Inside The $440 Million Startup Building The Brain For Physical AI

本站添加: 2026-05-28

[00:00:00]There are already millions of robots today. There will be billions of robots soon. They will all need a brain.

[00:00:05][music] >> [music] >> The artificial intelligence boom has revolutionized the digital world, but a massive [music] barrier remains, the physical world.

[00:00:23]While large language models can write code in seconds, getting a robot to reliably fold a shirt or pack a box [music] with the correct amount of force has been, well, difficult.

[00:00:36]>> [music] >> Roboticists have previously been trapped by the limits of hand coding, but the three co-founders of Generalist >> [music] >> are mining data in new ways, enabling humanoids to navigate and manipulate the physical world as fluidly and intuitively [music] as AI currently navigates text.

[00:00:54]I was working [music] on applied physics research to design new types of solar cells. Most of my day was spent doing repetitive [music] manual labor in the lab. And that was one of my realizations that at some point all this repetitive manual labor probably going to be able to have robots help us with all this.

[00:01:11]>> [music] >> I did a PhD at MIT in Russ Tedrake's group, and Pete, our CEO, and I were there together. So, at that time we were working on autonomous drones. Then I went to Boston Dynamics. He had worked with Andy Zeng at Dmine for a long time. When we talked about starting a company, Pete said, "You know, Andy, we have to have Andy Zeng." And so, we just called Andy Zeng, and he was like, "Let's go."

[00:01:37]One of the amazing things that's happening now >> [music] >> is that robotics has really started to enter the pre-training era. So, the way this works for large language models, right, is you go and you you take as much of the internet as you can possibly find, all the text that you can possibly find, and you train a very large model on this [music] massive array of this very diverse textual information.

[00:01:59]Robotics has started to enter >> [music] >> very analogously this pre-training era.

[00:02:04]And we haven't had the [music] data to go do this, right? There isn't just a ton of robotics data all over the internet that you can download and train a model on.

[00:02:15]We're making it all ourselves.

[00:02:19]>> [music] >> The startup raised $140 million at a $440 million valuation in 2025.

[00:02:29][music] We met all the groups and there was really only one group who truly believed in the scaling hypothesis for robotics and then had the capabilities and like the skills, [music] the research capabilities, the robotic capabilities to go and breathe life into that. And that was Pete, Andy, Andy, [music] and the rest of the generalist team.

[00:02:49]I think the interesting thing is the data has allowed them to pre-train the first [music] general robotics foundation model. Uh and the presence of having, you know, the first robotics foundation model trained from scratch means that they're able to actually advance the frontier of uh of AI research into this area.

[00:03:08]What Spark Capital and [music] other VC firms have realized is that Generalist's biggest selling point is their proprietary data collection system.

[00:03:17]They've built [music] specialized hardware specifically to capture human dexterity.

[00:03:23]What Data Hands is [music] is the uh handheld devices that we've constructed.

[00:03:28]They're very simple by nature, but that's by design because we want to be able to scale it to um uh many different scenarios. And so, [music] the the purpose of it is to really bring in all these experiences into something that we could then put on robots to be able to express and be able to interact with the physical world in ways that we'd never seen robots do before.

[00:03:50]We have half a million hours of dexterous robotics data.

[00:03:55]Well, now we get to ask the question, what happens if you take half a million hours >> [music] >> and you train a model to do X or you train a model to do Y?

[00:04:04]Well, General is the only place you can ask that question in robotics. We're making algorithmic advances that nobody else can make because we have the resources to do it.

[00:04:14]What we're seeing here is a robot and it's [music] running uh one of our models, one of our Gen 1 uh models and it's it's doing this task here of of servicing this robot vacuum.

[00:04:25]The bigger picture here is this is just one example of a pretty challenging, honestly, dexterous task. More importantly, it's the type of task that in reality, nobody would use traditional programming to try to do.

[00:04:39]As the robot goes through all these behaviors, >> [music] >> it is, in a very real sense, drawing on all of its experience from all of its [music] pre-training to know how to do some of these delicate and complicated little maneuvers, right? Like how much force to apply as it's um you know, removing these pads from the robot vacuum or how to use both of its hands in coordination [music] to accomplish the dynamic maneuver of of, you know, spinning it around as it's flipping it.

[00:05:05]>> [music] >> This massive influx of physical data is leading to something researchers have sought for decades, a robot that can think on its feet and recover from unexpected scenarios. There's this incredibly interesting and subtle way in which people push things and move things around in the world, which is this idea of physical common sense.

[00:05:34]>> [music] >> And it's this reactive intelligence that we have at and and we take for granted actually as people. And what we're starting to see is these glimpses of this kind of improvisational intelligence that [music] can react to these new situations and still be able to perform these micro [music] corrections and subtle air recoveries so that it can continue to do the job without necessarily landing on the mistakes. We have this task where we are working on these belts, and the belts are flexible, and they're really difficult to model. I thought to myself, I can't program [music] this. Like, I do not know how to write software to do this. And then we trained the model, right? So, we showed the model how to do it, and we used the whole pre-trained base, and then the [music] robot is actually really good at doing it. To the point that you can smack the thing with a hockey stick [music] halfway through, and the robot just fixes it.

[00:06:25]So, there was a question on what would happen if we you know, knock the pad out of its hands.

[00:06:40]But, the goal isn't to perform impressive lab tricks.

[00:06:44]By stepping away from complex [music] hardware design, the generalist team is pushing toward mass adoption, ensuring their efforts are not stymied by global supply chain [music] issues.

[00:06:54]One key thing to drive all of that scale is simplicity in design. And so, a really big piece of that is just thinking about [music] how can we pair down, step away from complexity, go back to grassroots, [music] first principles, and figure out how we can build something that is simple but captures the Pareto distribution [music] of 80% of capabilities, and move from there.

[00:07:16]Building hardware's hard, right? So, we have to really think about how do we build the right hardware that scales in the right way. We're making a product where lots and lots of people can use these things because, you know, it's not a bundle of wires going everywhere, and you need a PhD to turn the thing on.

[00:07:37]>> [music] >> We have forever thought that strong technical founders who have like a legitimate breakthrough in either research or science or technology, if given the opportunity, find ways to bring it to market, right? And so, like, yes, hardware is tough. Getting robots into the real world is going to be tough. If you empower these strong technical founders, they find a way to breathe life into what it is that they're doing.

[00:08:09]With [music] their new Gen 1 model rapidly improving, Generalist is poised to be the intelligence layer for a massive [music] hardware revolution.

[00:08:18]We just announced about a month ago this Gen 1 model.

[00:08:21]>> [music] >> The rate of improvement in just a handful of months from where we announced our Gen 0 model back in November, then April, we announced our Gen 1 model.

[00:08:30]Because robotics is now in the pre-training era, we can very predictably improve these models.

[00:08:37]Where I want Generalist to be is in a place where we are enabling loads of robots in the world. And [music] maybe that's Generalist is inside other people's robots, and because, you know, we can't build all the robots, right? So, so Generalist is powering a whole bunch of robots. Maybe Generalist is also has our own robots in in certain areas, [music] but fundamentally, like, what do we want to see? We just want to see robots be useful for everyone.

[00:09:06]>> [music] >> We want to think about it as the application layer of robotics. There's going to be, uh, I don't know, a Cambrian explosion of people, entrepreneurs, and developers, and engineers building creative ideas [music] on top of these robotic foundation models.

[00:09:22]Just as hyperscalers [music] need new power sources to fuel the digital AI boom, the physical AI boom will require universal adaptable brains capable [music] of infinite tasks.

[00:09:38]Just like how today you can interact [music] with a language model and it gives you five different essays and you get to pick your favorite. I think there will be a future where one day you can [music] communicate with your favorite AI system and it would one day give you five different physical prototypes to choose from.

[00:09:54]In the future, [music] there will be many robots, millions, billions potentially. And I think that what I'm excited about is a model powered by generalist [music] that powers them all.