One demo. New skill. — GEN-1.5 learns physical tasks in seconds

Generalist AI just dropped GEN-1.5, an embodied foundation model that can pick up a new physical task from a 3–12 second demonstration — no fine-tuning, no gradient updates.1

Across 10 short-horizon manipulation tasks, one-shot prompting hit 59% average success. With about five minutes of data and ten gradient steps, that rose to 83%.1

The part worth watching: after a demo, the robot does not only replay it. It improvises — clearing a paper lid off a bowl, peeling a stuck Lego off a fingertip, turning a jar lid with two hands when the demo used one, even using a banana as a brush.1

Physical AI is starting to look less like scripted demos and more like “show once, then figure it out.”