XIRRA raises hundreds of millions of yuan to build a physics model for robots

Why it matters

XIRRA's $500 million company-stated valuation rests on a bet that Li Yin can turn large-model training experience into robot intelligence that works across tasks and machines. The financing funds that test; it does not settle it.

Li Yin's XIRRA has completed seed and angel rounds totaling several hundred million yuan, according to financing details reported by Jiemian News on September 23rd. Founded in July by the former Huawei Cloud large-model CTO, XIRRA is based in Shenzhen and says it is valued at $500 million. Neither the precise amount raised nor the basis for that valuation was given.

Li spent his previous job working on Huawei's Pangu models and embodied-AI projects for business customers. In Huakong Fund's account of a forum it co-hosted, Li identified three obstacles to selling useful robot intelligence: poor generalization across tasks, difficulty executing long sequences of actions, and weak learning from the results. XIRRA is his attempt to solve those problems at the model layer rather than build another robot.

A model bet backed by hardware money

Dunhong Asset led the financing, with Huakong Fund, Sanhua Holdings, Yinxinggu Capital, Zeran Capital, Benjian Fund, Shanghai Angel Association and Biaopu Investment named as participants. The list puts a manufacturer alongside financial investors: Sanhua makes thermal-management and control components and has identified robots as an area of expansion. Its investment connects XIRRA's model ambitions to an industrial sector that will need to test models on physical machines. The investment itself does not establish a development or supply agreement.

The September 23rd disclosure supplies a cumulative figure for two rounds, rather than the size or closing date of each. An angel financing involving Dunhong, Huakong, Zeran, Benjian and Yinxinggu had already appeared in an August 27th report based on investment-database information, with no amount attached. The new announcement therefore gives investors a larger financing total and a company-stated valuation for a process already partly visible in August; it does not mean both rounds closed this week.

According to 36Kr's account of the financing, XIRRA plans to spend the proceeds on large-scale physical-AI pretraining, datasets from real-world interactions, research hiring and testing commercial uses. Those are distinct expenses. Training a model is only part of Li's proposition; obtaining interaction data and showing that the resulting system works outside a controlled demonstration will determine whether XIRRA can sell a reusable foundation model.

Li wants the layer beneath the robot

Li calls the planned system a large physical model, or LPM. In Huakong Fund's account of his presentation, he described an approach that learns both how a physical scene changes after an action and which action to take next. He tied the design to the Bellman equation, a framework for breaking long-term decisions into steps. XIRRA's proposed customers would build specialized systems on top of that base, rather than buy a robot body from XIRRA.

That distinction puts the burden on transfer. A model trained on one collection of scenes or machines must help with unfamiliar tasks, hardware and conditions for the foundation-model strategy to pay off. Li's stated concern about generalization is commercially specific: a system that needs extensive retraining for each deployment remains a custom robotics project, however broadly its underlying model is described.

XIRRA also says Zhang Hanwang, whom it identifies as Huawei's multimodal-AI chief scientist, will become co-founder and chief scientist. Zhang is an associate professor at Nanyang Technological University whose research spans computer vision and causal inference. NTU records a 2024 young-investigator research award for his work. Jiemian describes his XIRRA appointment in the future tense, so his contribution to the venture is best understood as a planned addition to Li's founding effort.

The evidence has narrower bounds than the ambition

WorldArena 2.0 evaluates world models on video quality as well as their usefulness for training policies and planning actions on simulated or real robots. The benchmark materials cited here do not independently verify a ranking for XIRRA. A video-quality result alone would not establish performance across the benchmark's robot tasks. XIRRA's financing disclosure does not identify a robot deployment or commercial contract.

There are established model-first competitors. Skild AI describes its approach as a general-purpose brain intended to work across robot types, while publishing examples of the applications on which it is trying to prove that claim. XIRRA is making its case much earlier in that cycle: investors have backed Li's experience and a research program before the reported funding is tied to a comparable customer deployment. Huakong's account of the forum makes the investment logic unusually plain. Its chief investment officer said early-stage assessments focus on the people, their large-model training experience and how quickly they improve the algorithms; data and real-world use become the tests as companies mature.