New Models 2026-08-21

Show the robot once — three to twelve seconds — and it gets the job right 59 times out of 100 with no training at all

Generalist AI published GEN-1.5 on 19 August. The demonstration is not training data; it is the prompt. A three-to-twelve-second clip of someone doing the task — recorded by a person holding grippers, or by the robot itself — is loaded into the model's thirty-second context window, and the robot then attempts the task with no gradient updates and no fine-tuning. Across ten varied jobs, including opening a jar, unzipping a pencil pouch and pulling money out of a purse, this one-shot approach averaged 59 percent success, with a spread of about ten points. Ten gradient steps on five minutes of data pushed it to 83 percent. The model takes in video alongside sensor, language and body-position data and emits movement commands a hundred times a second. The company says the abilities it is showing off — transferring from simulation to the real world, using unfamiliar tools, working around obstacles, using both arms — were never trained for as such, and emerged from eight months of continuous pretraining on physical interaction. No weights and no product were announced.

Why it mattersThe reason useful robots stay rare is not that they cannot be programmed, it is that every new task has traditionally meant a fresh round of data collection and retraining, which prices out anything but high-volume factory work. Learning from a single demonstration, in the time it takes to show someone how to do something, is the difference between a machine that has to be engineered for a job and one that can be told. Two cautions belong with the number: 59 percent is a long way from reliable, and this is the company's own evaluation of its own model, on tasks it chose. But we covered an open-source Chinese one-shot robot learner three weeks ago, and a second group reaching the same idea from a different direction is how a research result starts turning into a trend.

#Robotics

✓ Verified · 3 sources

Nevada just cleared up to 8,000 driverless taxis for Las Vegas — Tesla's ceiling went from 10 to 5,000 in five weeks
2026-08-21 DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21 The doctors and lawyers who grade AI answers are now a half-billion-dollar business — and this one grew fivefold in eight months
2026-08-21 The robot dog China sells the world was measured, to the millimetre, from one the US Army paid to invent
2026-08-20 The robot did a four-month internship on a car line. It went from missing one nut in ten to missing one in fifty.
2026-08-20

Show the robot once — three to twelve seconds — and it gets the job right 59 times out of 100 with no training at all

Related stories