Zero-shot learning

Yes, sure.

So in our paper, the goal is to train a very big world model with real robot data and then to train reinforcement learning policies inside this very big world model.

So we don't use any simulator.

We don't use mock-up.

We have no demonstrations to imitate.

And here in this video, you can see the deployment of one of our policies on the real robot, ZeroShot.

The task here is to grab this box by its semi-soft handle and lift it in the Here, importantly, our policy only has access to the egocentric view that you can see in the top left corner.

And we apply our method to three different real robot tasks.