After 700,000 NIO Cars, a Tsinghua Team Wants to Teach Robots to Assemble Servers

肖中阳 (Xiao Zhongyang) spent the last few years inside NIO building the world-model software that today runs in more than 700,000 of the carmaker's electric vehicles. Now at 日冕开物 (RIMBOT), he wants to find out whether the same approach can teach a robot arm to plug a component into a server chassis. His first customer, Dongguan-based server manufacturer 远图未来, will be where the question gets answered.

日冕开物 is building a foundation model for robots, in the same way large language models underpin chatbots and coding assistants. The work is in embodied AI, the field of building AI systems that control physical robots rather than just generating text or images, and the foundation is a "world model" trained to predict how the physical environment will change second by second and to use those predictions to plan a robot's next move.

Every founder comes out of Tsinghua's ADAS program, and the team's résumé reads as a roll call of recent Chinese autonomous-driving history. Per a 36氪 exclusive that first reported the company this week, 肖中阳 was the world-model delivery lead at NIO and says a NIO-era world model he led was deployed across more than 700,000 of the carmaker's cars. The base-model lead, 钟元鑫, was part of Huawei's "genius youth" recruitment program and is described as having led the next-generation ADAS world-model base design and its mass production. The post-training and delivery lead, 王云龙, is cited with prior work on NIO's "Super Spark" autonomous-driving effort and at Agibot on large-model and real-machine reinforcement learning. The market lead, 戴亚奇, came from 武岳峰科创, a semiconductor-focused investor.

The technical centerpiece is LaMPA, a world model the company built and is now selling as a full-stack solution. The architecture is organised around three representations of a scene: the environment, the robot itself, and the robot's accumulated experience, all three feeding a shared latent space, the compressed numerical layer where the model stores its understanding of the world. On top sits a Block Diffusion base, a generation method that produces a sequence of predicted next states in chunks rather than one at a time. For post-training, the company has distilled a "World Reward Model" meant to replace the hand-tuned reward functions robotics teams normally write by hand for every new task. The company has also open-sourced PalmDex, which it calls the world's first multi-scene dual-hand full-palm tactile manipulation dataset.

日冕开物 explicitly frames LaMPA as an extension of JEPA (Joint Embedding Predictive Architecture), Yann LeCun's framework at Meta for learning abstract representations of the world rather than predicting each pixel. The company's argument is that JEPA leaves two questions open: how to build the latent space, and which base architecture to use to predict the next state. LaMPA, in the company's telling, is its answer to both. None of this is third-party-validated; nothing in the public source set independently shows that LaMPA's triple-representation approach outperforms any rival world model on a standard robot task.

The deployment test is at 远图未来, a Dongguan-registered server manufacturer that holds its own assembly-line IP. The partnership is scoped first at high-precision server assembly, with the stated intent to extend into front-end and back-end process steps across the line. Server assembly involves tight-tolerance insertion, component verification, and frequent product changeovers, all of which break a robot pre-programmed for a single SKU. A world model that can predict the consequences of a slightly misaligned connector might, in principle, let a robot recover instead of fail. That is the question the pilot is meant to answer.

Two seed rounds are closed, totalling "hundreds of millions of RMB" (an amount that, at recent exchange rates, is equivalent to roughly $40M-$110M USD), with a further round simultaneously closing, per the same 36氪 硬氪报告. Investors named include 鼎峰科创, 远图未来 itself, 百度风投, 沃衍资本, 武岳峰科创, and 万林国际. Aggregator pickups from Sohu and dtm.com.cn re-report the same facts with no added mechanism. The company's official site at rimbot.com is a brand statement only.

The 36氪 piece says 日冕开物 was founded in March 2026, while aggregator pickups citing corporate registry records point to an October 2025 entity-registration date for the legal entity 北京日冕机器人有限公司; both dates are in circulation. 王云龙's prior affiliations at NIO and Agibot are described in the source, but his current role is at 日冕开物, not at either previous employer. The 700,000-vehicle figure is a founder statement carried in the 36kr article, not a Type0-verified fleet count, and LaMPA's "first" and "extension of JEPA" framings are company positioning until a third-party benchmark or peer-reviewed result lands.

By late 2026, if LaMPA can move from a single server-assembly task at 远图未来 to a second production line at a different customer without task-specific reward engineering, the ADAS-to-robots transfer starts to look real. If not, the team will have a strong demo and a much harder fundraising story. The 36氪 first-exclusivity piece is dated this month. The next concrete milestone is whatever the company can show on a 远图未来 line in 2026.