Robot deployment requires solving the whole stack

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 3:38

   And then now we need robots, but actually getting robots into the factories, into manufacturing lines, isn't that straightforward? What are the biggest problems there? There are a million problems, unfortunately. So we have to solve all of them, right? Everybody talks about the model. The model is obviously a really big problem to solve. It is true. We've had robots for the last 60 years, but they were these dumb, repetitive machines.

2. 3:59

   right, like Rube Goldberg, special constructed machines that are the factory, I would argue in the last 60 years, we haven't really had any technology that's actually a real robot, right? If you go in your mind's eye and you think about kind of an intelligent machine that's reactive and can do a wide diversity of tasks in a wide diversity of places, we actually don't know how to build that yet. And one of the key limiters there is intelligence. But even if you have the intelligence, how do you get the hardware, right? The United States today, because of this intelligence bottleneck only deploys about 38,000 or in 2025 deployed 38,000 industrial robots in the United States, that's a rounding error against a million unfilled jobs in manufacturing logistics just in the United States. And so we don't have a supply chain for robots in the United

3. 4:40

   States. We have to build that from scratch too. And then thirdly, there are no customers, right? There are no robot bodies waiting in these customer sites for the software because the robots can't do those jobs and those customers can't afford robots. And so to make robots work, In our view, you have to walk and chew gum, right? You have to do all three layers of the stack. You have to train the model, you have to build the hardware, and you also have to deploy.

4. 5:02

   Yeah. Tell me about this third piece, deploying and talking to manufacturers, talking to companies. Yeah. What are other people who are building robots and intelligence getting wrong about what it actually takes to get all of that intelligence and then have that diffused into the actual manufacturing? Yeah. Well, I think you hit the nail on the head there, which is like step one is talk to the customers.

5. 5:24

   And that's very important and not necessarily obvious to everybody. But you have to really understand what the customer's problem is. And we have an enormous market of tens of thousands of manufacturing logistics businesses here in the United States. Most of them are small businesses. Almost none of them have robots. Nine out of ten U.S. factories and warehouses don't have a single robot. And so when I started this business, I didn't know anything about factories.

6. 5:48

   And so I started cold calling factories. I started visiting factories with my co-founders. and we started to understand what problems they actually have and kind of what shape of model, what shape of hardware actually works to solve real customer problems today. And our thesis is if you can get a foothold, right? If you can build a model and hardware that solve real customer problems today, you can start to run a flywheel.

7. 6:08

   You can deploy more robots. That makes your robots cheaper. That means you have more data. You can deploy more robots and so on and so forth. Yep. Tell me a little bit more when you're talking to customers and you're identifying the actual problems they're facing. Yeah. Just summarize that for me. What were these big problems? Yeah, so it's labor, right? It's how do you staff a factory? And there are lots of different jobs.

8. 6:28

   And some of the jobs we can do today with robots and some of them we can't. One big job that people don't know a ton about is palletizing. It's stacking cardboard boxes of stuff onto pallets. And if you think of the United States industrial supply chain as a directed graph, Every edge in that graph is a shipment on a truck of stuff in boxes on pallets. And the movement and stacking of those boxes on the pallets, of boxes off pallets, of moving pallets around, is done almost entirely by hand, which is crazy, right?

9. 7:04

   So we kind of looked at that problem. We're like, okay, this is not super dexterous work, right? It's basic pick-and-place work. We can solve that with foundation models in the 200 to 400 million parameter range. We can build at a reasonably low cost a robot worker that can, you know, do that task in the hardware. That's our first robot called Cassie. And we can deploy that today and, you know, save factory workers from backbreaking tasks, help factory owners hire, and generally make goods cheaper for the American consumer.

10. 7:31

    Wow. Okay, so now how does this actually solve?