An applied research lab building generally intelligent, affordable robots.

Fully general robots will transform what is possible for humans to do in the world. The first step in building capable robotics models that can act autonomously is to scale every part of the stack.

We're building inexpensive and durable robots that can be easily manufactured and deployed, in order to greatly increase the amount of physical interaction data available. Our first goal is 10M hours of real-world interaction across a fleet of several thousand robots.

We're also making simple, scalable methods for learning behavior. We are developing methods to use internet-scale video to train models that know how to act in any visual environment, including video games, computer use, and real world settings like robotics.

We believe that when put together, scalable pre-training and post-training will produce robotics models that are robust, capable of performing long-horizon tasks, and capable of acting autonomously for hours without human intervention.