GEN-1.5

GEN-1.5 is an AI model that teaches robots new movements. Show a robot arm a person sweeping blocks into a bowl just one time, and the robot stores that short clip in its 'memory window' and immediately repeats the same motion. It was released by Generalist AI, a startup building AI for robots.

Normally, teaching a robot a new move means repeating the same action tens of thousands of times while very gradually adjusting its internal settings. GEN-1.5 cuts that process way down. Feed it just one or two short demo clips into a 30-second memory window, and the robot can perform the task right away; even with a bit more practice, the actual changes to its internal settings are tiny. Think of it less like rebuilding an entire library and more like moving a few books to a different shelf.

What's more interesting is how well it generalizes. A robot that learned to sweep blocks with a brush was able to do the same motion holding a banana instead, and it could also reproduce a demonstration it had only seen in simulation on an actual physical robot.

GEN-1.5 comes up when articles describe demonstration-based teaching, such as "a person sits in front of a robot arm and sweeps blocks into a bowl just once." But there's a catch that's easy to miss: it doesn't nail every task perfectly. Success from watching just one demonstration was only around 59%, and it took about five minutes of additional practice to push that up to around 83%. Also, this isn't a consumer product yet — it's a research-stage model currently accessible only through partnerships.