Xiaomi and Baidu invest in embodied AI startup Chuguang Yueqian

Shanghai Chuguang Yueqian Intelligent Technology, an embodied AI startup founded five months ago, has taken investment from funds tied to Xiaomi and Baidu, corporate records show — a rare joint bet by two Chinese tech rivals.

The registered capital increase is tiny: from 500,000 yuan to 600,000 yuan (about $70,000 to $85,000). The shareholder list is not. Xiaomi's Hanxing Venture Capital and Baidu's Sanya Baichuan Zhixin fund now sit alongside founder and employee holding vehicles, a sign the two groups are making early defensive bets on the algorithms that connect AI models to physical machines.

A small filing with heavyweight backers

Filings tracked by Tianyancha, a corporate records database, show Hanxing Venture Capital and Sanya Baichuan Zhixin Private Equity Investment Fund Partnership among the new shareholders. Zhou Zijian stepped down as legal representative and was replaced by Zhang Ying.

The company was registered in May 2026. Its approved business scope covers smart robot sales, smart robot research and development, service and consumer robot sales, general-purpose AI application systems, and AI基础 resources and technology platforms.

After the change, Chuguang Yueqian is held by Shanghai Xihe Yizhi and Shanghai Chuguang Chengyao — employee and founder partnership vehicles — together with the two corporate funds.

A self-driving veteran and a world-model researcher

Zhang Ying, the new legal representative, was a core executive at a leading autonomous driving company. The team worked through the shift from reliance on high-definition maps to mapless perception and one-stage end-to-end driving systems, including mass production delivery.

Zhan Fangneng, the team's other central figure on the academic side, has focused on world models, physical AI and 3D generative representation.

That pairing — an algorithm veteran who has shipped at scale plus a frontier world-model researcher — is scarce in a robotics sector short of engineers who can bridge both.

The physical data bottleneck

Boston Dynamics, Tesla Optimus and domestic humanoid robot teams face the same problem. End-to-end models that flourish in digital worlds do not transfer to physical bodies.

Language models can learn endlessly from hundreds of billions of words online. Robots must learn to grasp objects in kitchens, mines and supermarkets, largely through teleoperation, which is costly, slow and cannot cover every edge case.

Chuguang Yueqian's answer is an architecture it calls Physical RSI, or robotic closed-loop self-evolution. The team is trying to use a pluggable, modular world-model framework and an automated research engine so the system can generate physical hypotheses in simulation, run experiments and correct motion plans in real time.

That kind of self-evolution, where algorithms improve algorithms, is the key to closing the gap between digital brains and physical bodies.

What the two giants want

Xiaomi's strategy spans phones, cars and home devices, and it is pushing deeper into manufacturing and consumer hardware. It runs a highly automated Xiaomi Auto plant, is iterating its CyberOne humanoid robot and ships robot vacuums and robot dogs in volume.

Its strengths are supply-chain control and industrial design. A general embodied "brain" that works across different hardware bodies is not. Investing in Chuguang Yueqian gives Xiaomi early access to pluggable, self-evolving algorithm modules for that hardware portfolio.

Baidu's problem is different. ERNIE and Apollo are strong on screens and roads, but the company lacks a foothold and a physical data flywheel in unstructured indoor manipulation and service robots.

Through the Baichuan Zhixin Fund, Baidu can track frontier embodied AI architectures without carrying the cost of heavy hardware research.

A steep engineering climb

The startup, valued in the tens of millions of yuan, still faces a difficult path from simulation to reality. Running millions of motion predictions and collisions in a virtual engine is one thing.

In the real world, reflective surfaces, objects that deform under pressure, gear wear in reducers and sensor drift can destabilise a robotic arm. Keeping world-model hypotheses running on real hardware at millisecond latency will test the team's engineering to its limits.

Then there is the pull of two backers with different priorities. Xiaomi prizes cost efficiency and scale, and will want technology adapted quickly to its robots and factory lines. Baidu cares more about data accumulation and algorithm barriers in multimodal models.

A seed-stage team must keep its independence while balancing both, and avoid becoming an outsourced testing ground for either.

The 600,000 yuan registered capital and the presence of two giant funds send a clear signal to the AI and robotics market: money in embodied AI is shifting from assembling humanoid shells to the underlying physical algorithm stack. Chuguang Yueqian has traded a clever technical architecture for a seat at the table with both giants. Whether it can turn self-evolution theory into industrial delivery is the question it must now answer.