piece of code is worth keeping, and why a model is worth special deployment adaptation—overseas peers have each given their own answers.

Looking back at each round of China's technological development and breakthrough, the answer is not only cost and speed, but also the connection, reuse, and co-creation brought by openness. This was true for the internet and for large models, and it is the same for embodied intelligence.

Understanding this, the annual World Internet Conference in Wuzhen is a vivid footnote to this thread.

Last year, when the "Six Little Dragons of Hangzhou" appeared together for the first time, Wang Xingxing recalled the story of attending the conference for the first time in 2017; what he brought was the company's first robot dog. DeepSeek senior researcher Chen Deli, who shared the stage, talked about how open-source innovation promotes technological inclusiveness and allows the community to give back to R&D.

This year, Ant Lingbo, after intensive open-sourcing, also appeared in Wuzhen with the full-stack LingBot 2.0 technology and open-source practices, and brought a real retail pharmacy into the exhibition hall.

These details together outline a shift in Chinese technology: from bringing products to the world to beginning to try to open innovative capabilities to the world.

Embodied intelligence is of course a long, steep, and snowy race, and going global is still in its early stages. Technology generalization, data compliance, local operations and maintenance, and commercialization are all hurdles that have not yet been crossed, but two paths for going global have already begun to become clear.

One is asset-heavy product and supply chain going global, such as Unitree, AgiBot, and UBTECH, which build factories or R&D centers overseas and export the robots themselves. On the next leg, they need to move from "product delivery" to "local operations," completing commercial validation through products, channels, and customer trust.

The other is technology platform ecosystem going global, represented by Ant Lingbo, for example, entering overseas teams' daily R&D workflows by opening models, toolchains, and interfaces. On the next leg, they need to continuously iterate models, adhere to an open-source strategy, and accelerate the establishment of an innovation ecosystem for foundation models.

The former enters customers and scenarios faster, while the latter expands technology reuse faster, but together they point to a future for Chinese embodied intelligence going global: only by truly solving problems, improving efficiency, and exercising innovation in frontier exploration and real scenarios can one seize the positional advantage in embodied intelligence going global.

This "great voyage" of going global has only just set sail, but Chinese embodied intelligence has already taken a key first step.

This article comes from the WeChat public account "Phoenix Technology"; author: Phoenix Technology; published by 36Kr with authorization.

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.

English text is being prepared.