Sector · Data collection for AI
The machine never saw a pair of hands work. The glasses show it
Almost everything on this page is a dataset, not a product. Human first-person video is what robot learning was missing, and the glasses are the cheapest way to record it. Also here: the day Amazon announced its delivery glasses, and what the Chinese industry signed.
A robot arm can be taught by teleoperation: a person drives it, the machine copies. It works, and it is slow and expensive, because every minute of training costs a minute of someone sitting at the controls. The alternative is to record a person doing the task with their own hands, from their own eyes, and let the model learn from that. For twenty years nobody had that footage at scale. Glasses with a camera, gaze tracking and hand tracking are how it started to exist.
Meta's Project Aria is the visible engine. It is not for sale: Meta lends the glasses to research partners, more than 290 of them running over 1,000 devices in 27 countries, and the partners have logged more than 8,000 hours. Out of that came Ego-Exo4D, where 740 people cooked, played sport, repaired bicycles and performed health procedures with Aria glasses synced to four external cameras, publishing 1,286 hours. And EgoVerse, a consortium of Georgia Tech, Stanford, UC San Diego, ETH Zurich, MIT, Meta, Mecka AI and Scale AI: 1,362 hours, 1,965 tasks, 240 scenes, 2,087 demonstrators. Microsoft built a different shape with HoloAssist, where 222 people worked in instructor and performer pairs and the performer's headset recorded seven synchronised streams.
The one number that measures a result
Nearly all of these projects publish hours, not outcomes. The exception is EgoMimic at Georgia Tech: about 90 minutes of everyday human tasks recorded on Aria glasses, co-trained with a low-cost bimanual manipulator, reported as a 400% improvement over training on teleoperation data alone. Read that for what it is: a benchmark on the tasks the lab chose, published at CoRL 2024 and ICRA 2025 with the hardware, dataset and code open. It is not a robot working in a warehouse.
The part that is not research
Amazon showed AI glasses for delivery drivers at its Milpitas hub on 22 October 2025. They read packages, guide the route, flag hazards and dogs, and take the proof-of-delivery photo without the driver touching a phone. Every one of those readings is also training data. The announcement came one day after the New York Times reported internal plans to avoid hiring more than 600,000 workers by 2033, with an internal goal of automating up to 75% of operations. Amazon published no productivity figure with the glasses.
The industry knows how this looks. In June 2026 a Chinese alliance including Lebird, Rokid and ZTE signed a 15-rule self-regulation pact on trustworthy vision for AI glasses: clear notice and consent before a sensor turns on, face, iris and voice processed locally where possible, data compliance along the chain. It is a pact, not a law. In the same season Chinese education authorities banned AI glasses from gaokao exam rooms, treating mere possession, powered on or not, as cheating.