Embodied AI Shows AIR, a Wheeled Humanoid With Soft Tendon Arms That Learns Factory Jobs From Workers

Most humanoid robot launches lead with legs, backflips and a video set to music. Embodied AI, a spin-out from EPFL in Lausanne, went the other way with AIR, which it unveiled on November 6. No legs. A wheeled base, because factory floors are flat. And arms built to be safe to bump into, before any software gets involved.

Safe by build, not by code

AIR's arms are tendon-driven. The motors, gearboxes and electronics sit in the torso and pull on cables, so the arms themselves carry very little mass and give way on contact. Most industrial robots get their safety from software: speed limits, force sensing, stop zones. Here the hardware is soft to begin with. The company designed AIR to ISO/TS 15066, the technical spec for robots working beside people, and plans CE marking in 2027.

CEO Francesco Stella built the arm design during his PhD at EPFL and spent time at MIT CSAIL before that. His co-founders, CTO Kai Junge and COO Max Polzin, are robotics PhDs with backgrounds at Cambridge, ETH Zurich and EPFL.

Learning on the job

The training model is the part to watch. Every new task starts with a human operator controlling the robot, and every run is recorded. The software learns from those demonstrations, and the company says AIR usually takes over a task on its own within about a month. The operator stays on call for anything new, and what one robot learns can carry over to others.

That is a practical answer to the problem every factory robot runs into. Programming a robot for a new task is expensive, and most small production jobs change too often to justify it. If a line worker can teach the task by doing it a few dozen times, the economics change.

Proof so far

The evidence is early but specific. In a test run by a large consumer electronics maker, Embodied AI's two-arm industrial cell, a stationary version of the same idea, hit 98 percent reliability on a cable-handling task against a 90 percent target. Five funded startups took the same test and none met the targets, the company says. It reports signed contracts with major manufacturers and paid factory deployments of the two-arm cell, without naming customers. The round was led by Faber, with TechShop, Look AI Ventures, Plug and Play and Vento taking part. The amount was not disclosed.

My take

Cable handling sounds dull. It is one of the jobs factories still do by hand because cables are floppy and rigid robots are bad at floppy things. A robot that is compliant by design and learns from a person's hands is aimed straight at that gap. Skipping legs is the sensible call. Legs make for better demos, wheels make for better uptime.

The open questions are the usual ones: price, how many tasks a single robot can hold, and how well a month of demonstrations survives a change on the line. AIR is still in testing with no launch date. Worth watching anyway, because it is the rare humanoid built backward from a factory job.