And what we're doing is the models that we're training are able to simulate forward in all these different types of environments. 

So what that means for robotics, and we think this is the biggest bottleneck to what we call physical intelligence, is for the robot to know how to plan its actions when it wants to do. 

And the way it can do that is to simulate forward what kind of actions you want it to take. So if you wanted to pick up a cup, for example, it will literally like simulate itself picking up a cup and then work backwards as to what kinds of robotics actions it should be taking. 

And that's only possible with a very strong role model so that it can generalize across