Embodied AI introduces a factory robot built to work without safety cages

Embodied AI is promising a factory robot that can work beside people without a safety cage and learn its job within about a month. The Swiss company introduced AIR in Munich on November 6, but its cited factory results come from an earlier two-arm industrial cell—not a disclosed production record for the new mobile robot.

Light arms, heavy components in the torso

Presented at Humanoid Europe 2026, AIR has a torso, two arms and a wheeled base. It is aimed at repetitive manual manufacturing jobs that change too often for conventional automation. Despite the humanoid label, it does not walk: Embodied AI says wheels offer greater stability on flat factory floors.

Its central hardware choice is to move the heavy parts out of the arms. Motors, gearboxes and electronics sit inside the torso, while cables transmit movement to the limbs. This tendon-driven arrangement reduces the mass of the moving arms. The company says they give way when they make contact with a person or obstacle.

That is the basis of AIR’s safety pitch: hardware that yields, rather than a rigid machine separated from employees by a barrier. Embodied AI says the arm design grew from CEO Francesco Stella’s PhD research at EPFL. It says AIR is designed under ISO/TS 15066:2016, with CE marking planned for 2027. The marking remains a future step, not a completed launch milestone.

The first shift still has a human operator

AIR does not begin a new job working independently. A human initially controls it remotely, a method called teleoperation. Those working sessions are recorded and become training data for Embodied AI’s own software. The company’s proposed deployment model therefore puts the robot to work under human control while it gathers examples of the assigned task.

Embodied AI says AIR typically learns to perform that task itself within a month. Even then, the operator stays available to handle situations the machine has not encountered. The promise is task-specific learning with human backup, rather than an immediate handover to a robot able to handle anything on the factory floor.

The company also says learning can transfer between units: what one robot learns, another can use. That gives its approach a second component beyond teaching a single machine—reusing the resulting capability instead of treating every unit as a fresh start.

Factory results precede the mobile robot

AIR uses the same AI software as Embodied AI’s two-arm industrial cell, which has operated in paid manufacturing pilots. Those included a consumer electronics manufacturer and an automotive and industrial technology supplier. Neither customer is named in Interesting Engineering’s account.

Embodied AI reported an average 95% task success rate across those deployments. It has not disclosed the number of tasks evaluated, the trial duration or how it calculated that average. Those missing details limit what the figure tells a buyer about sustained operation or performance on a different manufacturing job.

The launch release offers a narrower example: a customer proof of concept in cable handling. The company says its industrial cell achieved 98% reliability against a 90% target after five other startups failed to meet the customer’s targets. The release identifies Embodied AI as the provider of that data; this was an earlier cell test, not a result attributed to AIR.

AIR itself is being tested with global manufacturers, according to Embodied AI, which is beginning plans for scaled production. The company said it would accept orders after the unveiling. A price and delivery schedule have not been disclosed, leaving the commercial timetable less concrete than the hardware and training approach presented in Munich.