In today’s rapidly evolving world of artificial intelligence and automation, the most exciting breakthroughs aren’t about replacing humans — they’re about augmenting us.
That’s the core message from Anthony Jules, Co-Founder and CEO of Robust AI, who joined Brian Bell on the Ignite Podcast to explore how the next generation of robots will collaborate with humans in the workplace.
A 30-year tech veteran, Jules has built a career at the intersection of software, AI, and human systems. From helping grow Sapient Corporation from three people to over 4,000 and taking it public, to leading robotics initiatives at Google, his perspective is shaped by both technical expertise and a deep understanding of human behavior.
“The future of robotics isn’t about pure autonomy — it’s about collaboration. The best systems use people for what they’re good at and robots for what they’re good at.”
— Anthony Jules, Robust AI
From Trinidad to MIT to Tech Pioneer
Anthony’s story begins in Trinidad and Tobago, where he first started programming at age 11 and built his first video game in his teens. His curiosity and drive led him to MIT, where he studied AI and robotics long before they became buzzwords.
After graduating, he joined the founding team of Sapient Corporation, a company that helped Fortune 500 firms build some of the earliest client-server and internet systems. Sapient scaled to thousands of employees and went public — an experience that gave Anthony a front-row seat to the dynamics of hypergrowth, organizational culture, and systems thinking.
“Things that use feedback loops perform better than things that don’t,” he explains. “That’s true for organizations and for robots.”
Returning to Robotics: From Redwood to Google
After Sapient, Anthony returned to his first love — robotics. He co-founded Redwood Robotics, a company that designed collaborative robot arms that were safe and easy for humans to train. The startup was later acquired by Google, where Anthony worked on next-generation robotics and machine learning systems.
Inside Google, he saw firsthand how deep learning was years ahead of what most of the world thought possible — and it changed his perspective forever.
“At Google, I realized we could build incredibly capable robots — but I also saw that full end-to-end AI systems lacked transparency. If we want robots to operate in the real world, they need to be modular and explainable.”
That belief — that robots must not only perform well but also explain why they act — became the foundation for his next venture.
Founding Robust AI: Building Smarter, More Transparent Machines
In 2019, Anthony teamed up with robotics legend Rodney Brooks (founder of iRobot and co-inventor of the Roomba) to create Robust AI. The company’s mission: build collaborative robots that are intelligent, flexible, and fundamentally human-centered.
Their flagship robot, Carter, looks like a smart mobile shelf — but under the hood, it’s a powerful AI system. Using eight cameras and advanced computer vision, Carter can see and understand its environment in 360 degrees. It navigates safely, collaborates naturally with humans, and reduces wasted motion in logistics and manufacturing.
Unlike traditional robots that rely on LiDAR or pre-programmed paths, Robust AI’s system uses semantic understanding — recognizing people, boxes, and forklifts instead of just obstacles. This allows for fluid, adaptable movement and safer collaboration on busy warehouse floors.
Why Collaborative Productivity Is the Future
Jules calls Robust AI’s approach “collaborative productivity” — the idea that pairing human intuition with robotic precision delivers the best results.
“Getting to 99% autonomy might cost you X,” he explains. “But getting to 99.9% costs 10X more. That last 1% can often be done better by people — and you gain human oversight in the process.”
This philosophy has resonated with major players like DHL Supply Chain, Robust AI’s landmark customer and the world’s largest logistics provider. Their systems help workers save time, reduce walking distances, and handle materials more efficiently — all while maintaining human agency and control.
The Ethics of Physical AI
As robots become more integrated into workplaces, ethical questions follow. For Jules, the answer lies in transparency, trust, and design.
He emphasizes three pillars for ethical robotics:
Safety – ensuring both real and perceived safety in human-robot environments.
Transparency of intention – making it clear what the robot is doing and why.
Honesty about impact – acknowledging when automation reduces labor needs and managing that transition with empathy and openness.
“No one objects to being more efficient,” Jules says. “But we must design systems that are not just productive — they have to be humane.”
Looking Ahead: The “Rainforest” of Robotics
When asked about the future, Anthony doesn’t see a world dominated by humanoid robots. Instead, he imagines a “rainforest” of diverse machines — specialized robots that work together like different species in an ecosystem.
From autonomous carts and robotic arms to AI-driven sorters and conveyors, the next wave of robotics will be modular, interoperable, and context-aware — designed to interact with both humans and other robots.
“It’s not robots replacing humans. It’s robots working with humans — and with other robots — to make the world run better.”
Staying Grounded Amid Exponential Change
Despite leading a company at the frontier of AI and robotics, Jules remains grounded through meditation and reflection.
“We live in a world that systematically destroys attention,” he says. “Meditation helps me reclaim it — it’s about staying centered so you can keep seeing clearly in the middle of change.”
Final Thoughts
Anthony Jules represents a new kind of AI leader — one who believes the future of automation lies not in replacing people but enhancing human capability.
His work at Robust AI is redefining how we think about productivity, ethics, and the relationship between humans and machines. As AI continues to advance, Jules’ message rings clear: the most powerful technology is the kind that makes us more human.
👂🎧 Watch, listen, and follow on your favorite platform: https://tr.ee/S2ayrbx_fL
🙏 Join the conversation on your favorite social network: https://linktr.ee/theignitepodcast
Chapters:
00:00 Intro & Origin Story
01:27 Founding Sapient Corporation
03:00 Lessons in Leadership and Culture
05:43 Transition to Robotics
06:12 Google Acquisition and Lessons Learned
09:00 Inside Google’s Robotics Vision
11:45 Leaving Google to Start Over
13:55 Founding Robust AI with Rodney Brooks
15:50 Early Challenges and COVID Pivots
19:00 Human-Robot Collaboration
22:00 Cameras vs. LiDAR
24:44 Sensor Debate: Tesla vs. Waymo
26:59 The Robust AI Tech Stack
30:07 Collaborative Productivity
34:14 Real-World Deployments
37:18 Powering Robots with NVIDIA
39:22 Practical Robotics vs. Humanoids
42:40 Partnership with DHL Supply Chain
46:15 The Next Five Years of Robotics
49:33 Synthetic Data and Simulation
51:07 Scaling Phase and Phase Shifts
53:02 Reflections on Growth and Systems Thinking
Transcript
(00:00:00)
Welcome to the Ignite Podcast, where we explore the intersection of startups, technology, and innovation. I’m your host, Brian. Today, I’m joined by Anthony Jules, co-founder and CEO of Robust AI. We’re going to dive into his fascinating journey—from MIT to Sapient, to Google, to building robots that redefine human-robot collaboration. Anthony, welcome to the show.
(00:01:06)
Thanks, Brian. It’s my pleasure to be on the show.
(00:01:21)
Yes. So I usually always kick off each episode with pretty much the same question. What’s your origin story? Tell us about your background.
(00:01:28)
Okay. So unfortunately, I have to go really far back in time for that one. I’m originally from Trinidad and Tobago, moved to the U.S. when I went to MIT, but started programming computers when I was 11 and made my first video game in my teens, then came to the U.S., I studied AI and robotics at MIT. Went on from there to be part of the founding team of a company called Sapient Corporation and helped build that from three people when I started to almost 4,000 by the time I left and helped take that public. And then came back to robotics, which is the focus, if you will, for really all of my life is about how people and machines interact in meaningful ways. So came back to that, have done a couple of startups, and I’m sure we’ll get into that in the next few minutes here.
(00:02:18)
Yeah, amazing, amazing career arc. For people who don’t know what Sapient is or was, maybe you can explain what you did and what you learned scaling to 4,000 people there and all the roles you probably held over the years.
(00:02:29)
Sure. So Sapient Build, large information systems for Fortune 500 companies. But the timing was such that when we started off, no one knew how to build client server apps. And we did that when kind of no one else could. And then we had an incredible run because we were a several hundred person, highly technical firm that understood the internet and saw the signal on it long before anyone else did. So we then helped typically Fortune 500 companies build their first websites, build their first websites that did transactions, which was a huge deal back then because there was no infrastructure. And we did everything from banking systems to building the backend for internet service providers and really built a bunch of the infrastructure that real companies used.
(00:03:21)
Almost like an on-prem AWS back in the day.
(00:03:24)
Yeah, there was no cloud back then.
(00:03:26)
Yeah.
(00:03:28)
Yeah, so everything was on-prem. And occasionally you’d have these things called co-location facilities where you could put a machine near to some trunk on the internet. But even those were pretty rare.
(00:03:38)
That’s amazing. And then what are all the roles that you held there? I mean, at employee number three, I mean, you probably did everything.
(00:03:43)
Yeah, so I, of course, wrote software when we started. I helped deliver the first thing that we ever shipped and then managed client accounts, ran large teams of developers working against customer contracts, and also then did strategy for customers. And then at one point in one of our leadership meetings, I went on a rant about how important managing intellectual, managing human capital was. So I ended up running what we call people strategy for a year and a half and figuring out recruiting, training, how we actually build the next generation of leaders inside of the organization.
(00:04:22)
That’ll teach you to open your big mouth in a meeting, get under a soapbox.
(00:04:25)
Yeah, well, actually, I think it’s a great thing when that’s part of a culture, which was part of our culture then, which is don’t complain about anything that you’re not willing to fix. And we really lived by that. I think that’s a great thing to have in a culture as a company because it means it allows you to constantly use feedback to improve internally. And I just think that that’s a process that’s critical, whether you’re building a company or designing a computer system or building a piece of mechanical hardware. Things that use feedback loops where you see how something performed and then you make changes and then you try again. Things that use feedback loops perform better than things that don’t.
(00:05:03)
Systems thinking, basically. Donella Meadows’ book, Thinking in Systems. I remember reading that. It must be 15, 20 years ago now. Really think about systems and everything I do, what is the system here and what is the outcome and what are all the positive and negative feedback loops and it’s a really good lesson I think.
(00:05:19)
Yeah, it’s something that I’ve just held close really through my whole career. Yeah, and then the other part of course of Sapient was growing this incredible hyper growth company and seeing the inflection points that companies go through. You know we had one vibe when you’re 25 people, a different one when you’re 80, really different when you’re 400 to 600 in six or eight offices at that point. And then, of course, quite different when you’re 3000 plus. And you have to change your communication mechanisms. You have to change your priorities. You have to change your org structure, kind of at each of those inflection points.
(00:05:59)
Kind of the internal ceremonies, you know, how do you manage goals and OKRs and cascade those and all that shifts. I’ve only been on a company like that once. I was Rocket Fuel and the ad tech company was doubling every year in headcount and revenue. And we’re just, you know, pretty crazy to watch. It’s like a new person every day walking around the office. And then so after that, you moved on to Redwood and then Google. Maybe you could walk us through that transition.
(00:06:21)
Yeah, so Redwood Robotics was a small robotics startup that we were building a collaborative robot arm. And what that means, it’s a robot arm that is safe around people and someone can actually grab it and move it around to teach it by demonstration. And it means that the setup of the robot becomes incredibly easy when you do that. And we were on a path to take over the world, all eight of us. And then we got bought by Google. It was a fantastic experience because we get the call one day and we say, no, we’re not going to sell. And then we start thinking about the fact that they’re so big that if they were doing what we do just by mistake, they could put us out of business. So maybe we should take this seriously. Maybe there’s more than one way to take over the world.
(00:07:09)
Something I always think about as an early stage investor is, you know, to what degree will the incumbents just make that a feature inside there and bundle it?
(00:07:16)
Yeah.
(00:07:16)
Right.
(00:07:16)
It’s really important.
(00:07:17)
Yeah.
(00:07:18)
And especially now, you know, I think there’s in certain markets, you can end up with an inadvertent race to the bottom. Because everyone looks at the feature set that they want to build and they decide some amount of things that they’re going to give away for free. And if there are enough competitors doing that in a market and the selection is somewhat random, there’s a free way to do almost everything. So the value that any one player preserves gets eroded by the free offerings of everyone.
(00:07:47)
This is why I’ve actually been doing a lot more robotics lately as an investor. I mean, that’s something I know a ton about, but I think it’s that the timing is right in the industry. Maybe you could talk a little bit about that, but I think there’s all these vertical use cases in robotics and some horizontal platforms as well. And I feel like you’re starting to get to the point where you can, you know, deploying some value out of these robots, but maybe we’re getting ahead of ourselves. We’ll talk about robust in a second, but I’d love to kind of wrap up on, you know, lessons or any standout stories working at Google.
(00:08:17)
Yeah. So there’s a few things there. I’m still a big fan of Google. I think it gets a bad rap compared to what it actually is on the inside.
(00:08:25)
Does it get a bad rap? Google gets a bad rap?
(00:08:28)
I think more so in Europe than here.
(00:08:30)
Oh, okay.
(00:08:32)
There’s more anti-Google sentiment, I think, versus the positive things that people on the inside are trying to make true. I’ll go on record for that one.
(00:08:41)
I always like their motto, don’t be evil, right? They’re always trying to do the right thing.
(00:08:45)
Yeah, they have a lot of data and they have a lot of power, but they were never trying to be evil about it, right? Or even monopolistic.
(00:08:54)
Certainly my experience. So I think really my big learning from Google was what I saw happening with modern machine learning and deep learning inside of Google was a few years ahead of what everyone else in the world knew was possible. And that really convinced me that we could build cost-effective, incredibly capable robots soon. I had a difference of opinion in how you do that versus how Google was trying to do it. Google, of course, really leans toward doing machine learning and deep learning and really AI completely end to end. I personally believe that you want to keep your robot systems modular, even if you’re using AI for parts of that architecture, just because we need to be able to answer deep questions about why these things did what they did. And right now, I don’t see a way to give that transparency and that ability to introspect without having a modular system that you can observe traces on. And I think until we get a system that can both perform a job and then very credibly explain why it did what it did, I don’t think monolithic systems make sense for physical AI systems in general.
(00:10:12)
So all these robotic systems that we’re now starting to build, step one is getting them to be functionally appropriate, accurate, and valuable. And step two is making sure that they’re transparent and they can explain what they did because nothing in the physical world is perfect all the time. And it doesn’t matter if we’re talking about, you know, a car, a robot dog, or a light switch. Stuff breaks, stuff works properly for one minute and then is on the fritz two seconds later and then starts working again. This is, you know, all of these things are fighting against physics, you know, it’s physics and entropy. Yeah. So given that, we need transparency in all of these systems and I think the only way I can see to get there is to have something that’s modular that you can actually trace what happens.
(00:11:02)
Good segue to talk about robust, I imagine. So what was the aha moment or the impetus behind going out and doing a startup again? That seems kind of crazy. Leave the cocoon of Google, you know, with all the resources.
(00:11:16)
It was interesting. And I can’t name names, but I had someone very high level person at Google say to me that they were jealous when I decided to leave. Because it was this moment when deep learning was showing itself to be this magical thing. And it was unreasonable, the results that we were getting. You know, now we all see, you know, five years later, all of these different AI products that were like, wow, this is pretty impressive. But at that moment, when you started seeing this signal, for me, a light bulb went off. And I said, you know, I’ve always been able to build anything I wanted. You know, as I mentioned, I’ve been programming since I was 11. And I’ve always had the perspective that if I can understand any problem to the point where I can decompose it into steps, I can make a computer do it. However, there was now this new tool, which was if I get the right set of examples of the thing that I want to do, we can now teach a computer to do it. And that’s wild.
(00:12:17)
So wait, I don’t actually need to know how to program nearly as well. I just collect bunches of pictures and bunches of examples or lots of different graphs and I can make the computer do millions of things that look like this. And that was just a fundamentally different paradigm. So I decided I needed to add that to my quiver in terms of the things that I can use to build systems, companies, et cetera. So I left Google and I spent about a year and a half doing, I joke now, every online class, a bunch of different online contests, consulting to companies because I find that when I use something that I’m learning, I just get better at it faster. You know, I’ve always had the ability to learn quickly. So during that time, I did like a week of every class in about a day and a half. So I could just do a ton of classes. That left me with a really strong belief that we could make incredibly capable systems. The capabilities we had for perception and for planning were amazing. But these are also numerical systems. There is, there’s actually not magic in there. No, we’re not tapping into the knowledge of some hidden universe. It really is a big matrix and you’re multiplying numbers. That’s what’s happening in there.
(00:13:32)
Turns out that might be the brain too, right?
(00:13:34)
Yes. I mean, and there’s, you know, so we like, you know, the fact that we’re getting such effective behavior from a relatively small set of mathematical operations is, it’s telling.
(00:13:49)
That’s what the book, The Master Algorithm, what we were looking for. I read it back in 2019, I think. I’d love to have him on the author on the top pod. Anybody make an intro for me? Back then, we didn’t know what it was, right? This is six years ago. But it seems to be large language models, which are basically softmax neurons stacked and stacked and stacked and all kinds of convolutions.
(00:14:13)
Yeah, I believe that large language models are just one example. So the transformer architecture in general is what we’re seeing to kind of be the next wave after the first wave of deep learning. But that architecture, whether it’s using large language models, visual language models, like there’s a bunch of different heads you can put on them so that you can look at language, you can look at pictures, you can even use them on tabular data. And so for me, transformers are the new magic. There’s something about that architecture that’s a combination of everything that’s good about deep learning, but then also using attention and attention at multiple levels that is unreasonably effective, absolutely unreasonably effective at everything we’ve pointed it at so far.
(00:15:00)
So you’re consulting, you’re taking a bunch of classes, 18 months goes by and then what happens?
(00:15:05)
And then me being me, I make a long list of business ideas and it ends up one or two lines per idea and it’s nine pages by the time I’m done. And one of those ideas, it turns out that Rodney Brooks, my co-founder, was also thinking about. He’s a legend in robotics, founder of iRobot, co-inventor of the Roomba. Ran the AI lab at MIT for almost 20 years and we end up starting at coffee and ending when someone needs to get on a flight and then talking when we land and deciding that I’m going to be a co-founder with him and some others starting a company by 6pm that night because there was just so much overlap across these things that we were thinking about and then forward that to now. Lots of twists and turns because of COVID. You know, you start a company in the second half of 2019 and eight months in, everyone’s in their living room and on Zoom all day. So, and you don’t have access to construction sites and customers and all these things that were initially part of what we were focused on.
(00:16:07)
But it allowed us to reset and look at what industries do we think would get the most value from the types of things we could build with these new technologies and logistics and supply chain and manufacturing kept rising to the top and then product management 101 you do a bunch of customer discovery and you talk to tons and tons of people and we did everything from normal customer discovery to a diary study of 120 people in warehouses, floor associates, supervisors, managers, directors. And we kept seeing recurring themes that we couldn’t solve with existing hardware and existing robots. We just kept coming at it and feeling like we weren’t solving enough of the problem for customers or it felt too clunky. And that really informed the design of where we ended up, which is a completely new class of robot that unlike any other mobile robot, you can literally just grab onto the robot and move it. And a person always has agency over a robot in those contexts.
(00:17:13)
So the insight here was let’s go after manufacturing and supply chain. We see the biggest need there, but the insight like, okay, there’s just transformer architecture coming. It’s enabling stuff. You were already seeing at Google and with some of your clients, right? It makes you wonder how like advanced the stuff at Google is right now. When, you know, I remember the guy three years ago rang the alarm bell, like, hey, there’s sentient AI inside of Google. Maybe he was right. You know, if they’re three years advanced from what’s currently on the market, a lot of people think it’s sentient now, right? You know, if you chat with it. So, and then you were, the big insight here was like, okay, we’re going to do these vertical applications, but we’re going to approach it in a different way because manufacturers, supply chain folks have already been doing robotics, right? So the budget is already there, the familiarity. But the unlock here is like the trainability where you don’t have to rigorously like code up these robots to do very repetitive tasks. They’re a lot more malleable, right?
(00:18:04)
So I would go even further than that, that the tasks that are now available are now more flexible than the tasks that have been automated in the past. Bluntly, it’s pretty darned easy to take a big industrial robot, lay out the paths it needs to follow, hook it up to a few PLCs that tell it when to go and when to stop, and let it rip. The applications that now exist where the value remains, you have to interact with people, the environment is dynamic. There’s times when you start, there’s times when you stop. It would be ideal to follow the rules of the road and the traffic signals, if you will, inside of a facility. And all of those things require more and more smarts. And as humans, we take it for granted. You know, of course, you walk, you look around the corner, you see if something’s coming or not. You learn in your safety briefing that at this one intersection, you’re extra careful because the forklifts go past there. Any warehouse, the forklift is the apex predator. It gets priority over everything. So that’s how you need to think. Your head needs to be on a swivel. And these things become straightforward and automatic for people. But getting robots to adhere to those rules, to essentially start following our social decisions, is non-trivial. I think it’s exciting, but it is non-trivial because there’s a long tail of behaviors and expectations of what the robots need to understand and do.
(00:19:32)
So what that allowed us to do is build a robot that was more capable in a few dimensions. One is perceiving the world. So instead of using LIDAR, we use cameras. And we actually use eight cameras around the robot so that it can see, use stereo vision to understand how far away things are. But also the eight cameras allow the robot to see 360 degrees around it. And then use a bunch of AI and neural processing to understand what the cameras see. That’s a person, that’s a box, that’s the dreaded forklift. And based on that, the way that the robots act is quite different than a robot that just uses two-dimensional LIDAR, which is what indoor robots use, that just tells them, hey, there’s free space in front of you and there’s an obstacle here. No idea what the obstacle is. So it can’t actually do any different behavior. So for example, if you’re supposed to interact with a person and present yourself so that the person can put something on the robot, most robots fake that by driving to the location where the person is supposed to be standing. We actually know that that’s a person. And in fact, we know it’s the person who badged in, even without collecting PII about them. So we interact differently and interact more appropriately. And that nets out to really how we make our customers money and how we provide value is reducing the amount of walking, making the activities that people do take less time. And those two things end up making people much more productive in any environment that we’re deployed in. And that means that the cost of doing any of those activities goes down for that customer.
(00:21:13)
Lots to unpack there. Really interesting. One thing you said is the eight cameras versus LiDAR indoors. So the whole debate between Elon, you know, the cameras on the Tesla and Waymo with their whole LiDAR system. Where do you fall in on that? Do you believe Elon’s on the right path with just cameras on the car? Or do you feel like you need the LiDAR? And I’m just talking about the car context first, and then we’ll talk about the robotic context.
(00:21:34)
Yeah, this is fantastic, of course, because many years, but I’m a Tesla owner. Ex-Google. I participate in many different parts of it. Bluntly, I think multiple redundant sensors on a car is critical because the speed at which cars move, I don’t think cameras should be the only solution. I think you need multiple redundant sensors because the velocity is— At least a forward-facing LiDAR, right? Just one little spinny, cheap LiDAR in the front.
(00:22:01)
Yeah, LiDAR, radar, like some set of sensors because there are other things. The other thing that happens with cars is you have an incredibly high dynamic range on lighting while, you know, so you... White-out conditions. White-out conditions. Tunnel to sun. Now, while you’re adjusting the parameters on that camera, there are features you’re not going to see. And you’re also in an environment where the cost of being wrong is extremely high. And the marginal cost for the device that you’re on for a couple of redundant sensors is actually quite small. Real cost is actually development time and some complexity in your software stack. So strong, strong proponent of some amount of redundant sensing in cars. But when we now translate it, unfortunately, to what I do, our robots move maximum 1.5 meters per second. And at that rate, I’m getting many, many camera images per second that I can filter through and do Kalman filters, do all of these fancy techniques that we now understand how to do to feel very confident in what we’re observing or knowing when we’re not observing and being able to adapt in our known stopping distance before there’s the chance for any type of obstacle.
(00:23:17)
So tell me about your tech stack and how it differs from some of your competitors and what’s unique about it.
(00:23:25)
Sure. So as I mentioned, we use only cameras and no LiDAR. So that means that we perceive the world semantically. We understand it in human terms. Like that’s a person, that’s another robot, that’s a forklift. The other big difference in our tech stack is we’ve built technology internally that allows us to describe workflows very, very compactly. And essentially workflows for us are content and the stack isn’t 100% tied to workflows. So if our robots are doing a point-to-point delivery, they’re just running an application or a workflow on top of the major part of the stack. If our robot, same robot, is doing a picking application where they’re driving to specific locations and guiding a person about where the next thing that they need to grab for an order is, again, they’re just running an application on top of the rest of our tech stack. And by application, it’s not even giving the right impression. If there is special code, it’s a very small amount on top of essentially configuration. Our ability to describe what the robot should do in a given context while working with people, think of like, you know, drawing stuff in Photoshop and the rest of the stack is Photoshop running underneath it.
(00:24:45)
Are the actual robots built by you guys as well, or is it off the shelf modular components? Tell us about the vertical integration if there is anything.
(00:24:53)
So what we’re really focused on is what we call collaborative productivity. And this is the idea that a person and a robot interacting appropriately, driven by the software that’s underneath, actually get a job done much more quickly and much more efficiently than a person alone or a robot alone. And that combination of using people for what they’re good at and using the robot for what it’s good at actually gets you a better ROI than trying to do everything with labor or, in most cases, trying to do everything with pure autonomy. I always draw the analogy between tolerances in mechanical engineering and autonomy in robots, that getting to 99% might cost you X. Getting to 99.9%—it doesn’t cost you X plus five, it costs you X. And bluntly, if you can get that 1% done by people, you get better ROI, but you also get human oversight, which I think is something that’s really overlooked. So our focus is really building these systems where people do part of it, robots do part of it. You get an incredibly efficient solution and you get the benefit of human oversight.
(00:26:05)
So it’s really about how do you raise the bar on what it means for robots? And for anybody listening and driving in the car, there was a video on the slide that the robot is basically a moving shelf. In this case, the worker, the human, is just grabbing boxes off the shelf that comes to them and then the worker just pushes the shelf away when he’s done and the robot moves on. And the next robot shelf comes along and repeats. And to your point, that’s 99% of the value right there. The human doesn’t have to move that much. And that last 1% of getting a robotic arm to grab the right package and put it in the right place on the shelf, 10x harder.
(00:26:43)
It is. And I always say, Amazon hasn’t done it yet. And they’ve spent a number that has nine zeros on getting there. Billions and billions of dollars.
(00:26:52)
But they have done the moving shelf thing, right? They have their own proprietary moving shelf.
(00:26:56)
Things that are similar.
(00:26:58)
No, I know because they bought them one robotics company.
(00:27:00)
Yeah, so they bought Kiva years and years ago. And those robots actually drive under a special kind of shelf and move it around. And they recently said they have about a million of those at this point.
(00:27:13)
Yeah, they are my example of what it actually looks like when a company really understands the advantage of autonomous mobile robots and how much efficiency you can get at scale by deploying these things properly.
(00:27:26)
Right, that 99% of the not human value add moving shelves around kind of stuff.
(00:27:32)
Yeah, moving packages around. That’s really interesting.
(00:27:34)
Yeah, so as you mentioned, this is a novel form of robot where the form factor is very much like a shelf and it allows people to be really familiar with the form factor and understand really easily how to use it and how it replaces the cart with wheels that’s in their facility. And just like that cart, at any time they can grab the handlebar and move it around. And therefore, it’s this thing that’s fully autonomous and it can go anywhere it needs to go and remove all of that walking, but it’s also completely collaborative in a way that most robots aren’t.
(00:28:07)
And so you actually literally sell the shelf with the robot system on it and then it can autonomously move around a manufacturing facility, any logistics facility. You have some stats here, but what’s the implementation look like? So like, okay, I have a need for this. How fast can I get going on it? What’s implementation look like?
(00:28:27)
The implementation can be incredibly quick on the tech side. So I have to find some wood and knock on it because we have an unbroken streak of shipping robots. First day that we show up at a facility, we unpack them, get them moving, map the facility and have robots driving on the first day. Because we use cameras, we use this technique called vSlam, visual simultaneous localization and mapping. And because we have this handlebar that magically makes the cart effortless, the first robot that we put together and power up, we put it in mapping mode, grab the handlebar, drive it, you know, just push it up and down all the aisles. That makes the map. It shares it with the cloud instance for that facility. All the other robots download that map. And now they can all move around in that map.
(00:29:13)
That’s amazing.
(00:29:14)
And it’s dynamic, I would imagine. So as the robot keeps moving around the facility, if they shift a line or there’s a crate sitting there in the path, it just kind of detects that, adds it to the map.
(00:29:24)
Exactly. And it adds it to the map. And then we always do a human decision on whether or not the map gets updated because we will have context that they may not have. So for example, we may know that bunch of pallets is laid out because there’s a special going on this week. We don’t want that added to the permanent map and the robots will see those objects and just deal with it even though they’re not going to stay in the permanent map. So we don’t need to update the map. So we leave because our job is making the customer more efficient, not necessarily being able to say, oh, our robots learn faster than any other robot. So we always put a human in the loop on that decision.
(00:30:04)
And I think this is an important point in general, that for real deployments, there are careful steps that you want to take to ensure that enterprise customers get the experience that they expect, know that you’re always hitting their SLAs, that there are no surprises. Very often, it is more important to not surprise the customer than to have your stats go up by 1% that day.
(00:30:29)
I love that. Anything else in the slides you want to cover?
(00:30:31)
Yeah. So I think the last thing I wanted to touch on here. So while the form factor is pretty simple and seems really familiar, there’s a bunch of affordances and capabilities on it that make it a general purpose device. So as I mentioned, there’s eight cameras in the top and this giant NVIDIA compute unit that allows us to use AI in the robot. There’s the magic handlebar.
(00:30:57)
Which compute unit are you using from NVIDIA? What’s it require?
(00:31:00)
Yeah, we’re using the Orin, the Jetson Orin series, which is what was in cars up until Thor came out last year or earlier this year.
(00:31:09)
But pretty powerful.
(00:31:09)
It is. And it’s mind-bogglingly powerful, actually.
(00:31:14)
How many, do you know how many like TFLOPs or anything like that?
(00:31:18)
I believe the ORIN that we’re using, you know, there’s always a, if you do FP16, so 16-bit floating point, I believe it’s 40 TeraOps. I mean, I did a calculation that that was the amount of compute available to the whole planet in about 2003. And that is literally what’s running our robotic shelving inside of manufacturing facilities now. It’s all the computing capacity 20, 25 years ago of the entire planet.
(00:31:45)
Most cars that have ADAS, by the way.
(00:31:49)
Yeah, I think my RTX 4090, which is last gen for my gaming PC, is like 80 teraflops.
(00:31:55)
It’s mind boggling.
(00:31:56)
Yeah.
(00:31:57)
Yeah, it’s mind boggling what we have access to.
(00:32:00)
Yeah, and that’s part of this trend, which is bluntly, we have no excuse to not have robots providing value in a consistent, explainable way for people at this point.
(00:32:12)
I think right now is the time for all these verticalized general purpose applications where it’s like, hey, you have this one problem. How do we take robotics and vision? Because it’s sufficiently advanced and solve that problem for you. Like, yeah, you got Figure and Tesla. They’re trying to do the complete humanoid robot that can do everything. How far away do you think those are?
(00:32:31)
I think it’s tough because the amount of complexity that you add when you make a humanoid with legs is enormous. You’ve added now, in the easy case, 20-something degrees of freedom. In the hard case, close to 40. To make that work is a giant technical challenge. The other piece, unfortunately, while the humanoid form is great because so much of the world that people use is designed for the human form scale, just because you’re in the human form doesn’t mean you can do everything a human can do. And just because you can put a humanoid robot for that use case doesn’t mean you should either.
(00:33:15)
Sometimes you just need a robotic shelf that moves around.
(00:33:18)
Yeah.
(00:33:19)
Or, you know, or you just need this one robotic arm that like grabs something off of a conveyor belt, puts it in a box. You don’t need a humanoid to do that.
(00:33:25)
Completely. You don’t need a hand with 30 degrees of freedom. It could be just a claw.
(00:33:29)
Yep.
(00:33:29)
So one of the things that is important for me to put a fine point on is robots like ours and humanoids don’t compete the same way that a person doesn’t compete with a truck or a shopping cart or a lorry. We figured out a long time ago that piling a bunch of stuff onto a platform that has three, four, six wheels is a really efficient way to move stuff around. I know it’s part of the reason, you know, essentially from the Roman Empire onward. We’ve put stuff on carts and moved it around because it’s so much more efficient to move a hundred things on a platform that has wheels on it than unfortunately wrapping it in your arms and carrying it around.
(00:34:12)
And that’s really what this class of robots is about. It’s about portage. It’s about moving stuff around. It’s about logistics. So humans aren’t carrying boxes around inside of warehouses. So humans are pushing carts or pallet jacks or other things that move boxes around. And then we put them in trucks. We put them on pallets. So we consolidate. And that’s the way you efficiently move stuff around. That’s what the logistics industry is, which is, by the way, 20% of GDP. 20% of everything that people do across the globe is moving stuff around. And they’re not carrying boxes in their arms. And it’s invisible to people who aren’t in that industry. But I think during COVID, we all figured out what happens when that slows down. Nothing works.
(00:34:56)
That’s fascinating. And so it seems to be working. You got some traction. You got partnerships with some big names. Some serious adoption here.
(00:35:04)
Yeah, no, we’re really excited about where things are going. You know, we of course have a great relationship and our landmark customer is DHL Supply Chain. They’re the part of DHL that runs warehouses for companies. So their customers are, you know, e-commerce or retail customers where they hold the DHL Supply Chain, holds the inventory and then ships it to the retail stores or ships it to customers, et cetera. And they’re gigantic. They’re the number one in the world that does this. This is called third-party logistics.
(00:35:36)
They’re the number one.
(00:35:37)
3PL, as they’re called.
(00:35:38)
3PL, yeah.
(00:35:39)
And yeah, so they have about 2,500 facilities globally. It’s incredible, the scale there. It’s awe-inspiring when you’re in a warehouse that’s 1.5 million square feet, big enough to have a three or four story building inside it. And that’s what these facilities look like at scale.
(00:35:58)
Yeah.
(00:35:58)
Anything else on those slides you want to cover?
(00:36:02)
The interesting thing on this last slide that I want to mention is the form factor is that of a shelf and it’s pretty standardized. However, because of the behavior of the robot, the handlebar, the screen, the marquee to tell people what’s going on, the ability for the shelves to light themselves up in different parts to tell people where something needs to get done. It’s one piece of hardware that becomes more general. So this one piece of hardware can now do many different applications.
(00:36:34)
And we’re excited because we feel like this capability, the software and AI capability and the ability to have different workflows really represents that switch from when telephony went from kind of flip phones to smartphones. You know, a smartphone is not just a phone, it has many different applications on it. And we feel like that’s the point that robotics is at now, that even these vertical applications that look like one piece of hardware can be generalized much more than at any point in the past, if they’re designed appropriately.
(00:37:10)
For us, those affordances were touch so that a person can change what the robot’s doing, lots of visual cues so that the robot can always indicate the next action to do or the next piece of input that it needs. And just by combining those, we have one platform that can do picking and e-commerce fulfillment for one customer and point-to-point delivery in a manufacturing facility or parts running to feed a manufacturing cell for another customer.
(00:37:41)
So looking out three to five years, what are you excited about for Robust? And then the second part of that question would be for the industry.
(00:37:48)
Few things. So for Robust, of course, I’m excited about scaling the business. This is one of the things that I enjoy most about businesses — when you start scaling and you figure out the things that are easy to make automatic and you’re problem solving every day, trying to figure out how to make the next hurdle automatic. So that’s what I’m excited about us. And the parts of that are getting lots and lots of Carters, our robot, out into the world, getting new versions of Carter — you know there may be smaller and larger ones in the future — and all the great customers and collaborators we’ll work with in doing that. For the industry in general, I’m not excited about one specific thing. I’m excited much more about the general arc that we’re on now which is meteoric. Like, I don’t think anyone can predict the AI capabilities that we’ll have five years from now and what that means for physical AI systems like we build and like other people build. I do know that it’s going to be — I expect it to be — at least as much change as we’ve seen in the last five years. And if that’s true it will mean that some parts of what is hard now, some parts of our difficulty now completely disappear.
(00:39:11)
Well, I think we’re also in a hardcore exponential progression right now. We’re not just in Moore’s Law anymore with AI. We’re in orders of magnitude more capabilities every year, if not like a half-oom or a full-oom. Did you read the situational awareness paper, Leopold, where he talks about orders of magnitude?
(00:39:29)
No, I did not.
(00:39:30)
Really, really good paper. It’s a little old now, but he started a hedge fund and was making billions of dollars from all his ideas, as one does. But basically, he lines up the orders of magnitude that are happening every year, both in the data and the compute and the algorithm efficiencies. You add those all up, we’re not just getting 2x better every year, we’re getting 10x better, which is why AI is now crushing all these benchmarks. Every benchmark is basically saturated.
(00:39:58)
Yeah, this is great. And I think that that is true in general. And I think the only bottlenecks are the data that we have access to. So I do think one of the bottlenecks for robotics is getting appropriate data so that we are in that regime. One of the things that made large language models explode the way we did is because we had the entire internet to mine as the data that we could train these things on. We don’t have a corpus like that for robotics yet.
(00:40:31)
Are you guys starting to kind of toy with simulated data, right? Where you create world models and let these carts go around in simulated game environments?
(00:40:40)
Yeah, so I think I won’t say specifically the techniques that we use. I’ll just say that one of my big learnings through my post-Google exploration was that simulation and synthetic data are a crucial point to any scheme that you have for learning. I think essentially, no matter how much real-world data that you have, generate, depends on the problem, but somewhere between five and X, that same data in simulation or synthetically, and your networks will perform better.
(00:41:12)
So you just get — you increase the sample efficiency on your real world data by adding synthetic.
(00:41:18)
Are you seeing gains on the kind of the test time compute layer for kind of these visual robotic systems the way we are in the large language model kind of text-based environments?
(00:41:27)
Again, I think the gains aren’t the same because the data, the access to data is different. But for the most part, all of these techniques have impact regardless of the domain that we learn them in.
(00:41:40)
Fascinating. Well, why don’t we wrap up with some rapid fire. What’s a small tweak you’ve added recently that made a surprisingly big impact at Robust?
(00:41:48)
So internally, we’re at the beginning of our scaling phase. So internally, I’ve started describing and pitching the idea of phase shift and what that looks like as you scale.
(00:42:01)
I use the... What is phase shift? What is that?
(00:42:04)
That means, you know, we’re moving from one phase is, you know, a small number of robots, small number of customers we can hand adjust everything.
(00:42:14)
Oh, you’re moving to the early majority. You’re crossing the chasm to like this new phase of scale.
(00:42:20)
Exactly. And in that new phase of scale, you have to use tools that look at things in aggregate. And the analogy, the metaphor that I use is, imagine if no one had ever invented a bicycle, but you knew how to ride one. And you can understand that it’s a movement from static stability to dynamic stability. You understand that once you’re moving, this thing feels pretty stable and you can do stuff with it, like weave between things that you can’t do any other way. Where we are when you do a phase shift in a company is, you have many people who’ve only experienced static stability. You’ve only experienced standing on your own two feet, walking, etc. And now we’re moving to this phase where it’s like riding a bicycle. It works, but you don’t know how. And you’re going to scuff some knees, you’re going to fall down occasionally, but you will get to this very different phase where the system works differently.
(00:43:22)
I’m kind of laughing at myself because we always wrestle with this with every generation of machine learning and AI, how to explain it. I was thinking about this 15 years ago when I first got into AI. It’s like, okay, how does the model work and how do we explain it to our customers? Every new generation of models and AI, we struggle to explain how it works, why it works. It’s really fascinating. And back then we were just doing logistical regression at scale. We’re still struggling how to explain it to our ad tech customers.
(00:43:52)
Anyway, if you could redesign an industrial workflow from scratch with no legacy constraints, how would Robust or one of your products fit in?
(00:43:59)
Yeah, so what I would love to do is create a warehouse that is fully reconfigurable and all it has is static shelves, static racks to hold products, and Carter robots. And dynamically, they switch between picking, replenishment, sortation, because you can actually use the robot as what’s called a sort wall, doing point-to-point deliveries. And you have a giant optimizer that decides what robot’s doing what when. And then you have this facility that can have a huge variation of different order profiles or order SKUs or volume changes. And in software, you adjust resources to move between the set points as your order profile and your volume changes without having to change any hardware. That’s one that I’d love to make happen in the next year or two.
(00:44:52)
I would imagine that you’re landing and expanding with a lot of your clients too. Do you find you throw a couple of these into a warehouse and three months later, like, hey, we need two more of those. And three months later, send us four more of those.
(00:45:03)
Yeah, yes. And it’s not even three months. We typically will have people come up with new ideas or start asking those questions a week, two, three, four.
(00:45:14)
Nice. That’s great.
(00:45:15)
In the spirit of robots should promise and fulfill, I think this is something you said, the most surprising thing a human partner has taught a robot to do.
(00:45:22)
So this one isn’t really teaching, but this is maybe my most surprising thing is we have a site where one of the floor associates actually refers to the robots as her babies. And she’s pretty protective of them. And she’s a big influence on training all the other floor associates and strongly affecting their perception of the fleet. So there’s a lot of positive affect around these robots making us more productive. They’re great to work with. And the degree of emotion and connection was surprising to me. You know, of course, we designed for it, but even then to see it.
(00:45:58)
And go, wow, we hit that mill. Essentially, I think what happened is maybe it happened in reverse, that it’s about the people learning. And that’s where the surprise came.
(00:46:10)
So you’ve seen startups, moonshots and scale. What’s a pattern or myth in robotics today that you wish more people questioned?
(00:46:17)
Great question. I think the myth, and this is true for robots and technology in general, but I’ll say it for robots. The myth is that robots by themselves solve problems and can scale without a huge amount of change management. Humans are the fulcrum to all change. Introducing any new technology and making it successful has typically more to do with human behavior than the underlying capability of the technology or the robot. So that’s maybe the myth that I would like to bust the most where people think, here’s this great new technology, of course this change will happen. And we have thousands of case studies to show that it’s really about change management and having people change behavior to make any system really be adopted.
(00:47:07)
How are you guys thinking, if at all, about ethics and physical AI?
(00:47:21)
Pretty much yeah we do this thing that we love doing which is letting robots just drive around and seeing how long it takes for customers to essentially start ignoring them and one of the moments that we love is when someone on the floor tells another person yeah it’ll get really close to you but it’s not going to hit you it’ll get like this far and it just try and squeeze through because it’s trying to do its job also so to answer your question I don’t think there’s any you know grand theory of robot ethics you know there’s some pillars that I look at and I think are really important and you know it’s safety it’s real and perceived safety making something that’s actually safe and that people perceive as safe and feel comfortable around I think making something where one of those things isn’t true is problematic I would say transparency transparency of purpose meaning people know what the robot’s trying to do transparency of intention which again means that what the robot is promising and evoking is genuine and transparency of implication so for example we talk about a lot about our robots empowering people but we’re also very open about the fact that in general the labor contribution in this facility for this task will go down you will have fewer people on your shift no one and the good news is when you’re up front about it no one’s surprised everyone’s like yeah we have to get more efficient and we’re going to do this in a way where it’s actually pretty delightful and empowering to work on this every day so people are willing to accept that trade-off and I think all of those come back to really one thing which is about trust and that’s about whether you’re a robot or a company or whatever consistently doing what you say you will over time and living up to whatever promise you’re putting on the table and in my mind those are really the pillars of ethics in terms of how they affect what we do what I like building etc but I also will recognize that that’s only part of the questions on the table.
(00:49:19)
And you know you can’t hold yourself responsible for job loss and manufacturing facilities as an entrepreneur.
(00:49:25)
Well yes and I think no one is going to argue that we shouldn’t be more efficient or we shouldn’t be more productive but there are myriad ways of getting there our position is just design with people in mind from the beginning so that you end up with a solution that is highly efficient and highly productive and humane and empowering for the people who are doing that work.
(00:49:52)
You know what I’m starting to see, having spent an hour with you now, is this future robotics where it’s not just humans and robots, but it could be robots and robots too. Where eventually there’s different kinds of robots interacting with other kinds of verticalized robots for specific use cases. Just like humans are pretty adaptable, we can kind of do a little bit of everything. But eventually you might have robots that carry pallets around and robots that carry shelving around and robots that pick stuff off conveyor belts.
(00:50:19)
Yes, I completely agree with that point of view. And my perspective is you’re going to have a rainforest, if you will, of different morphologies and different capabilities. So the vanilla varicose version of Carter, it’s a shelf and it needs a person or a fixed robot arm to put stuff on it or take stuff off. We of course have designs for peripherals we can add to do that, but in a lot of cases you don’t want that. So for us, whether it’s a fixed arm that puts stuff on and off of Carter, a conveyor that slides stuff on and off of Carter, a humanoid that puts stuff on and off of Carter, we don’t care. You know, at some point, there’s the need for a thing that you can put 20 boxes on and can drive 500 meters in a facility to get those things there on time the most efficient way possible. And that’s the core of what we provide. It’s portage of multiple things at a time.
(00:51:18)
Given the pace of change, how are you guys staying up on staying resilient as not just leaders, but also technologists? You know, everything’s happening so fast.
(00:51:27)
Yeah, so I’m going to take this maybe in a different direction. But I think staying resilient for me ends up being about personal power. And that has to come from the ways that you fill yourself up. One of those for me is meditation. I just think it’s, you know, we all need some set of things that help us stay grounded, especially in a world where we have devices and systems that systematically destroy attention. So for me, it’s about meditation. It’s about maintaining human connection with friends and family. It’s about pursuing and recognizing the quiet moments that help you kind of consolidate again, because it’s only in those quiet moments that we really come up with brilliant ideas, that we really see the terrible crystals themselves, you know, the heart of the thing. So it’s about building practices that give you those moments.
(00:52:21)
I love that. Plus one for meditation. We’ve got a transcendental meditation teacher on episode, I don’t know, probably 15 or 20 of this podcast. So yeah, everybody should try it. It’s, you know, whether or not you’re religious, you can just go do whatever. Go do some meditation. Try it out. And it’s not about religion. It’s about taking back your attention, especially in this world.
(00:52:42)
Yeah, and releasing stress. Because you’ll be sitting there and you’re like, oh yeah, that person said that. And you’re like, oh, I’m really stressed out about that.
(00:52:49)
Anthony, thank you so much for coming on. This was a really fun conversation. I learned a lot.
(00:52:53)
Thanks. I appreciate it. And thanks for having me on the show, Brian.
