General Robotics Is Betting on Modular Intelligence, Not One Robot Brain

0:00I'm Mike Witzman, co-host of the Robot Report podcast. If your company wants to reach decision makers in robotics, automation, and AI, you're in the right place. Each week, thousands of industry professionals listen to this podcast for the latest news, expert insights, and conversations with leaders building the future of robotics. This spot could be yours. To learn more about podcast sponsorship opportunities, visit therobotreport.com slash podcast.

0:37What's up, everybody? Welcome in episode 263 of the Robot Report podcast. I'm Steve Crow, executive editor of the Robot Report. Thank you, as always, folks, for tuning into the show. Hope everybody's having a great week. I'm joined, as always, by Mike Oitzman. He's the senior editor of our robotics group, or Grandpa Mike. Is that the new name for you, Grandpa Mike? Yeah, indeed. I'm a new time. Grandpa's been an amazing experience. We're hanging out with mom, dad and the newborn. And I forgot how exhausting it was to be a new parent, Steve. That's why I got these these big bags under my eyes. Right.

1:15Yeah, we all have those. That's what kids do to you. Yeah. So, yeah. Oh, that's great. I'm happy for everybody. Wish everybody well. So good times. Thanks. That's great. Awesome to you and your family. Hope everyone's doing well. We have a great show for you on tap. Our guest today is Sai Vamprala. He's the CTO of General Robotics. We'll play that a little bit later on in the show. But Mike, why don't you set up that conversation for us? Yeah. In this conversation, Si is going to talk about how their new product called Grid helps bring together models, simulations, and robot-specific skills to help robots move out of the lab and into real-world deployments.

1:53And we're going to talk about robot agnostic intelligence, which I still don't believe you can do yet. But, you know, Cy is going to talk about what their view of that is and, you know, the role of agents, which everybody's also talking about and how enterprises can build and deploy those skills and keep ownership of their data, which is another thing that everyone's starting to talk about as well. So those are all, you know, great topics to share. And we're going to hear all about general robotics in this conversation. Awesome. So we'll play that a little bit later on in the show. We'll recap the news here in a minute as we do every week.

2:28And there's some juicy stories out there. We're going to break some more news, Mike. Yeah. So I think that's our new motto on this podcast is breaking some news. And I think it's going to prove that, you know, we might know what we're talking about every now and again, right? Which is kind of fun. But before we get to the news, I did want to tell folks about RoboBusiness. This is coming up quick, three and a half weeks away, October 20th and the 21st out in Santa Clara, California. Our 20th anniversary show, which is a lot of fun. So we'll have some fun photos throughout the show looking back at not only 20 years of RoboBusiness, but also 20 years of the robotics industry as well.

3:05We'll have over 2 ,000 people, over 100 exhibitors on the show floor, 70 plus speakers on the floor talking all things commercial robotics development, talking the business of robotics. So if you want to head to robobusiness.com, you can check out the agenda. We're still adding a couple people here and there. But some amazing people will be on our stages over the course of the two days. And as we've done for the last few weeks, make sure you stay tuned to the end of the show. where we'll share a discount code exclusively for you, our loyal podcast listeners, to help you save a couple bucks on RoboBusiness passes.

3:39But coming up quick, October 20th and 21st out in Santa Clara. We hope everybody here listening to this will join us for the 20th anniversary of RoboBusiness. All right, on to the news of the week. And we have you covered with all the latest news and information on our websites, The Robot Report and Automated Warehouse. Make sure you follow us on social media channels. Make sure you sign up for our weekly newsletters so you don't ever miss a beat. Again, we're going to break some news. We need to have like a breaking news sounder, Mike. And was this last week that we talked about this? Obviously, humanoids are in the news every day, all the time now.

4:13We talked about agility last week. I believe it unveiled the new leg design for Digit 5. So now it has, instead of having the backwards facing legs, it has forward facing legs like most, if not all other humanoid robots out there have. But one thing that we noticed in the video was a short snippet at the end of the video that unveiled Digit 5 was there was a wheeled humanoid in simulation operating in a couple different environments. I think one looked like a grocery store. I think the other one was a data center. And I remember you and I talking about this and I left a comment about this on my LinkedIn post about Digit5 was that it's curious that a company that has spent so much time pumping legged robots and talking about why this is the way, why they would showcase a robot that has wheels that just seemed to be contradicting what they've spent so much time and effort building.

5:06So that got my curiosity going. So I did what any journalist does. I simply asked them the question. I reached out to Jonathan Hurst. He's the co-founder and chief robot officer at Agility. I just asked him, are you guys releasing a wheeled version of Digit? And it's a little up in the air as to whether it's going to be a wheeled version of Digit or some other completely new product line. But they are, in fact, exploring different form factors, including wheels for future robots. So I'm just going to read you the quote that he sent me. Quote,

6:12leg design, which was a big moment for them, right? Now it seems like they're not beholden to just legs for every robot that they introduce. So I reached back out, I responded, followed up and said, Hey, is this, are you thinking about whatever this is, if it comes to fruition, is it simply a wheeled version of digit and you're keeping the upper torso the same with the two arms, the head and all the capabilities? Are you looking to maybe go down to a one-armed mobile manipulator in a completely different form factor, any arms. And Jonathan wrote back, quote, too early to say, I have guesses and inclinations, but we're really just starting, end of his quote.

6:51So interesting stuff, but it seems like they're at some point going to release new robots, different form factors that have wheels, whether or not that happens or not is obviously TBD, but that's some pretty important news, right? One of the leading legged robotics companies on the planet is at least opening the door to, hey, wheels make sense for certain applications, right? And you and I have, and I'll let you get in here in a second. You and I have talked about this a lot over the years on this podcast is that some of these companies get to beholden to their technology. They try to pigeonhole it into customers and environments where it doesn't make sense and ultimately kills their business.

7:30So now that we have the financials for agility, right? They reported it a couple of weeks ago in preparation for the SPAC IPO with Churchill Capital, is it 10 or 11? Agility made 1.8 million bucks in 2025, which is not great, right? I don't think any humanoid company is making a lot of money on the industrial front. Again, as we've said many times, it's apples and oranges compared to the unit trees and the agibots of the world that are really building an R &D platform. 1.8 million dollars is not a whole lot of money. They're burning. upwards of 140 million bucks a year or something like that. This was a smart move, right?

8:09If they want to stick around, make money, increase the stock price, create excitement, and ultimately deliver solutions, robotic solutions for customers to solve some customer pain points, they should be open to doing whatever they need to do at this point, right? That's my takeaway. So a pretty big moment, I think. And we'll have a story about this by the time this podcast comes out. You'll see a story about this on the robot report.com as well. But what's your reaction to this, Mike? Well, I think we got to come back to the fact that they've been one of their first humanoid companies to get real work at real customers and get paid for it.

8:48And you got to think that if any smart product team is going to be listening to those customers, And so I would have to guess that this is a highly educated pivot based on what, you know, their initial customers have told them. Again, that's what led to the leg redesign. And it's probably what's led to this opportunity as well. And, you know, good reasons for having both in the product line. I think, you know, as we're going to hear from Cy today, I mean, again, getting the manipulation and perception and all that stuff done. Once you get that done, it should be able to be embodied in whatever the physical aspect of legs or wheels or whatever the case may be.

9:34So I think that's really the lower of the hurdles to reach through in the future is getting the whole intelligence behind this thing to work once it works. But, you know, I think it makes perfect sense. I think it's a smart decision. To me, and he kind of says it in here, I'll go over this part of the quote again. And quote, as the safety, reliability, and capabilities of legged robots continue to expand, I expect they'll become the right solution for an increasingly broad range of applications. Meantime, we're interested in exploring other approaches that could help us bring robotics to more environments sooner.

10:09End of his quote. I think what he's saying there is that they're trying to bridge the gap. Yeah. Right. I think that's, again, I don't want to put words in his mouth, but I think that's what they're trying to do here is if you read the financials that they've released, they claim agility does. that it has, I think, a$300 million order on the backlog from a customer for the legged digit five version. They need to get there, right? They need to, and maybe this doesn't matter. Maybe this, again, they don't have a wheeled product that's anywhere near ready to be marketed. But can they somehow bridge the gap by telling people, hey, we're coming out with a wheeled solution that'll solve your problems in a much more efficient way.

10:55So it makes sense to me. So we'll see. We'll continue to report on this if they have more to share. Jonathan's going to be on the opening keynote of RoboBusiness where we're talking about the future of robotics. It's Jonathan, it's Rodney Brooks, and it's Ken Goldberg from UC Berkeley. I'm sure we're not going to focus this on wheels versus legs. I know how each one of those guys on the panel feels about that, but I'm sure Rod's going to bring it up anyway. And then Pross, the CTO of Agility Robotics is on the humanoid panel on the second day of the of Robo. That's my panel. Yeah. Yeah. Yeah. Which Mike's hosting.

11:33So all this stuff is going to come up in some way, shape or form at Robo business. But again, the big takeaway here is that agility is exploring other types of robot form factors, including robots that use wheels, which is really the big, the headline takeaway there. And again, I applaud them from that. I assume a lot of people are going to say, Hey, I told you so. Right. I get that. I applaud them for making this move. I can't imagine it was easy for somebody like Jonathan to go this route. And maybe it was, but he spent his whole career researching legged locomotion and legged robotics. I applaud them for making this decision.

12:13It's not the easy one to make publicly, right? It's the right decision. So I hope it works out for them. So kudos to them. Sticking in the humanoid realm, Mike, Boston Dynamics earlier this week opened up its Metaplant Application Center. This is to train its Atlas humanoids on site with Hyundai Motor Group, which is the parent company of Boston Dynamics. It's near Savannah, Georgia. and the center serves as a permanent test bed and data collection hub to train and integrate Atlas eventually into real-world automotive manufacturing and logistics tasks, again, on-site at Hyundai. Eugene Dimitri wrote about this over the weekend.

12:52Great story on the Robot Report, which you can read and we'll link to in the show notes. Your thoughts on this? Well, man, I'd love to go down and visit. That's my first thought. Sounds like a cool place in... Robotics Summit, Georgia. Here we come, baby. Yeah. And, you know, again, I think they're training using teleoperation and reinforcement learning and simulation. And so, you know, this is a hopefully it's a playground for doing all of that, both in simulation, out of simulation. and I think it should hopefully help push them down their roadmap as well because I think that's one of the next companies we haven't heard from in terms of the product roadmap.

13:36I'd be curious to see what's going to happen with Atlas. We had a stationary version of one last year or this year, I'm sorry, earlier this year at Robotics Summit and we asked why it's not moving and hopefully we'll be able to have one moving this year. We'll have to see. That'd be cool. Yeah, that'd be awesome. But this is, you know, I don't think we knew exactly when this was going to happen, but this was sort of the next step in their journey to eventually deploy. They want to deploy up to 25 ,000 Atlas units across Hyundai and Kia manufacturing facilities over the next number of years. They have never said that this is going to be done tomorrow, next year, by 2027.

14:19I think they said they're aiming for 2028 to have a manufacturing line that can produce up to 30 ,000 of these. So I think it's still obviously a number of years away. I think that's realistic. I think they've been nothing but realistic with their timeline. They're saying that this operation center, it's going to move to a different building next year. It's approximately 10 times larger, right? And expands into other testing areas other than just automotive. But this is exciting. You know, I think we've seen Boston Dynamics release some of these videos over the last maybe year or so where they're showing how Atlas is capable of learning through tele-op and other demonstrations and how quick they can do that kind of stuff now.

15:03I remember one of the videos came out where it was doing gymnastics. We've seen it doing the, what was the rowing move after the Norwegian soccer team. They're able to much, much more quickly now train it and teach it how to do new things. Those tasks appear to be quite simple compared to some of the things that I would imagine Hyundai eventually wants Atlas to do in its facilities. But they got to be on the ground. We always talk about how Tesla has an advantage, how Amazon has an advantage when it's developing its robots because it has that instant feedback loop. It has that real world working environment.

15:39This was a necessary step for Boston Dynamics to try to keep going in its journey to commercialize Atlas. So this is a great story. Maybe we'll get a tour of – we know some people down in Robo, Georgia. Yes. Ward, come on. Hook us up with a tour of the Metaplan. That would be great to see. That would be cool. And then finally, quickly, two acquisitions to talk about this week. We haven't reported on this one yet. Maybe by the time this airs, we'll have a story. I'm talking to the Cognex CEO on Friday morning. But this was a big one that happened. Cognex is acquiring RealSense. what, 14 months was it, Mike, after RealSense spun out of Intel as a standalone company.

16:20Here comes Cognex, a massive company. I think it made close to a billion dollars in revenue in 2025 here to acquire RealSense, which is making 3D cameras for robotic perception and navigation. I think it's, what,$500 million deal. So what's your take on this deal? Well, you know, Cognex is a name anybody in real manufacturing knows Cognex. They've been around as long as I've been in robotics, three plus decades, right? They were using Cognex cameras back in the early 90s. And the key there is that Cognex has made a name for itself in quality inspection, measurement, all of those types of operations on the manufacturing plant line, you know, measuring things as they're built, making sure everything's assembled, all those kind of things.

17:10So this is a really interesting expansion. They haven't broken into perception, right? They've never broken there. And so it's quite interesting that this acquisition brings them one of the leading technologies in the perception with the depth sensing cameras. And it's quite interesting. Also, it gives them a play now in physical AI. So I think it's quite interesting. I'm torn when I heard the news because, you know, we know lots of folks like Chris Matthew at RealSense. And I was really pulling for them to have a little bit more time as a startup to, you know, maybe innovate. I hope it doesn't stiple innovation, in fact, and give them an opportunity to breathe and innovate even more, because I think that's one of the reasons they were looking to get out of Intel was the innovation was sort of got long tracked rather than the way a startup should operate.

18:08So, and as you just said, Cognix is another big corporation. They really have built a name for themselves. They know how to build cameras. And so I guess from that standpoint, if they're going to, you know, open up the capability to build these things and to, you know, bring the cost out of it and bring the cost down, Cognix knows how to build cameras, right? So that's the good part of this whole acquisition. I think the real sense is expected to generate 80 to 90 million revenue in 2026. I think that's 50 % growth year over year. And so, you know, it's quite interesting. They say it should close here in the fourth quarter.

18:48So, you know, turn it around. So, yeah, I think it's an easy marriage. There's nothing there. It's just, you know, now Cognix has a, they call themselves a physical AI company. Yeah, I had the same initial reaction that you did. Like, why would RealSense do this? They just got out from underneath. Again, all the stuff, you know, I don't want to relive the story, but, you know, Intel has its well-known struggles. and it initially said it was going to shut down real sense we reported on this we had people telling us this was happening and then they got out there intel quickly changes its mind oh we're not shutting it down and like like it pissed a ton of people off it made a lot of people nervous and confused people left intel over this um and throughout that whole you know snafu when we're chatting with people over there you hear these discussions of it's too slow to release products It's too slow to get marketing going.

19:43There's too much red tape. Even despite its recent struggles, until last 2025, made$53 billion, right? Massive company. They're 53 times larger than Cognex. So it's still a way different company. But Cognex is a large company itself. I think it was over almost 3 ,000 employees, billion dollars. there can be red tape and problems and issues and, you know, politics that happen at a company of that size too. So, you know, I love the folks at real sense, right. They've, they've been at our shows sponsoring our stuff. Chris is great. Um, you know, so I have my ideas as to why they're doing this. You know, if you're making 80 to$90 million, which is 50 % growth year over year, I would think, and they have funding already from Intel and other, uh, funding avenues, I think they're a healthy company financially and it seems like they're doing really well.

20:40So I don't know why they would do this. I know there was some talk of when Real Sense was owned by Intel of it being spun out in a publicly traded company and people making tons of money. And they were pissed off, rightfully so, when that did not happen. So I think that probably comes into this as well for folks who have been at this for a number of years developing the RealSense product line. Yeah. I just hope that they don't run into the same snafus that they just got out from underneath of. Right, right. Because they have great products. The market's super competitive for 3D perception. There's a lot of low-cost Chinese players out there as well.

21:20But, you know, so again, I'm hoping to learn more. Again, by the time this podcast airs, maybe we'll have an update on that. I'm talking with the CEO of Cognix on Friday morning about this. So stay tuned for more. And then another one, I was excited to see this one. Qualcomm is acquiring Picnic Robotics. Mike and Dave Coleman, a good friend of the show over at Picnic Robotics. Yeah. Happy for Dave and the rest of the team. I assume this is a great deal for them. I mean, let's be clear. Qualcomm also wants to be a key robotics player. They make chip sets for autonomy and automation. so bringing Picnic into the family is great.

22:03I mean, Picnic was founded in 2015. They've been, you know, building a piece of technology called Move It, right, as part of the robot operating system, and, you know, they've agreed to, Qualcomm has agreed to buy them, for this technology and integrate it into its Dragonwing platform. So I think that'll be a great addition to Dragonwing, the movement platform, and help move them more into the center stage as well as a robotics technology provider. So again, they get to now take advantage of the vast sales force and channel that Qualcomm has. yep so picnic obviously works very closely with ross the robot operating system you should go watch i watched a little bit of it dave coleman gave some opening comments ross con launched yesterday in what toronto i think it is so the annual ross con global event kicked off in toronto yesterday dave was there announced this on stage and got into a little bit about maintaining that open source mantra moving forward um so go check that out if you haven't seen his comments, but yeah, excited for Dave.

23:22Picnic, they'll be at RoboBusiness as well. So we'll get all the juicy details from Dave then. Maybe he'll have more time. I don't think anything's changing based off what I know, but I know Dave does a lot of work on Space Ross as well. Maybe he'll turn his attention to Space Ross. I don't think so, but cool stuff there. Happy for Dave. All right, that's going to do it for the news of the week. And again, we have you folks covered with all the latest news and information about the Global Robotics Network with our websites, the Robot Report in Automated Warehouse. Make sure you follow us on social media.

23:52Make sure you sign up for our weekly newsletters so you never miss a beat. All right, Mike, now on to our featured interview. We're going to hear from Sai Vramprala. He's the CTO of General Robotics. Set this up for us. Yeah. So what I tried to do in this conversation with Sai was to explore the idea of robot agnostic intelligence, this idea that they're pitching about reusable skills, because I think that's interesting. I'm really curious, you know, how that works. And also the importance of enterprise data ownership, you know, with a Gentic AI type thing. So we're going to learn about their product called Grid that relies on this universal brain concept, combines specialized capabilities, perception, grasp, motion planning, etc.

24:39So let's hear Si's side of this.

24:46Right. On the show today, we have Sai Vamprala. He's CTO of General Robotics. Sai, welcome to the show.

24:54Sai Vemprala: Thank you for having me. I love the podcast and your work, and I'm thrilled to be talking with you today. Yeah. Hey, we're excited to have you on the show, and we're excited to learn everything that General Robotics is up to, because I think you guys are attacking a very interesting problem with robotics today and AI. and so I think we should start there. Some folks may not be familiar with General Robotics in the audience. I haven't heard of you yet. They should have, but maybe they haven't. So why don't you just start with an introduction briefly about yourself and then tell us about the overarching mission of the company.

25:31Sai Vemprala: Absolutely. Great to be here. So I'm Sai. I'm the CTO at General Robotics. And prior to General Robotics, I got my PhD in 2019 from Texas A &M, mostly working on robotics in the aerial domain at that point of time. So localization, path planning for drones. So it was kind of my body of work. And then I spent four years at Microsoft Research at the intersection of simulation, deep machine learning, foundation models, and robotics, where I've had the luxury of kind of exploring a lot of different approaches and techniques for creating what we thought at the time would be generalizable robot intelligence.

26:04Sai Vemprala: So that kind of segues into general robotics itself. So we started about three years ago. At the highest level, the overarching mission for general robotics is to make robots useful fast. We all know that the robotics landscape is a highly fragmented sort of ecosystem. The systems are extremely complex from an engineering point of view, and it remains a pretty hard problem in a full stack kind of a way, all the way from the hardware itself to the models and the techniques. And this is essentially resulting in a massive, I would say, innovation to deployment gap. There is a lot of research that's happening, but not a lot of it is actually seeing value in the production in the deployments.

26:43Sai Vemprala: So we started the organization to kind of help simplify the deployment of intelligent robots and really reduce the gap from what we think is the idea to execution gap by making the technology more capable and also more accessible and also help with the lack of potential expertise across all of the possible industry domains that are interested in doing robotics. So at the end of the day, we are a platform company that is pulling together pretty much all the tools that we believe are essential for creating robot intelligence. This could be hardware integrations, AI model integrations, simulations, data collection, pretty much every part of the tool chain that we believe is essential.

27:23Sai Vemprala: So our product is called GRID. It stands for the General Robot Intelligence Development Platform. It gives robots a shared intelligence layer where they can access different skills and the ability to create new skills. So roboticists, users and operators alike can come to the platform and essentially kind of break the bottleneck that's typically associated with robot deployment, move really fast from prototype to production at a pace that was not previously possible. Well, I think I hear a lot of stories today, tell a lot of stories today about companies that are trying to build one massive monolithic foundation model as a universal robot brain of sorts.

28:01I know a lot of humanoid companies are going on this approach, other physical AI companies as well. So tell me in your opinion, what are the challenges that that presents in your intelligence grid? How does that overcome it?

28:16Sai Vemprala: That's a great question. So when I think about the notion of a one-brain model, historically, I think even today, in a lot of situations, that notion of a one-brain model is tied to a single robotic form factor. So it's a vertically integrated system where, for example, the creators of the model have full access to the robot and they're limited to that set of sensors and actuators. And the challenge typically arises from different locations, different domains, and different tasks and so on. But even if we were to assume that as a single robotic form factor we're working with, the one brain model can only improve as fast as the one lab behind it can retrain it or gather data from all the scenarios and tasks that it's supposed to tackle.

28:57Sai Vemprala: And we all know how bad the long tail situation is in robotics. So that means in order to get to a really reliable, really generalizable model, we're looking at an immense amount of data and an immense amount of deployments to go with it because it's hard to collect all the long tail data kind of like a priori, essentially. And add to that now the complexity of having to work with several different robotic form factors or several different robots in a given form factor. The hardware ecosystem is expanding very rapidly. And at the moment, while this approach can still mature in the one brain model and provide immense value down the road, organizations need robots and value today.

29:33Sai Vemprala: And it is not yet there in order to scale to different robots and different domains very, very quickly. But what we are doing is we're actually taking a different approach and we are taking, I would say, fundamentally a modular perspective on the problem. So as opposed to training an end-to-end model where the bottleneck is the fact that you have to experience every single task and permutation yourself, we're saying, why don't we orchestrate the best of breed models that's out there? Why don't we leverage the entire field of AI innovation that has been driving computer vision forward, it's been driving planning forward, pretty much all the different categories that can really come together and provide, I would say, this modular intelligence fabric so that all of these models can get absorbed into our platform.

30:17Sai Vemprala: and by breaking the problem down into this, you know, the perception part of it and the action part of it and making it separate entities in the same kind of loop allows us to generalize much faster because we're no longer tied to the exact robotic form factor we're going to be working with. Even we sometimes don't know which robots we're going to be working with for a given customer or a given industry. But historically in our work, we've been able to transfer some of our skills almost in a, you know, zero shot, like out of the box way from one robot to the other and so on. So while we're not necessarily building that one giant model that kind of rules them all, we are harnessing the entire field of AI innovation, which I believe gives us a lot more compounding sort of gains.

30:57Sai Vemprala: And we are, you know, very quickly trying to absorb the pace of innovation that's happening out there. Yeah, I mean, I find that interesting. I mean, the whole concept of zero shot, everybody's talking about that is maybe a bit of a holy grail problem at the moment. And you're talking about this in a way that you're sort of breaking down skills into sub-elements that enable any robotic application or embodiment. Tell me a little bit more about your approach here to this idea. So let me take maybe a couple of examples to ground this. And, you know, the most common things that people think about when we think about robotics are, you know, manipulation, things like pick and place or, you know, navigation, which is an understanding of space and moving through it to go from point to point in a semantically meaningful fashion.

31:50Sai Vemprala: So even if I were to take something as simple as pick and place, if I have to pick up an object and place it in a target location, one approach is to teleoperate a robot or generate data of thousands of episodes that teaches the robot what pick and place really means. But at the end of the day, a lot of the existing architectures are really trying to memorize what is happening in those demonstrations that are being shown to the robot, and they try to imitate it, which is basically the imitation learning paradigm. And the argument there is that the more data we have, the more it would generalize.

32:20Sai Vemprala: But what we would do in Grid for something like that is that we take this task and we think of it as a recipe. What do we want to really do when we pick and place? We recognize an object. We recognize how to grasp it. And then we kind of semantically try to figure out where does it need to go if I want to pick up a pen and put it on a plate for whatever reason, right? So something like that. So what we do then is internally Grid essentially maps the fact that, okay, maybe I need visual descriptions of the scene and then semantic segmentation to ground where these objects are and grasp prediction and motion planning.

32:55Sai Vemprala: And I want to call out motion planning because it's kind of a very well-known, very well-engineered part of the pipeline, which has worked reliably for decades. And there is no reason to throw that away if you're working in kind of like the manipulation planning sort of domain. So rather than training a policy to do everything, Grid will try to compose all of these individual capabilities together. And an interesting thing to note here is that all of these individual models, such as segmentation or grass prediction or motion planning, they have all been advancing in their own way. And they all get better and better every single day with more data in their limited domain.

33:31Sai Vemprala: So the faster we pull in those sorts of advances, the more generalizable we become out of the gate. So combining all of this, what ends up happening is that because we are modularizing it into segmentation and object pose prediction or grasp prediction and motion planning, this approach actually scales out of the box to any arrangement of objects. And it also scales to any kind of object because these segmentation models have been trained with a lot more data than the end-to-end model would have been trained if you were to focus on a single task. So this also makes it robot agnostic because we are allowing the robots to understand the space in a more holistic way than saying you're going to get trained on your trajectories only.

34:10Sai Vemprala: And then good luck if you move to a different robot, something like that. So this decomposition is what makes it really robot agnostic. And it also makes us allows us to build reusable, composable little units that will all accrue to larger and larger kind of purposes, I would say. And similar same thing goes for navigation as well. We try to break down navigation into mapping, planning, object recognition, semantics, and all of that really comes together to allow any robot, regardless of a humanoid or a wheel or a mobile manipulator, to really, again, holistically understand the space and understand what it means to move through it and experience it.

34:46I guess what I'm having a problem understanding in it, when you talk about being truly robot agnostic, somehow, again, equipping a robot with a camera and being able to look at the scene and see something that you can pick up, but then to translate that to a five-fingered humanoid hand or a two-fingered gripper, parallel gripper on a cobot, the kinematics are certainly way different. You've got to make that connection. Is that a little more than just a setup, a definition, a transfer function of some sort that maps the kinematics to what we can then perceive and act on?

35:27Sai Vemprala: Absolutely. I think that's a great segue into maybe a little bit more detail into how grid treats robots at the fundamental level. So there are several layers in grid. I'll maybe focus on the first layer, which is essentially what we call the robot ingestion layer. So we treat the robot ingestion as a very fundamental and important component in grid, because that is what allows grid to understand the kinematic structure of the robot, the constraints of the robot, what kinds of end effectors are compatible with this robot, and so on and so forth. So, for example, we talked about robot agnosticism from a pick and place kind of perspective, and you bring up a great point, like how do you go from a two finger to a three finger to a five finger, for instance, right?

36:08Sai Vemprala: because we are breaking down the problem into, again, perception and grasping and then subsequent motion planning and so on. By understanding what the robot is and is not capable of, I think we also have the luxury of kind of adapting the skills. And like grasp prediction can be very different for a five-fingered hand versus a two-finger gripper. But as long as we are aware of the fact that we're working within this constraint, we can either tune those models to be more adaptable to a five-fingered grasp versus two fingers. Or maybe it's a completely different model. Like, you know, two-finger grasp models are different from five-finger grasp models.

36:40Sai Vemprala: And out there in the research, we have several of those today which have been trained specifically for grasp prediction. And they can be a very, again, a composable unit in the pipeline. And we can pick and choose. Grid can say, oh, if you're using a two-finger gripper, there are leading models to this grasp gen out there that's a part of our stack. And then if you're using a five-finger gripper, there are other models that have been very recently released on that. But we also have the luxury of minimally fine-tuning some of these models now that we know what the robot is capable of and what this end-effector configuration kind of looks like.

37:12Sai Vemprala: So I would say all of this really ladders up to kind of like a closed-loop approach where we have the ability to deploy certain models out of the gate and then get an understanding of how they perform potentially in simulation as a starting point. And then we can choose between different models. Grid can help the user benchmark and evaluate multiple models, even within a sub-component. Or if none of them actually do the trick because your objects are very niche, or maybe your robot is fairly non-standard, in that case, we do have the luxury of training some policies or adapting or fine-tuning these models, if that makes sense.

37:45Interesting. So let's continue down the rabbit hole of Grid here and trying to understand, and I'm trying to put myself in the ears of our audience who might be trying to consider whether grid is the right solution for their robot. Where does grid run? Does it run in the cloud? Does it run locally on the edge, split different functions in different places? Talk a little bit more about the architecture of this And then, you know, the combination that you need from simulation, training, deployment, those kind of things. So folks can put this in their mind about where you fit into their potential architecture.

38:30Sai Vemprala: So as an overarching sort of description, I would think of grid as a kind of a shared intelligence fabric that robots can connect to. So it is actually primarily deployed on the cloud, but we have a very distributed sort of paradigm. So it's not limited to only the cloud. It can also run on, for example, data center GPUs that are on premises for a certain facility or some kind of a network edge or things like that. But we do believe in separating the workload so that not everything is running on board the robot just because of the compute constraints of the robot. So in practice, the way grid runs is kind of like a system one system two architecture.

39:07Sai Vemprala: So we have the higher level, heavy reasoning, development friendly sorts of parts of the platform that are running on the cloud, which is all accessible to APIs. So this can be model inferences, this can be simulation, this can be policy training. These are all things that you want to do before you get to deployment, like when you're prototyping, what kinds of models do I want to use or what permutation of the skill would apply for me? So a lot of that runs on the cloud. And then on the robot, we have essentially some lightweight clients that are either acting as interfaces to the cloud platform so they can receive, you know, send sensor data up, receive some model inferencing and then actions back and things like that.

39:45Sai Vemprala: But they also run safety critical operations, such as obstacle avoidance or in the loss of connectivity, like what do you do as an emergency maneuver? That logic still runs on board the robot. So we do believe in having lightweight workloads on the robot because, you know, for emergency or safety kinds of situations. So if I were to dive a little bit deeper into how exactly is the grid stack set up, I would think of it as four primary components. There is the robot layer. I mean, we talked about robot ingestion briefly before. But Grid, we actually pre-integrate with several leading OEMs in the market.

40:19Sai Vemprala: So a lot of the representations of manipulators, humanoids, quadrubits, most of the popular OEMs, if you've heard of it, it's very likely integrated within Grid. So we try to maintain representations of the physics, of the sensor placements, of the sensor characteristics, and the visuals and so on and so forth of these robots. So that's kind of like layer one. So Grid has a very good understanding of how the robot is supposed to be working, how it's set up. In many cases, we also go all the way to the system identification level where we ensure that the actuators and the configuration is kind of matching what the robots do in real life, which is pretty essential for policies, for example.

40:55Sai Vemprala: And then the second layer, we call this essentially like a world experience layer. It's primarily driven by simulation, but can also be augmented by real data. So now that you have a robot, you have all the sensors and the control interfaces and so on, where does it go? Like, where should it live? It depends a little bit on the task of interest that you're trying to solve. So you can put this robot maybe in a warehouse or maybe in some underground mining facility. If it's a navigation robot, maybe it's an aerial robot flying for delivery missions or something of the sort. So we have a very extensive simulation stack that can allow us to simulate, you know, areas of interest, you know, locations, representative facilities like industry facilities and so on.

41:36Sai Vemprala: objects, scenes, assets, all of those things are kind of like very formally taken care of within grid. So, you know, customers can bring in assets of their own, their SKUs. They could also do digital twinning if you want to represent a factory and tie factory within grid. So that is a fairly rich layer, which really grounds the robot in the experiences we want it to have before it gets to deployment, essentially. And then the next layer is where it really comes to life, the skill creation layer. So skill creation, again, this is where we don't necessarily take a bet on a single approach, but we would much rather allow the roboticists and the users to really test and evaluate multiple approaches and pick what fits the bill.

42:15Sai Vemprala: So we allow zero-shot composition, like we briefly talked about earlier. We also allow for reinforcement learning, data collection, and policy training. All the well-known paradigms are definitely available through GRID for in evaluation and policy or skill creation. And last but not the least, we have the deployment layer. And deployment is, again, it's a very critical thing for us because we want to be a very deployment first kind of platform. So we want to close the gap from the time you get the robot to the time you actually deploy something, even if it's like a representative skill, because the sooner we close that gap and the sooner we are interfacing with the actual physical robot, the sooner we can establish that flywheel that is allowing us to evaluate, generate more data at deployment time, and then use the information about the failures and so on to really drive back into the simulations we're doing or the policy training that we're doing.

43:08Sai Vemprala: And I specifically mentioned that because I know there's a lot of discussion about what kinds of data is the most useful in robotics. And for the most part, I really don't think it is just the volume of the data, even if it's deployment, but it's rather the signal in the data. It's like the signal to noise ratio has to be somewhat higher, right? So we believe the deployment data is something that should be taken and used as interventions within simulation. So you can actually go simulate what you're seeing as a long tail failure, expand upon that, and maybe increase the performance and so on. So these four entities, you know, robot ingestion, the world experience of the simulation, the policy and the skill creation of deployment, they all work pretty much in lockstep within grid to establish a pretty close loop pipeline rather than a one way pipeline.

43:54Sai Vemprala: So all the evidence from deployment kind of flows back, adjusts the simulations, adjusts the skills, and then you're continuously going back and forth. So I just want to replay what I think I heard you say. and I'm trying to visualize in my mind, if I was a developer that wanted to build a new application and I wanted to pull from the current segment of humanoid robots from a potentially evaluating, maybe an open system, more open system like a Unitree G1 or all the way to, I don't know, maybe more closed robots, I would think of like a figure or Tesla or Boston Dynamics, for example, at the moment.

44:35Is your intention that you have fitted all of those robots with access to grid at some point in the future and be able to say, hey, you've got this robot model is now 100 % compatible. You can go ahead and deploy grid on it where maybe some other robots, or if I have a custom built robot, I'm working with one of those manufacturers that I want to deploy, I've got a little bit more work to do with you to make my custom built bot compatible with grid.

45:09Sai Vemprala: Yeah, that's pretty, that's actually spot on. It's pretty precise. So some of the robots that we integrate with out of the box are kind of like some of the leading manufacturers out there today. So those are almost deployment ready. So all the control interfaces are mapped, all the sensor information are mapped. So it's almost like, it's almost akin to like you get an API key and then you are directly kind of accessing some of the perception models or the skills. You know, we train a lot of skills ourselves in-house using reinforcement learning or VLA kind of paradigms and so on. So there will be a certain set of out-of-the-box skills that, you know, folks can start working with.

45:42Sai Vemprala: And of course, they're always, you know, they can leverage things like our teleoperation pipelines, data collection pipelines to train new things. But the deployment is definitely unlocked for those robots. But if you have a custom robot or if you're a new OEM or something, for example, you know, Typically, they work with us to understand what this template kind of looks like to be a part of grid. And then I think just like any good software insuring kind of a problem, the better defined your contracts are, I think, like inputs and outputs and so on in structured fashion, the faster it is to kind of expand your portfolio of devices or robots and so on.

46:17Sai Vemprala: So we have a standard method through which we kind of onboard robots into grid. And then once they're connected, all the robots have access to these navigation workflows, if they're humans, for example, or manipulation workflows and so on. Very interesting. Okay, so that sort of covers the question if I was an OEM or wanted to build a new application. I think the other interesting thing I wanted to ask you about that I would envision, are you thinking of like a skill set marketplace at some point in the future where more skills I think will come into the system, but can they be developed by third parties?

46:50Maybe even someone makes a business around positioning skills that they create into the general robotics ecosystem?

47:00Sai Vemprala: Well, from a technical point of view, it's definitely possible. I mean, for example, the way we work with enterprises now, I mean, they're kind of like our target customer audience. I mean, they are large organizations, right? So and then they have internally, you know, typically they have several teams that are, you know, looking to subscribe to certain kinds of skills. and they would go deploy them in maybe an inspection point of view versus a navigation point of view versus something else. So typically the process is that there will be some developers within the enterprise, within the industry, that would be working on A, maybe evaluating existing skills that are already in grid, and then B, then building upon that to create new skills that are more in line with their operational context or their constraints and so on.

47:42Sai Vemprala: And then they would have their own customers in-house, essentially, of those skills. So they would go use those skills and deploy them. Could be operators, could be other developers and so on. So I think we're already seeing that paradigm somewhat in the enterprises that we work with. And it's no different than I think, Mike, what you described as a marketplace, right? So we allow the users, if they're roboticists and developers, to obviously go through the documentation. They can very quickly create new skills, especially in the navigation manipulation domain. And because everything is kind of like written as a composable unit, you can always, you know, you can combine navigation with manipulation, right?

48:18Sai Vemprala: That's kind of what robots do at the end of the day, right? So somebody might be, some nav team in your company might be using grid to create a navigation, you know, object recognition and navigation kind of skill. And then somebody might be creating a VLA for picking up certain SKUs once you get to the destination and somebody could wire them together very quickly. And one thing I have not mentioned earlier is that one thing that caps all of this off inside of grid is actually an agentic layer that allows even non-technical operators for example to compose skills that somebody has written so that's kind of like your marketplace idea but now with like an mcp and agent on top uh so we provide mcp servers that can parse through all these skills and tools that are in grid either user created or gr created and they can on the fly start composing and commanding these robots to do certain things so we're already seeing a very similar paradigm and I think the customers that we work with today.

49:13Yeah, well, so I'm curious, again, things are happening so fast that, you know, things we didn't think were possible six months ago are now possible and that window keeps closing. And I'm curious from the seat that you sit in, in the industry and in the company, you know, what are some of the milestones or what are some of the things that have changed recently in this whole AI world? Even you talked about the agentic AI growth as well to make this approach that you're taking with general robotics possible now.

49:46Sai Vemprala: Yeah, that's a really good question. I would say, yes, agentic AI has been very transformational in the industry. I think we are all seeing so many flavors of it on every single day. At the end of it all, I think we now have access to LLMs and agents that can do an excellent job at kind of like engineering problems. So they're doing well at coding, they're doing well at math. And we are, again, our perspective is that robotics, first and foremost, is a complex engineering problem. And the reason it has been somewhat inaccessible in the past is just that, of course, the fragmentation itself, but also this lack of, I would say, structure in necessarily creating a single platform that unifies all the entire tool chain that we need for robotics.

50:28Sai Vemprala: So what we realized over the few years that we've been working on this problem is the better you architect these individual pieces and the more cohesively you put them together, I think the more naturally we realize that agents can take up each of these functions and then really advance how quickly we can, for example, we can ingest new robots or how quickly we can pull in new models and how quickly the platform is able to evolve. So it's all about designing this entire structure we talked about, the four-layer kind of structure, in a way that's agent-first. And that's what we've done at General Robotics.

51:02Sai Vemprala: So we actually termed this approach auto-engineering. So what we're doing now is we have given access to agents each and every piece of grid. And then it's reaching a threshold where essentially the agents are allowing grid to engineer itself. So if you give it a task today, it can assemble or suggest the right robots and the right models. It can build the simulations. It can receive some input from the user on what objects that are interested in or what kinds of worlds are they looking at. I mean, we are seeing, again, a lot of advances in, for example, the latest GPT models, which are really good at 3D reasoning and structured kind of semantic thinking and so on.

51:40Sai Vemprala: And we are seeing a lot of that help with grid also. And they can go reuse all of these engineering pipelines they've written with obviously good documentation, good structure and so on to build simulations, to even train new skills by themselves, evaluate them, deploy them. So overall, I think, of course, the human still has to provide the intent and some of the agencies coming from them. But we are seeing, you know, what we want to do for robotics, what harnesses and agents really did for coding and math. And through Grid, we see that we have a really good shot at creating this entire loop that can be very autonomous.

52:17Sai Vemprala: And then it also has the right checks and balances at each segment because everything is a harness where the agent is learning from its outputs and refining itself for the next run, essentially. Well, that sort of leads to my next question, which I wanted to sort of get you on the record for, which is what's your policy for data ownership and enterprise trust here? What's the corporate policy for your clients in terms of the use of customer data for specific applications? That's a great question. So the way we've architected grid is that grid is primarily a provider of capabilities, right? So the capabilities can be subscribed to by the users and they are free to train their own skills or kind of like use existing skills with their robots, their sensors and so on.

53:04Sai Vemprala: So as they go through the process, our customers, they retain complete ownership of the data of any custom skills they've created. So we essentially want to work, I mean, we want to work in a fashion that, you know, the organizations that are working with us, I mean, they are able to create their own robotic strategy, right? So they are able to retain like their own competitive advantage in the domain that they work with and work on. So yeah, we don't use any of their data ourselves. I mean, all of the data can live even within on-premises, depending on how grid is deployed to them. Most of the deployments we do today are also on kind of private tenants for the specific enterprise that we're working with.

53:44Sai Vemprala: So, yeah, and then every capability kind of resolves into these reusable skills that they can inform using their own custom data, private data, and then whatever fine-tuned versions of the skill that they create, they can own all of that. So are you saying that if I train a skill, teach a skill, compose a skill for a specific application that the grid doesn't absorb that understanding into its greater consciousness in a way that it could know how to do that in the future if another customer attempts to do something similar? I mean, again, here's your balancing between sort of generalizable skills, picking, observing, that kind of thing versus, you know, capabilities.

54:30I think it's sort of equivalent to we all know how to walk and ride a bike and what have you. But for me to be able to, you know, do construction, I have to know how a house goes together or whatever. And if I was trying to build a robot to build a house and I'm teaching it how to put up walls and what have you, and I want to maintain some competitive advantage in that skill for that robot that's a construction worker or helper, right? And I don't want other people there. But yet you, in the big picture, want to know how robots carry hammers, hit nails. But the point is, what can I protect or what can I keep as mine that doesn't get absorbed into the greater system?

55:15Sai Vemprala: Absolutely. I think you put it the right way in a sense that, again, taking this somewhat of a modular paradigm. For example, if we have any of these LLMs in the loop today, I mean, the LLMs are generally aware with common sense on how certain tasks are being done. So similarly, grid has a general awareness of how, for example, hammering a nail needs to be done. So for instance, but what we don't ingest, what we don't have to ingest are exact appearances of the objects that are being used in industrial context or some specific operational kind of protocols or manuals and things like that. Maybe the 3D representations of proprietary objects and so on.

55:52Sai Vemprala: So we don't ingest any of that in the sense that we don't keep track of any of those. They're all owned by the customer. But the general know-how and the knowledge are typically encoded by a lot of these large models. So they operate in a very abstract, common-sensy kind of space, if you use that word. But when I mentioned the data that is typically owned by the customers and that does not flow back into grid would be actual videos or actual demonstrations from their facilities, their objects, their images. That can all stay within the purview of the organization so they can maintain their sovereignty, I guess.

56:23Before I let you go, I want to also touch on a very important aspect of the company, which is that a lot of it came from Microsoft, Microsoft Research. And obviously, the concept of building a large scale system that supports needs, I see a lot of parallels here to some of the Microsoft software philosophy, whether you're building Office or other tools is a similar analogy, I suppose, in the decades ago. So tell me a little bit about the genealogy with the team coming out of Microsoft and how that's infused into the organization and how you approach the problem in the delivery of the solution.

57:04Sai Vemprala: Absolutely. That's a great question. There is a lot of things that I think that definitely inspired our vision, our execution from our Microsoft days. So I, for example, I worked for a few years as a senior researcher in the autonomous systems and robotics research group. I was mostly focusing on research at a fundamental level on the idea, somewhat abstract thoughts on how robot intelligence should be created. Our CEO, Ashish Kapoor, was the one who actually led that group. And then he was at Microsoft for about 16 years. And one thing that both of us worked on was actually a tool called AirSim.

57:39Sai Vemprala: This is a leading robot simulator that was built on Unreal Engine, pretty widely used across the board for drones and real robots and so on. It's an open source library. And then it kind of hints a little bit of our roots that we believed, you know, it's important to give robots these sandboxes, you know, in which they operate and they fail and they evaluate. And then, of course, it was all somewhat done by humans back in the day. But it's also applying to our agentic or the auto ensuring efforts even today. So we are big believers in how those evaluations and harnesses should be set up, for example.

58:10Sai Vemprala: But coming back to the infusion of the Microsoft philosophy, one thing it's taught us is to not to shy away from building large-scale software, naturally, because we believe that's what robotics needs. But also, we had the luxury of actually learning from a lot of the enterprise customers that Microsoft was working with back in the day, into what are the automation challenges? What are the robotics challenges the field is facing today? And it was specifically interesting for me because I was in a research position at that point of time, And then I could see the gap between the wavelength that we were thinking at versus what the industries and the enterprise are actually struggling with on the ground because it was somewhat even like very simple things were just not accessible to those industries.

58:52Sai Vemprala: And that led us to really think of it from a platform standpoint and the accessibility standpoint, as opposed to just another model that's just kind of like a research advantage rate. So this is why we are very much focused on shipping capabilities that have clear paths to deployment, a clear path to value today. And then once we get robots into production, which is what we're just kind of built around, not winning benchmarks or having two percentage points extra on something. But then we have the luxury of learning more about that industry and then refining the approaches and getting better and better.

59:25Sai Vemprala: So yeah, overall, it's a mix of, I would say, large scale software, kind of like ethos, but also the startup speed, naturally. That's why general robotics is pretty nimble. and also the fact that we want to be deployment first and work with large enterprises in a way that enables their strategy to succeed. Well, great. It's been very interesting conversation to learn more about general robotics. Before I let you go, I just want to ask, I mean, you're shipping today in the sense that anyone that wants to begin to interact with grid, how can they access that today to begin playing with it, learn more about it?

1:00:06Sai Vemprala: Yeah, that's a great question. So we work directly with enterprises, large enterprises today, but we also work through several partners, including OEMs, through which we kind of give their own customers access to the grid platform. And then once the customers, the users have access to the platform, the entire kind of like roster of capabilities are accessible through agentic means or kind of like a command line interface and so on. So very easy to get started kind of a tool. So we believe in the low barrier, high ceiling kind of approach, but that's how one would kind of maybe get started with using Grid.

1:00:38Great. Well, Cy, it's been a pleasure to talk to you and look forward to following the growth and success of general robotics in the future. So thanks for joining us today.

1:00:48Sai Vemprala: Thank you very much, Mike. Thanks for having me. And it's great chatting.

1:00:55Awesome stuff. There's so many ways to pull those models together and General Robotics obviously has their own approach. It had an interesting quote in there. Rather than training a policy to do everything, Grid will try to compose all of these individual capabilities together. Do you think that composite approach is the right one, Mike? Certainly. I don't know how they're doing it. I think that's the thing I'm trying to wrap my head around is based on what little engineering skills I have left. But I mean, the point is that, you know, the ability to, you know, bring these things together that are, you know, all AI based and, you know, for a complete end model.

1:01:36It seems like black magic to me, but obviously the science is moving forward so fast I can't keep up anymore. And, you know, I tried to, as I asked them, are they going to build sort of a app store like thing? And that's not really their vision yet. in the early days, but, you know, the ability to have different types of sub-models that you can put together in the skill-like piece for whatever embodiment you're going to put together, I think that somehow wraps up their secret sauce. So, and splitting workloads between the cloud and the edge, you know, I think that that makes sense as well. How they're going to do that efficiently is also left to be seen.

1:02:20But again, that's their overall architecture. So we'll continue to follow them and see how this technology matures. And it was a fun conversation with Cy. Yeah, great stuff. Go check out General Robotics if you want to learn more. That's going to do it for the show. Just before we sign off, again, one final reminder about RoboBusiness coming up October 20th and 21st out in Santa Clara, California. This is our 20th anniversary show, which we're super excited about. We'll have over 2 ,000 people and 100 exhibitors on the shelf floor. And again, this is the moment everyone's been waiting for. We're going to throw a little drum roll into post, Mike.

1:02:56Enter discount code ROBO15, R-O-B-O 15. At checkout, you'll save 15 % off your tickets for RoboBusiness. Again, over 2 ,000 people from the global robotics ecosystem are going to be there. It's a great show. Super excited about it. We'll have over 70 speakers on our stages over the course of a couple of days. So we hope to see everybody out at Robo Business October 20th and the 21st. Head to robobusiness.com to learn more and enter code ROBO15 to save 15 % off your tickets. If you have any questions about the show, reach out to any one of us here at The Robot Report. But that's going to do it for episode 263.

1:03:36Thanks again to everybody for tuning into the show. Thanks to Syver and Prowl, CTO of General Robotics, for joining us on the program today. new episodes of the Robot Report podcast drop each week you can find us anywhere you listen to your podcast please subscribe, leave us a rating, give us a review this episode was produced by Mike Oitzman edited by Tyler Kubisti, our executive producer is Amanda Buhner, I'm Steve Crow, thanks so much folks for tuning in, have a great weekend and we'll talk to you again next week