Chinese automakers are betting that robots need their own artificial intelligence foundation models, not hand-me-downs from self-driving car technology. Li Auto has hired a veteran world-model researcher to lead development of a specialized "brain" for its robots, marking a deliberate departure from competitors like Tesla and Xpeng who plan to share AI models across vehicles and robots.

What's the Difference Between a Robot Brain and a Self-Driving Brain?

The distinction may sound technical, but it reflects a fundamental insight about how robots learn. Li Auto's new hire, identified as Liu Yu, brings eight years of autonomous-driving experience and was among the earliest researchers in China to explore world models, a type of AI that learns to predict how the physical world works. Rather than repurposing the company's existing autonomous-driving foundation models, Liu's team will build a separate model specifically trained on robot data.

This approach contrasts sharply with industry giants. Tesla has stated its neural world simulator can power both Full Self-Driving (FSD) and its Optimus humanoid robot. Xpeng similarly plans to use its second-generation Vision-Language-Action (VLA) model across vehicles, robots, and flying cars. The VLA model is a type of AI that can understand images, language instructions, and predict physical actions robots should take.

Liu's philosophy is that the reusable knowledge from autonomous driving lies not in the data or models themselves, but in the underlying training infrastructure and data-handling techniques. His team plans to first train the robot model on general-purpose spatial and world datasets from the internet, then layer in robot-specific data to improve how well the model generalizes across different robot types, tasks, and environments.

How Is Li Auto Building Its Robot Foundation Model?

• Dedicated Team Structure: Liu oversees the "embodied behavior unit" within Li Auto's foundation-model team, with headcount deliberately kept below 50 people to prioritize quality over scale.
• Multi-Model Strategy: Li Auto is developing three separate foundation models: MindGPT 3.1 for intelligent cockpits, MindVLA-o1 for autonomous driving, and a new dedicated model for robots.
• Data-First Approach: The robot model will be trained first on internet-sourced spatial datasets, then refined with robot-specific data to enable the model to work across different embodiments and scenarios.

The hiring signals Li Auto's acceleration of its robotics ambitions. The company launched a robotics project code-named "Nexus" in late 2024 and plans to unveil its first two-wheeled robot this year for factory manufacturing applications, with a bipedal robot also in development. Liu's recruitment also helps Li Auto replenish its AI talent after losing 13 key executives across autonomous driving, chips, and robotics over the past year, including former autonomous-driving chief Lang Xianpeng, who left to found an embodied AI startup.

Why Does This Matter Now?

The shift reflects a broader industry realization: as motor and reducer component supply chains mature, the competitive battleground is moving away from robot hardware toward foundation models and execution capabilities. This timing aligns with a surge in robotics funding. Xpeng announced approximately $900 million in financing commitments for its robotics subsidiary Dogotix, implying a post-money valuation of around $6.3 billion.

Meanwhile, the global robotics sector is seeing explosive valuations. Generalist, a startup founded by former Google DeepMind and Boston Dynamics researchers, reached a $3 billion valuation after raising nearly $200 million in additional capital led by 8VC. The company is developing an AI foundation model that works across various robot types and claims its Gen 1.5 model enables robots to master new tasks from video demonstrations as short as 3 to 12 seconds long.

Generalist faces competition from Physical Intelligence, valued at $11 billion, and SoftBank-backed Skild AI, valued at $14 billion. The funding surge reflects investor confidence that robotics may soon reach its own "ChatGPT moment," where robots can perform general tasks without being explicitly trained for each one. However, some venture capitalists caution that a truly general robotics model may still be years away, since robots cannot be trained on the entirety of internet data the way large language models can.

On the commercial front, AGIBOT, a Chinese robotics company founded in 2023, ranked number one globally in humanoid robot shipments in the first half of 2026 with a market share exceeding 43 percent. The company's founder and CEO, Deng Taihua, was named to TIME's 2026 TIME100 AI list, recognizing influential figures in artificial intelligence. AGIBOT's architecture integrates locomotion intelligence, interaction intelligence, and manipulation intelligence into a unified embodied system, and the company announced that its 15,000th robot had rolled off the production line in June 2026.

"Our goal is to use robots to solve real-world problems and advance practical technology deployment in a grounded and pragmatic way," said Deng Taihua, Founder and CEO of AGIBOT.

Deng Taihua, Founder and CEO of AGIBOT

Li Auto's decision to build a dedicated robot foundation model reflects a strategic bet that embodied AI requires different training approaches than autonomous vehicles. While self-driving cars operate in relatively constrained environments with predictable road rules, robots must navigate diverse physical spaces, manipulate objects with varying properties, and adapt to unpredictable human interactions. By investing in a specialized model rather than reusing existing technology, Li Auto is positioning itself for a market where robot "brains" may become as valuable as the mechanical bodies they control.