Brian Gerkey

And that's what we think is going to really change how the willingness of people to be able to deploy these applications is knowing that they've got the freedom.

Sometimes it takes some adaptation.

Like, you know, you could have two robots that are about the same size, but the joints have...

different limits, the workspace might be a little different.

There could be some tweaks that are required.

But by and large, being able to write an application once and then ideally only configure it in order to run even on a robot hardware from a different company, that's a game changer.

Hmm.

I think it is.

So it depends on what problem you're trying to solve.

So I think on the training side, we're really, really good now as a not just intrinsic, but as an industry, we're really good at generating synthetic image data.

If what you want is to be able to, you want to train a system to understand images or even video, boy, we're excellent at that.

Now we've got, you know, we've got these loops where I can start with some data.

I can train a generative system that will then generate synthetic data, which at a higher volume, which then I can train, use to train the next step, right?

That's working really, really well.

The training data for physical contact with the world, that's harder.

And that's because we understand from first principles how to, like, we have an understanding of how physics works at different levels, right?

But it turns out getting very detailed models that are accurate is pretty hard.

And most simulators take what's called a rigid body dynamics approach.

So I just imagine infinitely rigid things that are just coming into contact and they probably have like a single point of contact, like a billiard ball model, right?

Like a simple...