We're building consented, rights-cleared, first-person datasets of skilled hands at real benches, labeled step by step, and the lab-action models trained on them.
Video from the researcher's point of view, the way hands, tools and parts actually look at the bench.
Sessions are segmented into steps, with parts, connections and timing.
Every session is opted in by the lab, with consent from everyone on camera.
General AI models still struggle to follow real lab work. We think the missing piece is data: labeled video of skilled hands doing real procedures. We're building that data, with consent, across many labs and kinds of work, and the models that learn from it.