Generalist builds foundation models for robot manipulation, training one model to drive many different robot hands instead of writing control software for each machine.
Generalist's GEN-1 model works across a wide range of end effectors, and GEN-1.5 can learn a new task in context from as little as twelve seconds of demonstration. Pete Florence, Andy Zeng and Andrew Barry left Google DeepMind and Boston Dynamics to found the company in 2024.
Manipulation generalizes across hardware. One model that drives many kinds of robot hands means every hour of data collected on any machine improves every machine, which is a different growth curve from writing control software per robot.
We backed Generalist because robotics has been held back by software written one machine at a time, and a single model that drives many kinds of robot hands changes what it costs to bring a new robot into service. Millions of robots already work in factories, warehouses and laboratories, and none of them share what any of the others learned. Pete Florence, Andy Zeng and Andrew Barry came out of the labs that did the underlying work on manipulation, and they went at the general case rather than one narrow task.