How Democratizing Access to Robotics will Benefit All: Brian Gerkey on Open-Source Software

Transcript

Brian Gerkey (00:00)

You can build a system that looks good in a spreadsheet, looks good on paper. If it's not usable by the people who are going to be on the shop floor interacting with it, if they don't understand the value that it brings to them, then they're going to reject it. And if you can't make the case to the management of a facility that their workers are actually going to get value out of this, then you're never going to make the sale.

 

Brian Heater (00:21)

Hey, everyone. Welcome to another episode of Automated. My name is Brian Heater. I am the managing editor at the Association for Advancing Automation. I am coincidentally also your host. This week, we are digging into Alphabet's physical AI play, open source hardware, robotics and a whole lot more with Brian Gurkey of Intrinsic and Open Robotics. It's a really great conversation. I learned a lot and I hope you will as well. So. Like, subscribe, and enjoy.

 

Brian Heater:

I like to ask this to people who have been around the industry for a long time, because I get this feeling myself. You know, I've been covering robotics for like seven, eight years now, where there's really this sort of like,  it's happening moment that we've been experiencing over the past two years. Does it feel like all of the things that you've been kind of talking about and working to are sort of starting to, in a certain sense, come to fruition?

 

Brian Gerkey (01:26)

I think I'm seeing the same thing that you are. And I've been waiting for this for 25 years since I've been working in robotics, waiting for it to happen. there is a certain, sometimes you feel a little bit like Charlie Brown going after the football. Like, this really the time? In retrospect, I can say that I think we had some big moments along the way. Like, know, circa 15 years ago, we figured out autonomous mobile robots like AMRs and warehouses, factories that that's working pretty well right now.  I think the next big step is how do we combine mobility and manipulation and figure out how to handle the much greater variety of things that people want robots to do. And is that about to happen? think we're getting, there's good signs. Yeah, I'd say that a lot of things have come together from the hardware is getting better, more reliable, costs. There's greater willingness on the part of especially some pretty old industries to like adopt new technology and kind of change the way that they think about solving some of the problems they've had for a long time. And then as a robot software guy myself, I'll say the software has gotten much better. There are just, we've got more greater ability to soak up a lot of the variability and the challenges that you have in these tasks in the software in a way that you just, couldn't do before.

 

Brian Heater (02:55)

Yeah, I guess something important to point out here is that this sort of thing, it doesn't just happen at once. isn't like, you know, this is like the moment for a box. Obviously, it's going to to come in waves. But this this does to me feel pretty different than all those those those moments in the past. And I'm curious at that point, at the point when you were really sort of starting your career early on, how would you describe the state of AI and robotics.

 

Brian Gerkey (03:26)

Man, I so I got into robotics in the mid 90s as an undergrad. There was a guy named Jim Jennings was a new professor to the CS department. I was at Tulane University in New Orleans and he showed up and started a robot lab and I you know as a eager undergrad urgent, know, budding computer scientist, I saw these robots and I volunteered to work in his lab and I just got such satisfaction out of writing code that made something move physically in the world. That was just fundamentally different from all the other things. Like all the other things you do with software, generally you get the right letter, number, or color on a screen. And with robots, you get to move things physically in the world. for me, it was just fascinating.

 

At the same time, the robots were capable of just about nothing. They were lab tools. They were experimental tools that we had to help us try to answer scientific questions about multi-robot collaboration or motion planning and things like that. And we wrote these papers saying, you know, one day there will be all these robots in these factories, in these warehouses doing all this useful work. But I don't think we really expected it to come, you know, anytime soon. It was really an academic endeavor. Like that was famously a time of, it was one of the AI winters that we talk about where there had been a huge level of interest in combination of expert systems, early versions of neural networks, and they had all kind of failed to meet expectations. And so the nineties was a time of kind of reconsidering like what kind of intelligence can we really expect from artificial systems, had a very different feel from how it does now. At the time, it was, I really thought about it as an intellectual academic exercise. Like we did it because we thought it was interesting, not because we were convinced it was going to be useful.

 

Brian Heater (05:19)

Actually, walk me through this concept of AI winter and whether there are any parallels or I guess potential parallels to where we are right now. it to a certain extent, is it kind of over promising and under delivering? Obviously, there's always going to be an element of that in tech and obviously like, listen, you no one's going to say that to a certain extent like. They're hyping the hell out of AI right now. Is it that AI didn't deliver on the things that it was promised and therefore people kind of just moved away from it for an extended period of time?

 

Brian Gerkey (05:53)

I think there are a few factors at play. Use the word hype, which I think is just right. And I think here, I mean, we don't have to reinvent anything. It's the Gartner, I think it's the Gartner hype curve is often used to describe how these things unfold. And it's not just AI, it's any new technology. There's a huge amount of excitement and expectations that are frankly unrealistic and then there's a trough of disillusionment and then eventually you figure out how to make it useful. So I think that that's applying to AI and to some aspects of robotics today, just like it does to any new development. I think an additional challenge that we have with AI is that, know, AI is a term that's been around since the 50s. Today, we use AI to describe a very, a relatively narrow part of what you know, over a broader time horizon you would consider to be AI. Like I studied AI.

 

Brian Heater (06:43)

Why do you say now? I think of AI as being incredibly broad, just the way that it seems to describe everything at this point.

 

Brian Gerkey (06:49)

Yeah, but if you if you say AI today, I think there's a good chance you're referring to a particular approach to building relatively large neural networks to solve problems in a particular way. And that's that's what I mean by narrow like that's that actually excludes a whole bunch of things that I think can genuinely be considered AI, which are related to motion planning and search. And there's all these other things that have happened along the way. And that's that's kind of the curse of AI is that once you solve a problem, it's no longer considered AI. Like once it gets into use in a serious way, I mean, you look at like a lot of things that you use today, whether that's, don't know, Google search or it's a Netflix recommendation system, those are all AI based now. We just don't call it AI anymore. so the label, it's like the horizon keeps moving. So the bar keeps getting higher and we kind of forget that the things that they start to work, they become reliable, we no longer think of them as having that magic aura of AI. And so that gives you the additional appearance that you feel like it didn't work because you stopped paying attention to it. And it took a while to get into production, but then it ends up actually being things that we rely on.

 

Brian Heater (08:05)

Yeah, I wanted to pull on a thread, something you mentioned earlier, when we were talking about hype cycles, unrealistic expectations. And I know obviously everybody has their own idea of what we're dealing with right now. Is there anything at the moment that we're looking at ? that just really feels a lot further off than we're being led to believe?

 

Brian Gerkey (08:25)

As a roboticist, I look for ways to make things useful - as close as possible to today. And so if I look at what are the problems that we can solve today, I look around and I see we've got like an intrinsic, we're working mostly with robot arms. They're newer and better than the ones that came before, but frankly, they look kind of like the robot arms that have been around for decades. The thing is they can do amazing things if we add better sensing and better software to them. And so I think that there's a huge amount of extra productivity, new problems we can solve with those systems. Combining those with mobile systems in some way opens up new opportunities as well. An area where I would, I personally think that it's going to take a little bit longer than some people are currently forecasting to get into use would be the much more mechanically complex systems. And here I'm thinking about legged systems. So legs are great for solving certain types of problems. There are lots of cases where it's not clear that you need them. And so the trade-off of figuring out how to make those work really reliably,  you've got to, I think you need to be realistic about how long it's going to take to do that and what's the benefit, the payoff that you're going to get.

 

Brian Heater (09:43)

So interesting because like 201, just about every single person I asked this question to says ? manipulation, says dexterity, that that's the real difficult thing to crack.

 

Brian Gerkey (09:53)

I think that's also a hard thing to crack, I think that it's where if you ask me where we're over promising, I think that I think as a a as an industry, I see greater over promising on legged locomotion than I at least I have the sense of maybe you could tell me as you know, as an observer of all this, like, I don't feel like as a as a community, we're over promising on the manipulation.

 

Brian Heater (10:20)

So you mean like in terms of like how actually sort of useful it might be day to day versus like how far along it is perhaps?

 

Brian Gerkey (10:28)

Yeah,I think that we're, mean, frankly, see that like legged systems are getting very reliable, but it's not clear to me where that high value use case is going to end up coming from.

 

Brian Heater (10:43)

I really did sort of want to pull on the complexity problem, you know, in terms of this is kind of been, this is something I've been really focused on lately. You know, we see a lot of really highly produced videos coming through robots that look like they're incredibly capable in the field, in warehouses and in homes. Mobile manipulation is obviously sort of a really big part of that. As somebody who works with the company that deals so closely with picking and placing, largely stationary, but why is that such a difficult problem to solve?

 

Brian Gerkey (11:29)

Picking and placing things, if you want to do the same thing again and again with really, really high precision and we want to do it really efficiently, that actually, that is solved. Anything that you're willing to put enough money into to build up a bespoke assembly line for, you can automate. You can automate just about anything.  The challenges are: there's a huge amount of production that happens. It has higher mix or it has different volume characteristics and so it's not economically viable to automate it. And it's handling that variability. That's really the challenge. If you, almost any production system you can imagine, if you're going to have the same parts coming through, we can build a system, we can build what I would consider like a classical automation system to handle the manipulation, the picking, the placing, the part handling. If you want to do what A lot of manufacturers want to do today, which is I want to run, uh, I want to run 10 of that part. And I want to run a hundred of that part. And I want to run a 50 of that part. And I want to do all that in a single shift. And I want to have it be automated. That's really hard today. And it's being able to react to that variability. That is the challenge. And that's where, you know, our approach is to soak that up in, in the software. That's where we can, we can build and we have built perception systems that are able to with, you know, basically just give me an image and give me a prompt in the form of a CAD model. And I can find where that object is in the image. And more to the point, I can have the robot go and pick it up. And it's a part that we never saw before. We didn't train on in advance. That's something, the result of that is that you don't have to fixture all this. You don't have to, and you don't have to bring in the system integrator who installed your robot and have them reprogram it for you when you're switching from handling one type of part to handling another type of part. And it's those sorts of barriers are what has made it historically difficult. Just the cost of changeover is so high that most manufacturers think about these systems as being basically single purpose. And you can see this in the business model. You look at like, what does it take for a factory to agree to buy a robot cell? They usually want to return on investment that's on like a one to two year timeframe. That's not because the robot is gonna breakdown over one to two years, the useful lifetime of these systems is easily five, seven more years. The calculation they're making is, well, I'm not buying a multi-purpose device. What I'm buying is like a fixed appliance that is programmed to build exactly the one part I'm building right now, the one product. And when I changed my product line next year, I basically have to buy and program an entire new system. And so the value of that system I bought before is basically amortized down to zero, depreciated down to zero. With better software, you can start to think about these systems in a totally different way.

 

Brian Heater (14:27)

Yeah, and even like seasonally, I'm guessing that there's a big change to a lot of these systems that they want to be able to shift them to a certain extent on the fly.

 

Brian Gerkey (14:36)

Absolutely. Look at how cloud computing has evolved, right? There was a point when you would use one computer, one machine to do a particular job and you would put your software on that computer. Originally that was on premises and then you might put it in a data center. Now you don't think about individual machines. You think about, you have a compute workload. You have a set of kind of tasks you need done. You kind of throw that at a data center and it has in-house a few different types of machines. You some of them have a lot of GPUs, some have a lot of CPUs, some have a lot of storage, some might have TPUs or other sorts of processing - technologies. And there's a way to allocate your workload across those machines. Those machines are multi-purpose. They can take on one job in the morning, one job in the afternoon, a different job next week. What if we could do that for manufacturing? What if instead of having these fixed assembly lines that just do one thing again and again, imagine a factory which is full of these repurposable resources. And so I can toss my manufacturing, my production job at the factory, and it's deciding, there's an allocation or an orchestration system that's deciding, well, Okay, that task can be done by that work cell over there, which was doing something different yesterday. It's going to do this today and it'll depend on where the inventory is available. And we can start to treat these as like a matrix where we can adjust the flow of product through this matrix in a way that looks very, very different from the traditional assembly line that we've had since the time of Henry Ford.

 

Brian Heater (16:17)

It strikes me that maybe kind of one of the key thesis statements for intrinsic is that a lot of these problems are software problems where we thought maybe they were primarily just hardware problems.

 

Brian Gerkey (16:32)

There are still hardware problems that we need really smart people working on. -  We can talk about grippers and hands in particular. - But yeah, think that there is, I think the robot arm hardware, the sensing is really good. Our ability to use that sensor data is getting better all the time. I mean, we used to rely more on depth cameras, modern AI driven perception. is relying increasingly on just single color images without necessarily explicitly trying to calculate depth data. So the hardware, I'd say the hardware is willing and historically the software has been weak. And now we can really make some step changes there. The area where we still do need significant advances, and here I'm just talking about robot manipulators, is the so-called end effectors, the hands, the grippers. We're still by and large using relatively simple like parallel jaw pinch grippers or suction grippers for just about everything that is done in a day in day out reliable way because that's the technology that's been really well proven. It's totally hardened and we often adapt processes to live with the fact that we have these simple grippers. There's interesting three finger grippers. There's interesting, you know, more dextrous systems that are modeled after human hands. They are not nearly as widely used yet. We know that we need those capabilities and we need people working on the hardware there. And then, you know, as soon as it's working reliably enough that we could put it into production, we're happy to take it on.

 

Brian Heater (18:10)

Yeah, it's striking me talking to you that maybe like a big part of the reason why manufacturing and logistics ended up being such a fit for robotics and automation is because all of these problems are math problems. One of the reasons why I've really enjoyed writing about robotics is every time you write about a different company, like you learn about a different vertical, right? Like suddenly, you you're learning about farming or you're learning about the ocean. And obviously, working at a place like Intrinsic, you know, I'm sure that you know a lot about manufacturing that you didn't before, but at the end of the day, abstractly, everything boils down to numbers. And that's, I assume, something that...your entire team knows really well.

 

Brian Gerkey (18:56)

I can tell you that I and my team know a lot more about CNC metal cutting and sheet metal cutting and bending ? than I ever thought we would know. Because you do have to go deep into these, into the verticals in order to build something that's useful. I see what you're saying about it being about the numbers. I will tell you it is also, I mean this sounds cliche, it's also about the people. You can build a system that looks good in a spreadsheet, looks good on paper. If it's not usable by the people who are going to be on the shop floor interacting with it. If they don't understand the value that it brings to them, then they're going to reject it. And if you can't make the case to the management of a facility that their workers are actually going to get value out of this, then you're never going to make the sale. so there's, especially in these situations where we're like, we're in the process of asking, in many cases, these established industries, and some of the customers are relatively small, you know, ? A lot of the a lot of the places where automation has the opportunity to be taken up the quickest is in small sort of mom and pop machine shops. You really need to meet them where they are and offer them something that they're going to they see the value in it and they need to have confidence that they're going to be able to understand how to use it. You know, they don't want to be on the hook for making being having to call you to come back and deal with every potential problem. So there's a there's a relationship and a business ? aspect to it, which I've learned, especially over the last couple of years, you just can't ignore. Otherwise, you can end up building what you think is the perfect technology and just nobody ends up using it.

 

Brian Heater (20:37)

It must also be an interesting problem to work with, as you said, like Monpop versus a large corporation. And I suspect that one of the things that they would love to have is one of these kind of multi or I hate to use this, general purpose robot, you something that can be like adapted to various things so that they don't have to buy all of these, all of these different systems and as seasons change and products change that it can be adapted as well. ? One of the things obviously that we're all talking about right now is physical embodied AI and a big point of contention, I guess, that I've been hearing from a lot of folks in the field is this conversation around creating ? you know, one large generalized model that kind of works across embodied been or or small specialized model. And I got the sense when I talked to Torsten on your team that maybe intrinsics philosophy approach to that up to this point has been largely more specialized. Is that fair to say?

 

Brian Gerkey (21:45)

Yeah, Tristan Kurgers, our chief science officer and you know, we look to him to really set the tone for how we how we think about pulling in these new technologies and the term we're using ? now is ? a AFI or artificial functional intelligence. And it is I'd say what it reflects is our approach to wanting to provide these these functional components. that are usable today. So if we think about how we've historically tackled programming these robot systems, usually it's a standard systems engineering approach, right? You take a big problem, you decompose it into pieces, and then the pieces connect to each other in some way. So you have a piece that… takes in the image from the camera and you have another piece that processes the camera to get depth data and then another one that processes the depth data to find the object and another one that decides where you're going to grasp and motion planning and so on. And you have this kind of distributed systems approach. What I think of us is doing is saying, well, look, there are these new ways that we can solve, though we can… basically provide those functions. We can solve those problems in different ways. And so often what we're doing is taking a modern machine learning AI based approach, and we're taking out one of those existing blocks and we're popping in an AI powered version of it, which is better and more reliable, more robust, lower cost, easier to use. Now over time, what we're seeing is there are these adjacencies, like, you you talk about grasping, do I need to explicitly compute the where an object is in the scene and then pass it to a grass planner or can I ask a slightly bigger AI system to just go ahead and tell me where to grasp, right? There that you can start to skip these steps and we're exploring that as well. Now, if in the fullness of time, these this coalescence, you know, keeps going on, the blobs keep getting bigger to the point where you end up with. more or less one blob that is doing solving the whole problem. If we get there, sure, that might be the right approach. I would, to me, that's not an end in itself. I think that, you there's a, there's a, you can try to draw parallels here with like, you know, the search for AGI, which is, but that's more of like a, that's like a philosophical goal to me. That's not, that may or may not produce useful technology. And I'd say we're really focused on producing useful technology. If and when there's this one model that can control the robot, especially if it can control all the robots, then I would like to think that at Intrinsic, we've laid all the groundwork so that we're ready to host that AI model and connect it out to the physical world.

 

Brian Heater (24:34)

Is that approach to somewhat in contrast to, know, ? obviously intrinsic is part of is a part of alphabet. Google is a part of alphabet. ? know, DeepMind is a part of Google. Are they sort of starting with a blob first approach with Gemini, would you say?

 

Brian Gerkey (24:53)

It's a matter of time horizon. if you look at, you know, Alphabet has a portfolio of ? efforts that it invests in. And I think about Intrinsic as being the much more commercially oriented product focused, getting these systems out to people today kind of effort. And that means that, you know, we're going to make those, we're going to make decisions like we, if there's a smaller functional block that we can swap in today, we're going to do that when our partners who I think about as being kind of upstream in the, you know, I think they're further back into kind of the, in the pipeline between research and development and technology deployment. They're, know, DeepMind is, is a world-class research organization that's doing a bunch of work that over time, we're going to figure out how to take on board and then put out into the world. I mean, in fact, our one-shot vision model is the result of a collaboration between Intrinsic and DeepMind. We took some of the work that they did, built on that, extended it, trained it in ways that are relevant for the kinds of problems that we're trying to solve. So I think that's going to continue.

 

Brian Heater (26:10)

Yeah, it's interesting. And I know obviously this predates your time with ? Intrinsic and Google. But ? I get the sense that maybe this is sort of a big part of what's happened to a lot of these, what had been ? moonshots or - Alphabet X companies is this push to get from sort of like the abstract moonshot to, hey, what can we do right now to actually make this like a company that is actually making a product. What can we do to actually graduate this from the lab? And perhaps that's what a big part of what started to sort of separate intrinsic, the idea of intrinsic from what would have been like Google Brain or DeepMind or everyday robots at the time.

 

Brian Gerkey (26:59)

Like you said, I wasn't here at that time. but the way it is worth recognizing that we've got we've got alphabet betting on what we're doing here. And so what that means is that for that, for this to be meaningful for alphabet success needs to be big here. I think of it as a very big game that we're playing. there's, you know, success for us here. It looks like having a very broad, very ubiquitous use of the kind of software platform that we're building. Now, to me, to get there, we have to go through these all the intermediate steps. You don't jump from the ground directly to the moon, right? You've got to, there are these steps along the way, some of which in the moment look like they look kind of pedestrian. You you're working with a small machine shop and you're figuring out how to get a robot in there that solves a problem for them so that they can run a third shift overnight because they don't have enough employees to be able to meet their production demand. And those, those sound like kind of, you know, one by one sorts of interactions, but you have to go through that in order to learn everything that you need so that you can scale up and figure out how to, how to meet, like actually win that bigger game.

 

Brian Heater (28:18)

This actually kind of gets back to, think, what you were talking about before when we were discussing bipedal robots. you know, I've seen, again, a lot of companies right now that are talking about speed to development of a bipedal system. And it's impressive, right? I mean, it's impressive to see a company build hardware in a year that, you know, from nothing that can suddenly walk. But that's a different goal than building a company that has ROI and can actually like has a saleable product.

 

Brian Gerkey (28:48)

We should not ignore how amazing it is that it's possible to do that in such a short time frame. And I think a lot of credit here goes to folks like ? Mark Ray Burton, Rob Plater, who really pioneered a lot of this work at Boston and AmEx for literally decades. And now the things that we can kind of take for granted that, yeah, if you wanted to go from scratch and build a quadruped or a biped, we know how to do that as a community of practice in the robotics industry, which is extraordinary. ? think you've raised the right question, which is, okay, that these systems, they're clearly going to be areas where they are useful, but what specifically are the use cases? What's the value proposition to get people to take these systems on and deploy them? And I think that's a case where we still struggle a little bit. And it's probably kind of like what we were talking about earlier, the kind of, let's say, one big artificial, generally intelligent. model versus a collection of smaller functional ones. I think this is also a question of time horizon. I have no doubt that the legged systems are, we're going to find the right places to use them. I think we're still working on that right now. And in the meantime, I'm perfectly happy to be providing value to people today with the, you know, kind of the arguably less sexy robot arms. But I would argue that they're, you know, they're, they're doing really useful, productive work.

 

Brian Heater (30:15)

Yeah, this is one of the many things that and people who watch the show will know that - Willow Garage is a longtime fascination of mine. And this is one of the many things that's interesting to me is there did again in hindsight from somebody who is not at all involved, but there did seem to be maybe a little bit of tension between the idea to be a pure research facility and the idea to commercialize to bring to bring something to market. that a fair way to kind of categorize Wheel of Garage's existence?

 

Brian Gerkey (30:49)

It was never exactly clear what… what the mission of Willow Garage was. what was amazing is the autonomy that we had to kind of discover and construct that mission as we went along. Like in the early days of Willow, I mean, this has been written about, but there were, you know, the early days there was actually an autonomous car project. There was an autonomous solar powered ocean going boat project. And there was the, became the personal robotics project, which, you know, that eventually created Ross, created the TR2. And there was a point where Willow was doing all those things. And that was kind of the birth of it. In like 2008, we narrowed it down to look, we think personal robotics, the personal robotics program, as we call it, that's the big bet that we want to take. And that was where we kind of constructed the mission on the fly. We said, look, we think the most the highest impact thing we can do here is to accelerate the state of technology development in this industry by providing the smartest, most talented people in the world with the right hardware, but not only hardware, they need software too. Let's give them the best hardware and software so that they can start to solve some of these hard problems. And that's, that was what we did with the PR2 and we gave those, gave the first dozen or so out for free, but then we sold them. - I mean, it wasn't, it was never going to run Willow Garage on revenue selling those robots, but we did sell another few dozen of those robots. The lasting contribution arguably has been the software. And in that case, we decided early on that a fundamental proposition is that we're going to make all that software available open source. And that became ROS. Along the way, we were driven by this desire to have impact as opposed to it. So I wouldn't say it was exactly like, it more of a research organization or is it more of a commercial organization? It was an impact organization. Where the debates were, were like, what type of impact should we have? And there definitely was tension there because impact is a powerful word, but it's also an empty vessel in some ways. You can pour whatever you want into it. And so there definitely was tension, especially in the later years to try to figure out like, okay, we've shipped the PR2s, Ross has achieved critical mass. Like basically what's next? How do we pick the next thing to have impact from?

 

Brian Heater (33:15)

Was there tension at the time around open sourcing?

 

Brian Gerkey (33:19)

Not at all. that was a, and that's a huge credit to Scott Hasson, who, ? who, you know, founded Willow Garage. He had had his own early commercial success in the first internet boom in the nineties. You know, he founded, ? e-groups and then when it was, ? instrumental in the early days of Google. And what he saw was that they were able to do the things that they did in the nineties with new internet companies because they started from a really strong open source base. It was what you would often call the lamp stack Linux, Apache, MySQL, and PHP, Perl, or Python, depending on your preference for programming language. And what that meant was that if you had an idea for a company like eGroups, which became Yahoo Groups, or even Google, you don't have to start by writing or even buying ? an operating system, a web server, a database, or a compiler or interpreter. That stuff is just available for you for free and at very high quality and maintained by a global community. And he, Scott, saw Willow as the opportunity to provide that same, you know, standing on the shoulders of giant starting point to the robotics industry. And so it was a fundamental proposition from the beginning that what we wrote was going to be open source. And even more so more specifically, if you get, if you're an open source licensed nerd, it was going to be permissively. open source. And at the time we used the BSD three clause license today along the way we switched to Apache two, but it basically it's not a copy left license. It's not a GPL like license. It's a, it's a license that says, look, here's the software. We hope you find it useful. You're welcome to take it, change it, incorporate it into your product, redistribute it. And you are not obligated to give anything back or share your results. We really hope you do. And you end up with a small fraction of people contributing back, but it was that permissive, open source, while at the same time being high quality and well supported and well maintained aspect of Ross that caused people to be willing to incorporate it into their products to build companies around it.

 

Brian Heater (35:25)

You had given away the first several PR2s and I remember very distinctly the first time that I met Melody Wise was in person at Fetch and she was showing me the research robot that they were working on and it really in a lot of ways it was kind of ? I guess like their successor or their take on the PR2 and very small overall percentage of, of what the company was doing and, and certainly not moneymaker, but what she said to me, and this has really stuck with me is like, you know, that's kind of like height. That's how you get in. Right. Like you get that, you know, you, you get people to work on your system. Right. And then, you know, you get, you get really smart, like, you know, kids in universities, you know, working on it and, and you've kind of like hooked them to a certain extent. then that perspective, made a ton of sense that, ? you know, the PR2 would be going out in the world and that you would be releasing Ross in that way.

 

Brian Gerkey (36:27)

And there's a bunch of other aspects. mean, another really important aspect of Willow's success and a lot of this credit goes to Melanie for kind of creating and running the program was our intern program. So we, and just in terms of involvement, so it's one thing to give the robots away, to give the software away. We went a step further and had at any time a huge number of interns coming from the best schools, the best labs in the world. We hosted them at Willow and the proposition for them was, you know, these are some of the best people come in the, you They're in grad school labs all over the world. They could go work at big tech companies for the summer. If they go do that, then they're working on something that's secret, it's proprietary. They're going to have to leave it behind when they go back. And our proposition to them, our requirement in fact, was you come here, everything that you work on here is going to be open source. You're going to be pushing it directly into open source repositories. And then we sincerely hope that when you go back to your lab, you keep working on it. And we started to think about them as kind of a contagion vector. They would take what they learned about ROS, take it back to their labs and get their labs using it. And that was, I mean, I did not appreciate it at the time in retrospect. mean, it was a master stroke how like the impact that that program had.

 

Brian Heater (37:45)

really wanted to get your take on, again, as somebody who has been in this a while and has been so closely connected, specifically through - Open Robotics, to Ross on what the state of the open source - robotics, I guess, seen at the time. There's been a lot of really interesting...  A lot of really interesting things have transpired over the last several years. Hugging Face has gotten involved. They've got that basically like a GitHub-like repository. ? Obviously, there are a lot more cheaper robots, so that's probably in some ways a little bit closer to that original vision, I guess, of a PR2 in terms of accessibility. ? We've seen some...Projects. Yeah, I was I was rooting for the K scale team over in the Bay Area. I don't think that quite worked out but ? Yeah, I'm curious. What what do you think about the state of it right now Nvidia? Obviously like a player in that scene too and what role? ? In intrinsic is sort of playing and I guess helping to facilitate some of that

 

Brian Gerkey (38:57)

Well, first of all, I think it's...The state of that community, think, is really strong. ? I saw this firsthand at Ruscon, the developer conference that we run every year ? that we did in Singapore this year. Now when I say we, I should also explain I'm CTO for Intrinsic, running the engineering team. I'm also still the chair of the board for the Open Source Robotics Foundation, which is that that was the original entity that became Open Robotics that we created in 2012. And that's still, that's the organization that really owns ROS ? as an open source project. So in my role there, I helped to put on ROSCON. We had approaching 1,000 people get together in Singapore for three days and talk all about the latest developments in the use and development of open source software for a huge variety of ? use cases. there was a...at unsurprisingly, a lot of attention paid to how do we take this system and make sure that it's it's fit for purpose when we bring in these new AI models. So when we're, doing this more like, ? basically learning, ? trained systems versus kind of classically programmed systems, where is where is Ross working well? And where do we need to improve it? And those are really great discussions to have. also, they're also presenting on, you know, what are the latest applications that they're putting out in the field. So I think the state of the community, I felt really good. That's one of my happy places each year is to end up at RossCon and just see all the different things people are doing with it. From intrinsic side, we had a booth at RossCon where we showed off some of the work that we've been doing to integrate Ross with the intrinsic stack with flow state, which is our, that's the intrinsic product that is a end-to-end ? simulation development deployment system for ? today for manipulators. So we were showing that off. We were also showing presenting on the work that we just do contributing to the ROS ecosystem. So I've got, within my engineering team, there's part of that team is just dedicated full-time to making open source contributions to ROS, to Gazebo, to OpenRMF, to the the ? build infrastructure, actually produces all of the binary packages that most of the users out of the Ross ecosystem end up installing and running on their systems. And that to me is a really important commitment. It's rational as well. I don't think of it as an act of altruism, because if it were, it would be fragile. If we think about it as a gift, it's always something you could decide not to do. And I think about it as a completely rational self-interested investment because the goodness of like we depend on that ecosystem and its health increasingly for our own business. And this is in the end why companies contribute to open source. They're only going to do it in a persistent, durable way if it is important to their business.

 

Brian Heater (42:08)

You and I spoke a number of times when the acquisition was happening. ? I'm sure that you got... I'm sure there was lot of freaking out happening in certain circles. I'm seeing that play out in similar way with Arduino and Qualcomm right now. And I get it, right? I'm sure that you get it too, as a nerd and somebody who's been involved in those circles as well. And you had to be sort of...You had to be an ambassador, right? I mean, you had to be one of the people who's out there sort of explaining like, here's how things are going to go moving forward. Were those were those difficult conversations? it was it difficult to explain to people why you were making that move?

 

Brian Gerkey (42:52)

Sometimes yes, sometimes no. And I remember, I think you interviewed Wendy, our CEO, me ? at an event around that time a few years ago. if I recall, what I said then was, look, in my view, this is going to be the best outcome for that open source community that we've created. But you shouldn't just trust me. You should come back and check the result later. And that's what I would say is, you know, now we're, ? let's see that we're three years now, ? actually, three, three, three years later, and I would say, let's go look at how's how's Ross community doing? Are people still coming to to Roscon and showing off their results. I would look at the Open Source Robotics Alliance, which is a new, that's a trade association, a membership based trade association that Open Robotics created. And I would look at the huge uptick in membership that we've got there. Companies, know, Intrinsic is a member, yes, but we're one of several dozen companies that have come and are paying to be part of the Open Source Robotics Alliance to have a seat at the table to help steer the future of the community. So Was it were there difficult conversations at the time? Did people did people freak out? Did people think like, my god, this is gonna you know, ? Google is buying Ross and everything is going to go to hell. Yeah, some people thought that we went out, we talked about it. We were honest, we were open and transparent. But and again, I said, look, the the proof of the pudding is going to be in the tasting, you know, come back later and check the result. And you can decide whether you think we followed through. In my view, we did. ? If somebody has a different opinion, I'd love to hear about it.

 

Brian Heater (44:41)

I'm not going to get all like flowers and sunshine about this, but again, one of the things that I've really appreciated covering this, obviously, you know, it's a bit it's a business. These are all business. There is competition, not not everything that happens ? in and around companies is is is great and mutually beneficial for everybody. But I do. Compared to a lot of other things that I've covered compared to. consumer electronics, for example, which I covered for a number of years, ? people seem genuinely impressed with each other's work. People seem to like working together, like collaborating. And a lot of that stems from the fact that at the end of the day, it's still like a pretty small community that everybody went to school together. Everybody worked at the same like four or five companies together. Everybody knows each other and everybody's probably gonna end up working with each other again at some point.

 

Brian Gerkey (45:42)

That's one of the things that keeps me in this industry and this community is all those parts of it. think it is a... especially if you've been working in robotics for a reasonably long time, like more than 10 years, then you almost certainly got into it not because you thought it was going to change the world or it was going to make you independently wealthy. You got into it because you love the technology and you're just fascinated by robots. And so because all those other things they didn't seem possible or realistic back then and I think that that the especially the folks who've been in it for longer We kind of we set a norm, right? we were we we worked together in a variety of different ways including through a lot of open source development and We established a kind of a standard of behavior and we got together at conferences and we got to know each other I I think this is frankly not not unusual among young industries. Perhaps we've just done it and held on to that spirit for longer than it usually happens. But I think that aspect of it is one of the reasons that we all do stay engaged is why we don't go off and try something else, right? You don't have ? people jumping off to other industries. They really stay committed to what we're doing here in part because of those relationships.

 

Brian Heater (47:05)

Brian, always a pleasure. Thank you so much for taking the time.

 

Brian Gerkey (47:09)

It was my pleasure. Happy to be here. Thanks, Brian.

 

Brian Heater (47:14)

Really great conversation. Thank you so much to Brian Gurkey and thanks to Leonard for helping set that up. Thanks to you as always for tuning in. Please like and subscribe and don't forget to subscribe to the Automated newsletter over at automated.fm that drops every Thursday morning with exclusive features that you can't get anywhere else. Thanks again and we will catch you just about this time next week with another episode of Automated.