Quant Group tests its Physical AI model in a working restaurant kitchen

Four tasks in a live commercial kitchen

The company says each demonstration was validated through repeated testing and deployed under real, dynamically changing kitchen conditions. The four most recent tasks follow earlier mid-2026 demos that included sandwich preparation, shopping-bag sorting, steak seasoning and coordinated dual-arm milk-tea preparation. According to Quant Group, its embodied-AI capabilities have moved from dexterous manipulation to handling a combination of materials.

• Hand-pounded lemon tea: a closed-loop precision task. One hand steadies the cup while the other pounds repeatedly. The company says an RSI self-evolving loop records trajectories, identifies deviations and adjusts rhythm during preparation.
• Flaky flatbread: a sequence of twenty-two consecutive actions. The robot must remember which step it has reached and which tool comes next. System 2 handles global planning and exceptions, while System 1 executes actions in real time.
• Stir-fried shredded potatoes: a test of Chinese wok cooking. The shreds slip easily, the materials shift in the wok, and dual-arm coordination demands millisecond-level timing, with one hand working the wok and the other holding the spatula.
• Egg cake: a test of generalization across materials, including high-viscosity batter, unevenly sized crushed peanut and a soft cake that needs continuous adjustment of contact force.

A software-first strategy

Quant Group says it does not manufacture robot hardware. It is developing an open physical-world foundation model meant to generalize across scenarios and robot platforms. The architecture separates software from hardware, so the same underlying model can run on different hardware platforms. The company describes the central test as whether one model can transfer learned skills across different tasks, materials and types of physical disturbance.

The company argues that selling hardware or project-based solutions scales project by project, while a foundation model that is trained once and deployed across scenarios has marginal costs approaching zero. It says the next step after the kitchen trials is cross-scenario reuse, extending these capabilities beyond food service to a broader range of hardware platforms.

Quant Group also argues that its moat is data. Language data can be harvested from the internet in vast quantities, it says, whereas physical-interaction data can only be generated by operating in the real world. That is why it tests in a restaurant kitchen rather than a laboratory, since such unstructured conditions cannot be fully reproduced in simulation.

Funding and market context

The physical AI work is funded by the company's consumer e-commerce business. In 2025 Quant Group reported revenue of RMB 1.035 billion and net profit of RMB 195 million, up 32.6% year-over-year. Gross merchandise value on its Yangxiaomie platform exceeded RMB 10 billion. The company says 30 percent of the proceeds from its July placing went to research and development of the Physical AI foundation model, and a further 40 percent to the commercialization of intelligent hardware.

The release cites Frost & Sullivan, which forecasts a compound annual growth rate of 47.2 percent for Physical AI, with the global market for related industries potentially reaching US$3.25 trillion by 2040. It also notes a report that Field AI, which does not build its own robots, has been valued in the region of US$10 billion.

What it means

The demonstrations are the company's own, and the release gives no independent benchmarks or success rates. For the market, the message is that Quant Group wants to be judged as a provider of a reusable capability layer rather than a hardware maker. According to the company, its price-to-earnings ratio is around 19 to 20 times and its price-to-sales ratio around 2 times, which it reads as a sign the market has not yet valued it as a physical AI company. Whether the model transfers beyond kitchens remains the open question the company says it will address next.