--- title: "Megvii repositions from face recognition to running whole physical spaces with agents" date: 2026-10-07 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-megvii-physical-ai-agents language: en --- # Megvii repositions from face recognition to running whole physical spaces with agents > The computer-vision pioneer is reframing itself around "Intelligent Agents" for physical space — selling autonomous operation of buildings, factories and city zones rather than standalone algorithms. Walk into a large campus or a logistics warehouse and imagine the lights, cameras, gates and robots not as separate gadgets but as one coordinated crew that perceives, remembers, decides and acts on its own. That is the picture Megvii (旷视科技) is now selling. Fifteen years after it became famous for Face++ facial recognition, the Beijing AI company is telling customers it no longer wants to ship "algorithm parts" — it wants to run the physical space itself. ## From "seeing" to "operating" Megvii made its name in computer vision (计算机视觉) with Face++, but its official self-description has shifted to an "AI company focused on physical space operations (物理空间运营)." The mechanism behind that slogan is a layer it calls **POA OS** — Physical Operation Agent OS, the operating layer of what it now markets as an AIoT (人工智能物联网) system spanning cloud, edge and devices. In Megvii's framing, POA OS is the "deterministic just-in-time middle layer" of the physical world: it standardizes how algorithm capabilities and hardware devices are adapted and scheduled, so that agents can perceive, remember, decide, execute and evolve inside real environments. The distinction Megvii draws is deliberately different from most robotics and model vendors. It says it does not sell robot bodies, does not train a general-purpose large model, and does not merely license algorithms. What it provides is "space operation capability" — letting an office tower, a factory floor or an urban district be autonomously sensed and run by agents. That is a services-and-stack play, not a hardware play. ## The stack underneath Three pillars make the pitch credible: - **Brain++**, Megvii's self-developed AI productivity platform, spans data management, training and deployment so customers do not rebuild pipelines from scratch. - **Megvii Taiyi (太乙)**, the company's large model (大模型) "brain" for the physical world, sits at the center of a perception-decision-action closed loop. - An AIoT product system — operating systems like Megvii Pangu (盘古) for large enterprise groups and Hongtu (鸿图) for smaller firms, the Jiuxiao (九霄) cloud agent platform, and edge devices such as Huanfang, Mofang, Shenxing and Minguan — turns the brain into deployable products across five target industries: city transportation, commercial workplace, energy and urban governance. ## A research signal, not just marketing Megvii's pivot is backed by peer-reviewed work, which matters for a company whose value is technical credibility. On its official news page it lists a multimodal in-context learning paper accepted at **ICML 2026** (presented in Seoul). Independent tech coverage adds that a spatial-perception method called **E-ViC**, co-developed with the Beijing Humanoid Robot Innovation Center, was accepted at **ACL 2026**, with reported gains on spatial-understanding benchmarks and notably strong results on fine-positioning tasks. We did not independently re-run those benchmarks, but the acceptances themselves are public and verifiable. The throughline: agents in physical space must understand objects, rooms, behavior and feedback, then act. Megvii argues its edge is doing all four — algorithm, hardware, system integration and industry know-how — which most competitors only do partially. For global readers this is the more interesting bet: rather than race the United States on frontier model size, a cohort of Chinese vision firms is competing on who can make physical spaces autonomously operable first, where the buyer is a factory manager or a city district, not a developer. ## Why "deterministic" is the keyword Megvii stresses that physical-world AI cannot fail the way chatbot AI can. A wrong recommendation online annoys a user; a wrong action in a factory or on a roadway risks injury and property. So POA OS hard-codes the rules, workflows and safety boundaries of physical space as system-level constraints, letting agents act autonomously only inside that frame. That is Megvii's answer to the central Physical AI (物理AI) question: how to balance an agent's autonomy with a system's certainty. ## Honest limitations Several claims here are company-stated, not independently audited. Megvii cites "nearly 1,700 AI-related patents," "200+ top-conference papers" and "59 international competition championships" on its own about page; treat these as self-reported. The report that Megvii "expects full profitability in 2026" comes from tech-media analysis, not a filed financial statement, and should be read as an outlook, not a result. The ACL 2026 E-ViC gain figures (e.g. an average 10.1% lift) appear only in secondary coverage we did not cross-check against the paper. Megvii remains under intense competitive pressure from Huawei, SenseTime (商汤) and others in smart-city and enterprise vision, and its past reliance on hardware sales and government projects has weighed on margins — a structural risk the agent pivot is meant to fix but has not yet proven it can. ## What readers can do now - Facility and campus operators evaluating "smart building" upgrades should ask vendors whether the offering is algorithm licensing or full-space operation — Megvii'sPOA OS framing is the stricter, more valuable claim, and the harder one to deliver. - Investors tracking Chinese CV (computer vision) firms should watch whether the agent model converts pilots into recurring operation contracts, not just one-off hardware deals. - Technical readers can read the ICML 2026 and ACL 2026 papers directly to judge the spatial-understanding claims rather than trusting summaries. - Buyers in regulated sectors (energy, transit, government) should probe the safety-constraint design explicitly — "deterministic physical rules" is a posture worth testing with edge cases. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-megvii-physical-ai-agents Free to quote with attribution and a link to the original.