Brian Gerkey

Newtonian kind of model.

And I can compute the point of contact and I can compute what the reaction force must be and then where they ricochet off.

But the real world is so much more complicated than that.

There are many points of contact.

The surface properties matter a lot.

The texture, the friction properties, like friction is famously difficult to simulate very well, which also makes it difficult to generate synthetic training data for.

So that's one gap is how do we get

data for that physical, for the in-contact kind of tasks.

And that's why you see some companies having, you know, essentially armies of human teleoperators who are moving robots around and then gathering data from physical systems because it's so difficult to generate synthetic data.

We're getting better, but that's still one gap.

This is where I've got some of my colleagues are smarter about this than I am.

But I think if in principle, we understand how, let's say, frictional contact works.

But in practice, it is very, very difficult to implement

way to compute it that is tractable it is if what I want to have is a if what I so if I'm studying like in like like there are great models for cars so if you you know you think about you're gonna design a new car or you're gonna tweak something about the tire design and you want to understand how does that behave as it's coming into a corner at a particular speed and there's this much water on the surface like those specialized models we have

Because of the need for them, the industry has a really, really good understanding of how that works, and we can compute that.

It might take a long time to compute, but that kind of specialized model we've got.

If what you want to ask instead is, well, I've got a robot, and it's in a factory or warehouse, and it's going to interact with all these different things, which have all these different properties, and I want to get the right answer for all of those interactions, that's a lot harder.

yeah there's just so much more it's just like i don't know the exact term but it's a greatly expanded i don't want to say exponentially but it's a it's a vastly larger space that you have to model for it so much more difficult and so the places where very very very detailed simulation has gone really really far is in those domains where you can you can narrow it down you can say like you know what we care a lot about tire to road surface contact that's super important let's build

A very, very deep simulator for that.

You can look in aerospace and say, you know what we care about?