Covariant

Contribution to AI

Industrial picking robots had always been told in advance what they would see: each item measured, each grasp point specified, every new product line a fresh engineering job. The alternative case made from Emeryville was that manipulation should be learned once, centrally, from what an entire fleet encountered. The proof was unglamorous and therefore convincing — a robot installed at the electrical wholesaler Obeta outside Berlin in late 2019, working a catalogue of more than ten thousand items, including polybags, blister packs and the transparent objects that defeat conventional vision, at better than ninety-nine per cent accuracy and with nobody standing by. Warehouse operators had been hearing about learned manipulation for years; this was it running a shift. The partnership with KNAPP put the same system into other operators' lines, and because every pick fed back into one shared model, a production fleet doubled as a data-collection instrument. RFM-1, released in March 2024, took the logic further: eight billion parameters, started from a pretrained language model and continued on text, images, video, robot actions and sensor readings tokenised into a common space, so that a video the model generated of a bin after a grasp served as its physics prediction, and a typed instruction could change how it lifted. Amazon licensed the models and hired the founders that August.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).