Giving Robots the Ability to Reason

We talk increasingly about humanoid robots, robotics broadly on this program. You are working on the software side, not the hardware side. Just explain the approach of covariant. So what covariance building is really the true beauty of robotics. So think about it as the general purpose brain that can sit behind any robotics hardware that gives it the ability to see the world. Think about what's happening around it and make decisions intelligently. And this is a big contrast with traditional robo

tics, which is really programming robots to do the same thing again and again, which just doesn't cut it in today's warehouse environment. Manufacturing environments where you have constant changes that are coming in. And the unique covariant approach here is we don't just build a single specific A.I. that can solve one task. We are really approaching it in a similar manner that, for example, like large language model is approaching chat bot are the language locations is a single model in multip

le use cases. But here's the thing that you don't control the hardware side, right? Take Tesla as a comparison. I don't know your thoughts on the Optimus program, but they are developing both the hardware and the software. They obviously have Dojo that they're working on in-house. What are the risks and benefits of just focusing on software if you're not too also doing purpose built hardware to match it? Yeah. So first of all, on Optimus, we think it's an amazing robotics program. Like it's a re

ally great human toy. Hardware that is being built is being iterated very quickly. For my covariant perspective, we believe there's not a single hardware form factor that would rule robotics to come because the physical world is very diverse. Like there's not going to be a single software, a single hardware at all. And we believe there are going to be multiple kinds of robot houses, some in the human form factors, some in the industrial and form factors, some with maybe a mobile robot with a man

ipulator on top of it. And all of these different hardware form factors still need to make sense of the same physical world. And covering this building exactly that, the same brain that can power multiple kinds of hardware to make sense of the physical world. So, Peter, talk to us about the inputs here. How hard what are the intricacies you need to go about building? What is the largest dataset ever to train this new robotics foundation model? That's an amazing question. So we have seen the expl

osive success of large language model and what is really behind its amazing generalization power is the fact that is trained on the whole Internet of text. And if you want to build a robot A.I. that is as smart as, for example, large language model. But in the physical world, you also need to build the same kind of data set. But there's no Internet to scrape in this case. So you need to deploy a large fleet of robots into the world doing diverse and dynamic things and collecting the video data,

images, data, robot actions and the outcomes of those robot actions in order to really train a model that understand the world in all kinds of settings and be robust, even in the where long tail scenarios and truly deliver a human level performance to our customers. And the reason you've been able to get such a wide, varied underlying dataset is because the amount of countries you're in, the amount of companies already working with. Can you just give us an idea of what this model is already bein

g applied? How are we starting to see it in our everyday lives? Yeah, so we have already to deploy the AI to more than dozens of customers in more than ten countries, and they are powering the warehouse operations, the e-commerce fulfillment operations in a lot of places. So very likely, like if you have order an item, for example, doing Black Friday, there's a really good chance that that item has been touched by one of the covariance robots that's operating around the world. Peter, very quick,

we have 10 seconds. Your favorite use case for the robotics you're working on. My favorite use case for the new Robotics Foundation model AFM one is really the ability to talk to it instead of robots that are just rigidly doing the same thing again and again. We now have the ability to communicate with robots in natural language, very similar to how you would talk to a chat ship or Gemini in natural language and ask questions. Really democratizing access to A.I. for a lot of people. We're reall

y doing the same with robotics, which I found one.