# Embodied AI Unveiled AIR, a Wheeled Factory Humanoid. Tendon Arms Are the Safety Pitch; Named OEMs Are Not. > Markdown twin of https://roboterminal.app/posts/embodied-ai-air-humanoid · https://roboterminal.app/posts/embodied-ai-air-humanoid.md **Published:** 2026-11-06 Swiss EPFL spin-out Embodied AI debuted AIR at Humanoid Europe 2026 on November 6, a Munich-dateline Business Wire release said. The wheeled humanoid has tendon-driven arms meant to work beside people on factory floors; the company says major OEM contracts and manufacturer tests are underway, CE marking is planned for 2027, and the release does not name customers or disclose round size. Swiss EPFL spin-out Embodied AI debuted AIR at Humanoid Europe 2026: a wheeled, tendon-driven humanoid meant to work beside people on factory floors, learn tasks from operator demos in about a month, and chase CE marking in 2027. The company says it already has OEM contracts and unnamed manufacturer tests, without publishing customers, prices, or a funding total. **Tags:** humanoids, manufacturing, factory, safety, europe, epfl, deployments, teleoperation Swiss robotics company Embodied AI unveiled AIR on November 6 at Humanoid Europe 2026, a [Business Wire](https://www.businesswire.com/news/home/20261106661068/en/Embodied-AI-Unveils-AIR-a-Physically-Safe-Humanoid-Robot-That-Learns-on-the-Job) release dated Munich said. AIR is a wheeled humanoid with tendon-driven arms, built to stand next to people on factory floors rather than behind cages. The company, an EPFL spin-out headquartered in Lausanne with an office in Rome, says the machine went from research to an industrial humanoid in nine months, that it has signed contracts with major OEMs, and that unnamed global manufacturers are already testing it. CE marking is planned for 2027. No customer names, unit prices, or round size appear in the release. Our view: **AIR is a bet that factory buyers will pay for a humanoid that yields on contact and learns on their line, not one that walks through a keynote.** Most 2026 humanoid coverage still sells bipedal demos. Embodied AI puts the motors in the torso, lightens the arms with tendons, rolls on wheels because factory floors are flat, and starts every job under a human operator until its stack can take over. That is closer to collaborative industrial robotics than to a warehouse biped race. The open question is whether named OEMs and shift metrics show up before the CE date, or whether "tested with global manufacturers" stays a press-release phrase. ## What AIR is supposed to be The release frames AIR for repetitive manual work that changes too often for conventional robots, so people keep the jobs that still need their judgment. Embodied AI's [site](https://embodiedai.ch/) lists picking, placing, plugging, and tool integration among the assembly tasks it wants the platform to handle. The company also sells two-arm industrial cells that run the same AI stack for jobs that do not need a humanoid form factor. Hardware is the safety claim. Motors, gearboxes, and electronics sit in the torso. Tendons move light arms that "give way on contact," the company says, so compliance comes from mass and mechanism, not only from software limits. Wheels replace legs. The design targets ISO/TS 15066:2016, the collaborative-robot technical specification, with CE marking planned for 2027 — planned, not granted. ## How it is supposed to learn AIR starts every job under human control so the first shift can still produce, according to the release. Each run is recorded. Embodied AI's stack learns from those demonstrations until the robot can run the task itself, "typically within a month," with the operator staying for cases the model has not seen. What one robot learns, the next can reuse, the company says. That stack already has a prior proof point the company is willing to put in a footnote: in a cable-handling trial for a leading consumer electronics manufacturer, five funded startups were tested and Embodied AI says its two-arm cell hit 98% reliability against a 90% goal. The customer is unnamed. Treat the number as company-reported pilot data, not an independent audit. CEO Francesco Stella's line in the release draws a hard boundary between chatbots and robots: digital hallucination can be reviewed before anyone acts; a wrong physical move can injure people or wreck a line. "Safety and reliability are non-negotiable requirements of physical AI to enter the real world at scale," he said. ## Who built it, and who paid Stella, 28, co-founded Embodied AI with CTO Kai Junge and COO Max Polzin. The AIR arm design grew out of Stella's PhD at EPFL; he also researched at MIT CSAIL. Junge and Polzin are robotics PhDs with backgrounds at Cambridge, ETH Zurich, and EPFL, the release says. Faber led the company's first funding round. TechShop, Look AI Ventures, Plug and Play, and Vento also participated. The release does not disclose how much was raised. Embodied AI says it reached OEM contracts and manufacturer tests with "just a fraction of the spending needed by other robotics companies," also without a figure. Market context in the release cites third-party studies rather than Embodied AI's own bookings: Deloitte and The Manufacturing Institute's estimate that as many as 1.9 million US manufacturing jobs could go unfilled by 2033; Manufacturing Leadership Council figures that about 22% of manufacturers plan to use physical AI within two years, up from 9%; and Deloitte's view that more than 80% of factory task hours will stay human-driven. Those are industry surveys used as framing, not measures of AIR demand. ## What is still unproven No OEM is named. No factory site, robot count, price, cycle time, or uptime figure for AIR itself appears in the unveil. The nine-month build claim and the "fraction of spending" claim are company statements. The 98% cable-handling result belongs to the earlier two-arm cell and a nameless customer. CE marking in 2027 is a plan. Until Embodied AI publishes a named manufacturer, a task list, and a reliability number for AIR on that line, the product is a well-argued safety architecture plus a learning story, not a verified deployment. ## Objections **"Wheels mean it is not a real humanoid."** Fair as taxonomy. Irrelevant to a plant with flat concrete and no stairs. The buyer question is dexterous work next to people, not stair climbing. **"ISO/TS 15066 and CE 2027 are table stakes."** They are. Naming the standard is not the same as shipping a CE-marked fleet. Watch the 2027 date and whether AIR stays in the collaborative envelope once payloads and speeds rise. **"Learning in a month is a demo script."** Maybe. The transferable-demo loop is the commercial claim. It needs a named line and a before/after scrap or cycle metric, not another unlock video. ## Three things to check 1. **Name one OEM or factory.** "Global manufacturers" and "major OEMs" are not references. A plant, a task, and a robot count would be. 2. **AIR-specific reliability, not only the prior cell.** Publish the same style of target-vs-result for AIR that the company published for cable handling on the two-arm cell. 3. **CE timeline and what changes before it.** If the 2027 mark slips, or if production hardware diverges from the unveil platform, the safety pitch needs a new sentence. ## Sources - [Business Wire](https://www.businesswire.com/news/home/20261106661068/en/Embodied-AI-Unveils-AIR-a-Physically-Safe-Humanoid-Robot-That-Learns-on-the-Job) - [Embodied AI](https://embodiedai.ch/) - [Deloitte press release](https://www2.deloitte.com/us/en/pages/about-deloitte/articles/press-releases/us-manufacturing-could-need-new-employees-by-2033.html) ## Citation ``` "{quote}" Source: Roboterminal · {url} ``` --- Source: Roboterminal · https://roboterminal.app/posts/embodied-ai-air-humanoid.md