Palantir's Ben Harvatine demos edge robotics: pushing ontology down to the factory floor

Uh, we are going to have you hold this microphone as much as you can. Um, but why don't you, uh, kick us off with an introduction on yourself and kind of, I'd love to know how you found your way to Palanteer. That'd be super interesting. Yeah, it's, uh, kind of an odd path.

Um, I studied mechanical engineering and architecture in college. So not what you would think for a software company. Uh worked for Annheiser Bush. Oh, no way. Beer company for a year. That was a great sort of transition from college. What were you doing at uh Annheiser Bush?

Yeah, it was a it was a management training program kind of rotation based. So yeah. Um after that ran a hardware startup for a bit. Okay. Went to another hardware startup. Um but I had some buddies from college who had worked here.

And thing about Palunteer seemed like everybody had um just kind of like more autonomy and authority than Yeah. I saw anywhere else. Yeah. Yeah. Amazing. Uh so what do you want to show us today? Can you uh give us a little tour of what's going on? I've got a little brought a robot. Yeah. One robot slide in.

Bringing a robot is a great sign of respect in our culture. So thank you. Well, you know, you can imagine uh you know, when we have uh you know, events like this, there are a lot of demos, it's pretty screenheavy with software stuff.

Um, and we've seen a lot of, I'd say, like increasing demand for our edge offerings, hardware offerings, really trying to push the technology further and further down to the shop floor and into the field.

And so I wanted to put together something, you know, just a little kind of toy demo that made that a little bit more tangible for people who are here. Yep.

Um so uh walk me from my understanding to how we get to the edge, how we get to robotics because uh my famous like the case study that comes to my mind for uh Palunteer in terms of like making things in the physical world is like I think the Airbus example.

So I and and and whenever somebody says, "Oh, what what does Palanteer do? " I'm like, "Okay, imagine a plane. There's a bunch of different parts. You got to have a certain amount of seat belts. You got to have a certain amount of engines. You got to have a certain amount of fuel lines.

You got to have a certain amount of chairs. And all those come from different places and they all have different lead times and strengths and they need different safety requirements. Did they get checked off?

And so you put all of that instead of just in a loose database, you put it in a database, but then you have Palunteer that's actually tying everything together. So you know if there's a lead time on engines, you need to order more seat belts in three weeks instead of two weeks.

And that's kind of how I explain Palunteer in terms of like make a big thing that's complex. Is that roughly right? And then how do you walk from that to like we need Palunteer to somehow interface with like a robotic arm? Yep. Yeah. I mean that's roughly right.

Like the way I think about it, it's like anywhere you go people have data scattered all over the place. So the first step is can we get that all into one place? Got it. Then can we model that data so it's as easy to work with it as it is to talk about the concepts that represents, right?

Just like make it kind of so so there's this big meme in Silicon Valley and defense tech right now that like there's a whole host of manufacturing guys. They're all aging out. they're 65 and everything that they know about how to make a widget, whether it's a chair or a rocket motor, it's in their head.

They haven't written it down. Maybe it's some some loose notebooks. And so, this is kind of a way to jump and start getting more data online, right? We're actually not throwing out the data. We're capturing it. Correct. Yeah.

And really, like the whole point of any of these data exercises is you just want to put the right data in front of the right person at the right time to make the right decision. Yep. And then just be able to close the loop and learn from it. Um, and so if you're looking across a supply chain, that's how you do it.

If you go down to a factory floor, the process is there. That's how you do it. And so when it comes to this robot, we're basically just like pushing that edge further.

So instead of um you know, popping up an alert on a screen that tells somebody to go do something, what if you could actually just tell the robot to go do it. Okay.

Um so again, sort of a simple like toy example here, but the basic idea is that, you know, this is a little work cell that we made with a a robot arm and a camera. 3D printed, right? Yeah, it's it's all Yeah, it's all 3D printed. Even the arms are Oh, wow. Okay. Yeah, I didn't realize that. Cool.

Um, and so, you know, it's it's kind of set up to be a dumb terminal that kind of works and looks like, you know, the robot arms you'd see on a factory floor. Y, you can give it moves to take, maybe you can ask it for a picture, but past that, it's not doing any heavy computation on board.

Um, but then you can push uh, you know, that data to an edge hub that can run embedded models, um, can run embedded ontology.

So you can actually take that that kind of model of the world in terms of objects, relationships, um actions and models and you can push that down to the edge and even if you have um say like a like a network sparse environment where you don't have that real-time uplink to the cloud, you can continue to run off of that ontology.

Yeah, we were looking at uh semi analysis. They put the the five levels of robotics. I forget exactly how many levels there were, but they were trying to map the self-driving car analogy to physical robotics.

And I believe like level zero or level one, like the most basic was you have a pre-programmed robotic arm that's doing the exact same move. It's taking the windshield and putting on the F-150. And it's this huge arm and you can't go near it because it's there's no cameras on it whatsoever.

And if you step in that work cell, it will kill you if you don't if you're not careful. Um, and this seems like uh a step towards like level two where we're able to actually understand what different products mean.

If there's, oh, this type of product shows up, there's going to be more likely that there's a defect or you need to adjust what the robot is doing. How can you actually get that data into something that's actionable? Yeah. Yeah.

And even in like this simple demo, we've got, you know, it'll trigger alerts on, you know, it tries to execute a move and you end up with like a block like jammed up here. It'll say, "Hey, you got a jammed hopper. You need to declare that sort of stuff. " Okay. Um, interesting.

Um, uh, where does this play in like the stack of other software? I know when we talked to what was it DRA, uh, our buddy Phil, he was saying that like he's working with automotive companies, but then they also have a lot of there's a lot of like lower level control software on machine lines.

Some of that's from German companies that I think we just talked about with Dr. Karp. Um but uh like where do you see Palunteer playing in the stack?

Uh you have a bunch of data the database you put Palunteer on top but then at a certain point there might be uh some robotics company that makes the robot and then they also might have some control software with kind of a messy API or something like that.

Yeah, I think we can be pretty agnostic about how far up or down the stack we go. So we've got I'll pull this box. Yeah, please hold this. This is uh this is the node that goes on the edge, right? So this is so this is an example of an edge node that um one of our partners Edgecale makes.

So this is that box that you can stick in the closet factory network to those existing machines that you have on the floor if you just need a turnkey solution. Yep.

And then I think at the other end of the extreme that's where we've got something like this where this really at the end of the day is an ontology defined piece of hardware in that the machine itself its entire configuration the state machine is running everything about it is defined in the ontology lives in the ontology and it's like really just like a bespoke piece of hardware running that ontology native software.

It's a monument. So yeah, you you you know, if you've got like more nason operations, more green field operations, you think about some of the companies we work with in um defense tech, it's like they can go all the way down the stack if they want to. Sure.

For some of the, you know, the larger, more established customers that we're working with, you know, the plug-andplay solution. Yeah. What's the sweet spot for the specs on an edge scale like uh edge node? Like something on the edge like do you need to be running like a large language model?

That feels like something that you could do on you could I'd say it depends on the application like we we've done we've done some um some like examples of that even like previous AIP cons.

It's like do we need the uh like the local app served up with a chatbot for the line operator who can just be like what's going on and it just talks to you. Yep. There's and it's not just purely deterministic. Okay. If if the block is blocked then send the error message.

instead it's it's actually interpreting a bunch of data in a kind of non-deterministic way.

So I'd say it's like you know I think like anything it really depends on the application and the users because again there are a lot of guys that are working on these lines guys and girls where they don't need another screen in their life and so it's really finding like what's the right way to interface with those operators to ultimately just drive the better decision making.

uh how much is uh like how much is the what is the role of the FDE in in this kind of new era new territory because it feels like Yeah. Are you graduated from being an FD yet or is it once an FD always an FD? I I think it's once an FD, always an FD. I I try to keep my hands on keyboard as often as I can still.

Um you know, still flying out to whoever axle factories in rural Kentucky or whatever. Awesome. Um yeah, I think the closer you can stay to that stuff the better. I think really like the role of the FD is like just like it always has been. Go on site with the customer. Yep.

Don't just understand but internalize their problems, their challenges, you know, and uh solve go create some value. Uh well, thank you so much for hopping on the stream. We appreciate that. Congratulations on everything. Thanks for bringing your baby. Yeah. Yeah, you can definitely take this out here.

I will grab this and we will have our next uh guest Danny Lucas uh from Palunteer coming in. He also has a demo. Um, do you guys know if the demo is uh is gonna need the HDMI cable? Is that right? Okay. So, uh, we will bring in Danny uh whenever we get a chance. Yeah. Let's let's bring in our next guest. Here he is.

What's going on? Welcome to the show. How are you? Great to have you. Did you do a live demo? Always. That is bold. Doing a demo is on a live stream. This is live. So literally anything you share on your screen potentially will go out to the internet forever to be baked into the future super intelligence the future.

Yeah. Baked into the training models of the future into the pre-training data.