Autonomous Construction Sites and AI-Powered Heavy Equipment with Bedrock Robotics
Boris Sofman is the CEO and Co-Founder of Bedrock Robotics, a company turning existing construction equipment into fully autonomous fleets through same-day hardware upfits. With over $80 million in funding from Eclipse, 8VC, NVIDIA Ventures, and former Waymo CEO John Krafcik, Bedrock is tackling a major bottleneck in the global economy: a massive construction labor shortage just as demand for data centers, clean energy projects, housing, and manufacturing is skyrocketing.
In this episode, Boris shares how his experience building autonomous vehicles at Waymo inspired him to apply similar AI and machine learning approaches to heavy equipment. He explains why full autonomy matters in construction, what it unlocks for efficiency and safety, and how Bedrock plans to accelerate infrastructure and industrial development through robotic automation.
Episode recorded on Sept 30, 2025 (Published on Nov 13, 2025)
In this episode, we cover:
[02:45] Boris’s background in robotics and autonomous vehicles
[04:50] Learnings from Waymo applied to construction
[10:09] Boris’s predictions for autonomous vehicles in the future
[18:44] Why he left Waymo to start Bedrock Robotics
[22:59] Choosing construction as the first market for autonomy
[25:26] How Bedrock upfits machines without permanent modifications
[26:25] Why excavators are the first target use case
[28:20] Training AI to navigate changing job site environments
[30:54] Skipping teleoperation and going straight to autonomy
[35:52] Bedrock’s GTM focus on heavy industrial sectors
[40:46] How to work with traditional industries effectively
[43:55] How autonomy solves labor shortages and safety challenges
-
Cody Simms (00:01):
Today on Inevitable. Our guest is Boris Soman, CEO and co-founder of Bedrock Robotics. Bedrock transforms existing heavy construction equipment into fully autonomous construction fleets through same day hardware retrofit installations. The company has raised $80 million in seed and series A funding led by Eclipse and eight VC with participation from NVIDIA's Venture arm and former Waymo CEO, John Kraft Chick. I wanted to have this conversation because we're seeing two massive trends collide. We're facing an unprecedented construction labor shortage just as we need to build more critical infrastructure than ever data centers for AI, clean energy projects, housing, manufacturing facilities. Meanwhile, Boris spent five years at Waymo helping put the first autonomous vehicles on public roads and now he's applying those same machine learning approaches to construction equipment. This represents a broader shift where breakthrough technologies proven in one domain are migrating to solve bottlenecks and building everything else. I have some fundamental questions. Is full autonomy even needed in construction? What does it actually unlock? How different are the technical challenges compared to self-driving cars and how do you drive adoption in a traditional industry like construction? From MCJ? I'm Cody Sims, and this is Inevitable. Climate change is inevitable. It's already here, but so are the solutions shaping our future. Join us every week to learn from experts and entrepreneurs about the transition of energy and industry. Boris, welcome to the show.
Boris Sofman (02:05):
AIPleasure to be here.
(02:06):
Let's dive in from the start. Tell us what is Bedrock Robotics?
(02:11):
Bedrock Robotics. We're a company that makes technology for enabling fully autonomous heavy machinery, so these are the sort of machines that you see in construction, excavators, bulldozers, wheel loaders, but they exist in tons of sectors, agriculture, mining, lumber, garbage movement, and so we were working to bring technologies, enable these machines to be fully driverless and be able to perform the work at exceptional quality and safety and that actually enables some pretty amazing benefits to these industries that are suffering both from labor challenges as well as just operational challenges.
Cody Simms (02:45):
Boris, maybe talk a little bit about the background that you bring to bear into the company as well as quite a few of your co-founders.
Boris Sofman (02:51):
The company's about a year and a half old, but a lot of us have spent most of our professional lives in autonomy and autonomous vehicles and robotics and AI. Those were my roots. I was PG in robotics at Carnegie Mellon, work on some of the early off-road autonomous driving work and to be, one of the things that gets people into robotics is cars was one of the holy grails that seems to always excite people. There's just something magical about it. Vehicle driving itself and went through a journey in consumer robotics with a company called Anki, and then I was at Waymo for about five years, which was very, very relevant to what we're doing today. I started in summer of 2019. I was there for five years leading autonomous trucking and then also various technology teams that supported both cars and trucks, perception, reliability, freeway driving, and we helped launch the Jaguars into San Francisco, which Waymo cars, the product's pretty incredible.
(03:41):
It's now well over a hundred million driverless miles. It's growing exponentially, millions of miles a week, superhuman levels of safety, and so we were able to really see this mature from early R&Dz to this exponential commercialization phase, which is on its own magical, but one of the most interesting takeaways was just how powerful these machine learning technologies were that we'd brace at Waymo, going from traditional robotics approaches where you're using search and heuristics and engineered solutions to effectively using data to explain how you solve these incredibly nuanced challenges like driving in San Francisco or driving a class a truck. And we saw not only did it solve these problems, but it had the ability to really start to generalize and the system would extend super well to other cities like Los Angeles, Phoenix, Austin, and then even from car to truck and you started getting more and more of a subsidy on the data side where it just became less and less expensive in terms of new information to become very competent and that was pretty magical, and so we started to think about how do you apply this to other areas where you have a lot of nuances in the type of work you're doing and you have a lot of different platforms that can benefit from these superpowers.
(04:50):
Automation for heavy machinery very quickly jumped to the top of the list where you have astronomical amount of work and demand that's growing. You have onshoring and manufacturing, you have data centers that have to be built, you have housing shortages, you have infrastructure needs, and at the same time, I've never seen such a divergence with the labor pool that's available where a lot of the GCs we talked to and are partnering with, they're telling us that they're seeing astronomically tough ratios of people coming into the ecosystem and leaving, and some have given assessments as many as 60% of workers are retiring in the next seven years, and so you have this incredible talent train because this is not driving where everybody can pick it up and inherently knows how to do it. An excavator has a minimum of six degrees of freedom and you have this nuanced complexity that takes years to pick up and you have huge safety challenges, so it's an exciting opportunity.
Cody Simms (05:44):
We were talking a little bit before we started recording. We met more than 10 years ago when you were building a consumer robotics company in Anki and I was at Techstars leading the Disney accelerator. The old phrase that the future, it always feels slower in the short term and quicker in the long term. I'm curious, when you were building essentially these little consumer robotic truck looking things, if you thought a decade from now you would be creating autonomous versions of the real thing and how fast in reality that has seemed, is that something you would've anticipated a decade ago?
Boris Sofman (06:17):
We were trying to be very practical and think about where can we have a really big impact on a product experience and value to customers given the technology of the day. And so we started to focus in entertainment spaces where you could do some pretty powerful things in software and leverage the technology. There was at that point actually getting boosted by smartphones where every component in a smartphone was getting manufactured at hundreds of millions of units, so suddenly something that should cost $15, you could find for $1 how this library of really covered things you can do using microcontrollers and wireless chips and memory and batteries and all these sort of elements. We started working on it, but at that point there were definitely gaps. You could not do the sort of things that we're doing with AI today at that point in time and we recognized that because we always thought that there's this progression of things you could do going into home and healthcare and more service applications and so forth, but the technology just wasn't ready.
(07:10):
And funny enough, Waymo has a little bit of that same trend where it started in 2009 out of the excitement coming out of the grand challenge in urban challenge, everybody in the industry underestimated how challenging it would be to actually get a driverless car on the road and a lot of that is because of the super delicate long tail safety challenges and nuances that aren't visible when you drive around for 45 minutes when you fast forward, there's something really amazing that's happened in the last six eight years where on the AI front you just have such an incredible tool set that feels like the only thing even remotely comparable to it is the internet where you now have this incredible firepower to solve problems that you can now use in all these creative ways almost everywhere, in every industry in the world, a lot of it is around the ability to capture incredible nuances when you have the right data to solve problems.
Cody Simms (07:59):
I'm hearing a trend of this growing ability to push the learning and push the control systems away from the robot itself in Anki. It was pushing it to the smartphone and with Waymo it's pushing it out to a full machine learning instance where the vehicles are all learning from one another. Am I following the thread there correctly?
Boris Sofman (08:18):
We definitely did that, so we took advantage of this amazing compute that was available and it became the brain for the experience and that was a very clever a way to approach it where you could have a fantastic interactive experience for a very attractive cost because we were justs of the ecosystem that was kind of developing around us. It was pretty fantastic for a consumer. In the case of Waymo, it's interesting because the price of compute started really coming down, capabilities increased, but one of the biggest shifts there was going from having engineers writing code to try to capture nuances of how do you optimize these cost functions and solve these problems to optimizing these massive scale models that could capture these nuances in a way that was never possible before. And to a degree, it required this ecosystem of more compute and cloud functionality and the ability to put a computer inside the trunk of a card that was just unfathomable 15 years ago, but all these trends progressed and then they completely flip the equation where instead of engineers writing code, it becomes around how do you get the right data to explain the nuances that you're trying to capture and the intuition around how you manipulate these models and evaluate them.
(09:25):
And so there's an element of keeping your eyes wide open and trying to embrace the trends that are happening. And I still remember inside the company there's some really brilliant people that started pushing in these directions, but it was very controversial at the time because there was a system that was working and driverlessly in a small scale in Phoenix, but there was a really good foresight from the team on thinking about what would actually scale to solving, driving across the whole country and the world. But you're right, sometimes the timing matters a lot. If you're too soon, there's no prize for being there 10 years early.
Cody Simms (09:57):
Curious, given your lived experience building that and then we're going to get into Bedrock and talk all about your current company, but where do you think we're going from just an autonomous passenger vehicle perspective? What are the next five and 10 years look like?
Boris Sofman (10:09):
Suddenly all these tools become available, they're quite incredible and things that were science fiction are almost easy off the shelf elements now, and this includes even hardware sensors that were able to use at Bedrock. They didn't exist when Waymo started. One of the reasons that Waymo had to invent the whole stack is because it was the first pushing on it. I think sometimes about the sort of things we could have done back at Anki as well if this technology was available on interfaces and the magic of these experiences, but where it's all going. What's interesting is that everybody's looked around and seen just how powerful LLMs have been with ChatGPT and Gemini and everything where they're taking advantage of this infinite corpus of data on the internet and other places and been able to train these models and then with a little bit of work you can now mold them into magical experiences that solve problems that felt impossible even three years ago and that's amazing.
(11:00):
It is very, very clearly on the way to more and more applications that will impact how we interact with the world and improve efficiency and so forth. What's easy to forget is this digital type of world that's purely digital pushing bits around that's maximum 20% of the world's GDP and the rest of it is the physical world. It's transportation, it's manufacturing, it's construction, it's agriculture, it's physically solving problems in the world and that's the frontier because now all of a sudden robotics has had a handful of these false starts where people thug got really excited and didn't work. I think this is the beginnings of not all at once, but a sequence of really meaningful jumps where the physical world is going to get smarter just like Waymo's are driving around multiple cities today and how machines and construction are going to start operating themselves. It's just the beginning and I think that's the part that actually may in some ways be more transformational than the pure digital version.
Cody Simms (11:51):
On the direct transportation side, do you think people who are less than 15 years old today are ever going to drive cars? Is it going to move to almost a pure autonomous world?
Boris Sofman (12:03):
I got three kids 11 and under, so curious about the same thing. When you look at San Francisco without exaggeration, I think there are areas where one out of 20 cars on the road is a Waymo and it's an amazing product. Most people prefer it to the traditional ride hailing approaches and it's over a third market share and I would guess not far off from being on its way to half and above because a lot of times it's limited by volume of cars that are available and so forth and it's a magical experience. There's a dozen cities have been announced feels like every week there's a new city that's in testing including giant cities like New York and London and Tokyo, and so it's very, very clear where this is headed, this is going to go and spread out. There's going to be robotaxis that become a really viable solution that market use of it might decrease and then eventually you go to personal car ownership and logistics and everything else that'll take longer than I think people expect because the technology could be ready, but there's life cycles of vehicles on the road.
(12:58):
There's personal preferences and fears and there's just the physics expanding like this because this isn't software, this is infrastructure, this is physical manufacturing. There's cost elements of optimization that are necessary, but if you fast forward 10 years, I think there'll be a sizable percentage of cities where the default is probably you get into a driverless car because it's just a superior experience. Nobody wants to deal with parking and headaches and a driver yelling at somebody on the phone and you have this very predictable safe experience and so I think that's powerful and then that changes where people live. It changes how people tolerate commutes. Long-term it probably changes what your experience actually is in a vehicle, especially if you have a longer drive. I don't necessarily know if the country will tolerate, maybe there're cities that will just not allow physical manual cars. I wouldn't be surprised by that, but that's a sort of transition where I would imagine it's like a decade to have this be super, super prevalent and in another decade to have this really be absorbed to where governments will start to constrain the alternate because it's just less say you can as a bunch of headaches associated with it.
Cody Simms (14:03):
It feels like a decade ago everyone was like, oh, it's going to be here tomorrow. Now it's actually really here. Truly here.
Boris Sofman (14:11):
San Francisco feels like five years out from most cities. It's already there.
Cody Simms (14:14):
It's here. I was just in Las Vegas a weekend ago and the Zoox is are everywhere, different brands, different makers, they're rolling out. Obviously Tesla has their full self-driving, you still have to sit in the driver's seat with it, but they're getting there with Robo Taxis too. And so it feels like it is happening and it's just a matter of how quickly and what the experience looks like in the future. I assume there's going to be mostly some kind of hybrid experience for a while.
Boris Sofman (14:38):
There's a lot of debate on when this becomes really amazing and a great experience and super versatile and available and also cheaper eventually. Does this mean that ride hailing goes from 1% of miles to 5% of miles? I'm not sure about that elasticity because there's also a convenience to having your own car and so forth, but it's very clear that the next step is you want this in your own vehicle. I would pay $25,000 more for a subscription for a car to drive itself. It's just transformative when you think about the amount of hours that go into the lifetime of a vehicle changes everything.
Cody Simms (15:08):
Shifting gears slightly, you were talking about everything going on with AI and ML and only some percentage of it is the pure digital world and over time it's going to expand heavily into the physical world that we all live in with Waymo and transportation maybe being the first real big use case of that. I read an energy analyst report today that was saying that they believe that industrial robotics will eventually equal and grow greater than from an energy usage percentage. The amount of energy that the traditional software data center world is using, which I thought was fascinating to contemplate given how much attention is being spent right now on the data center energy footprint.
Boris Sofman (15:51):
I believe it because it takes energy to move physical things. Obviously a lot of machines are still diesel that'll eventually go electric, but you look at the sheer volume of productivity in manufacturing, assembly, logistics, fulfillment, warehousing, every one of these has an opportunity to automate and when you hit the gas and push on onshoring of manufacturing, there's no choice you want to manufacture in the U.S. The United States can't sustain manufacturing the way it's been done in Asia. The physics don't work. There's not enough people, even if you set aside labor costs and everything physically, you can tariff t-shirts as much as you want. You can't pay people in a feasible way in the United States to go and make clothing. But when you introduce autonomy, it's the only way that this works because when you have unemployment at 3% and there's pressures to get jobs higher, of course it has to go towards automation.
(16:43):
But what's interesting is that it's very expansionary where we went through this in a few points in time in the history of the US where industrial revolution completely upended the way we produce everything from farming to manufacturing. There was all the same concerns that people had about, well, what does that do? What about the jobs? You can always isolate it and see what might be disrupted in the way things are today. But what was always hard to predict is when the productivity skyrockets not just by 2x but by 10x or 40x and suddenly things that were economically and feasible become totally feasible, the amount of work that gets created that just fundamentally didn't exist before is astronomical. And what happened when the dust settled is that the average salary more than doubled and the productivity skyrocketed by some absurd amount and it became expansionary from a GDP standpoint.
Cody Simms (17:32):
I guess that's a Jevon's paradox in the physical world I guess.
Boris Sofman (17:36):
Exactly. And there's an interesting element of that is when costs and power usage of AI gets cheaper, is this a asymmetric exponential growth that becomes possible, but the energy we got to get through and it's interesting because now the traditional means just aren't keeping up. And so there's obviously a big push for reopening nuclear as a capability. There's an interesting statistic. So we studied Caterpillar closely obviously because they're a very close company in our space and we're starting to get to know 'em well, but their energy division is past the revenue of their construction division and most people would be shocked by that. A lot of people don't even realize they do energy solutions, and so they make these incredible turbines that basically convert gas to electricity and have all these various forms of solutions because a pretty sizable percentage of data centers are just not on the grid. They make power from something else, whether it's solar batteries or gas or whatever the case is. And so people are just getting really creative about how to meet the energy needs, but it's only going one direction and I really hope the US can keep up because it's one risk where I think the ingenuity in the US just surpasses any country, but we've dug ourselves in a hole on the energy footprint compared to other countries.
Cody Simms (18:44):
What was the AHA that led you to go build bedrock? You were obviously building something incredibly impactful at Waymo that was truly changing the world in terms of how we all navigate the world itself. What led you to say, oh, I've got an idea for something else I want to do?
Boris Sofman (19:02):
It was a very, very natural checkpoint where almost like a capstone to what we've been working on for five years, which was driverless on freeways. And so our team was leading a lot of that effort and that was the main focus for a big portion of my time at Waymo. And then we went driverless at the end of 2023 and then that's been scaling since and then eventually gets to enough volume where it plugs into the mainline operations. So that was such a fantastic zero to one that it became almost a natural point to just look around and say, okay, there's a hundred percent an opportunity to have incredible time over another five years or 10 years that Waymo. But there was almost this opportunity where I started to think where else can these technologies be applied? And I was always excited about other areas like manufacturing and industrial machinery, like what we ended up doing and part of me also just missed the building of a company.
(19:54):
So I think I had a wonderful time at Waymo in a lot of ways. We were almost like a startup within Waymo as we were assuring the early stages of a lot of the technologies and programs, but there was just a very natural point in time where there's so much interesting opportunity with these technologies that once we started looking around and a few of us that started the company really grip onto this idea, it felt really logical. They hit the gas and do it and really happy we did because a year and a half in, but there's so many carryovers in the learnings and also just amazing people that joined us from our old teams and a lot of our old colleagues that there's relatively few moments in life where you get a chance to take its wing at something so big and it feels like the timing's right that technology states right teams, right pulls, right.
Cody Simms (20:36):
So as you were looking at different potential things to go tackle and you're a leading expert at commercializing autonomy solutions at this point in robotics, what was it pulled you into construction and what was the work you did to say this is the one we're going to go do?
Boris Sofman (20:53):
So we were looking at a lot of spaces and we studied agriculture, construction, mining, manufacturing. There's a few things that really stand out. I really started appreciating across both on and Waymo how important the business model and the physics of an industry and the overall business model are to the success of a company. And so you look at other industries, they might have machinery, but the percentage of the cost of that industry or the cost of the customers that goes towards it is relatively small. So for example, in agriculture, you can do a lot with a few machines and a relatively small amount of percentage of the labor goes towards operating that machinery. A lot more goes towards fertilizer and water and seed and then manual labor and things like that. When you look at construction, it's higher. There's just a lot of work that has to be driven by these sort of machines and you also had machines that already existed, which is a deceptively huge advantage where you look at the sort of things that are happening in humanoids and people have to go and build their own machines. They're incredibly expensive, they don't exist. There's no ecosystem, so you have way more complexity, cost structures. You have to find a way to sell something way more expensive.
Cody Simms (22:01):
We're recording this in the week when we're seeing videos all over the internet of Neo trying to load a dishwasher, being human piloted. We're clearly still a little bit away from that.
Boris Sofman (22:12):
It's funny, I'm the biggest skeptic of these. I kind of know how hard consumer products are and how price sensitive people are where okay, you create up your technology, but it's just why am I going to pay this much money? It's got to be pretty flawless. But you have these machines in construction that are actually beautiful machines. They already are incredible manipulators of the world and they're even better. They're actually really well designed for upfit, being able to turn 'em into autonomy capable in a way that's actually better than cars and trucks. And so we were able to see a really clean path to market. We can bypass a lot of these complexities and if you're able to solve that problem, it's astronomical value because construction is like 13% of global GDP. And we talked a little bit earlier about how you have this astronomical spikes in demand that are geographically concentrated and insatiable.
(22:59):
In the case of data centers right now there's $170 billion of data center construction per year in the US and $250 billion of manufacturing construction. So just those two is unsustainable volume on top of infrastructure and housing and everything else. You had this pretty fascinating split of supply and demand and a value proposition that as we started talking to customers in the space, it goes far beyond labor. You start to see opportunities to compress schedules by working 24 hours a day, you have safety issues that are at the core of the whole industry. Predictability is very hard and it's one of the most difficult things for general contractors to bid on a project and you end up having 5% contingencies oftentimes because the earth work is so unpredictable, even though a lot of projects are won by a third of a percent. And so you have this really interesting flow in an industry that's so constrained and at the same time when you have this deviation in supply and demand, you see prices skyrocket. You see projects not get done, all the pieces align, but we're very thoughtful to try to build these systems in a way that even if we start with excavators in construction, we're designing fundamental systems that can generalize and create this snowball effect and flywheel that jumps to new capabilities, new machines and even new sectors down the road.
Yin Lu (24:13):
Hey everyone, I'm Yin a partner at MCJ here to take a quick minute to tell you about the MCJ Collective membership. Globally startups are rewriting industries to be cleaner, more profitable and more secure. And at MCJ, we recognize that a rapidly changing business landscape requires a workforce that can adapt. MCJ Collective is a vetted member network for tech and industry leaders who are building, working for or advising on solutions that can address the transition of energy and industry MCJ Collective connects members with one another with MCJ's portfolio and our broader network. We do this through a powerful member hub, timely introductions, curated events, and a unique talent matchmaking system and opportunities to learn from peers and podcast guests. We started in 2019 and have grown to thousands of members globally. If you want to learn more, head over to MCJ.vc and click the membership tab at the top. Thanks and enjoy the rest of the show.
Cody Simms (25:15):
Just to get into your technology, you're not building net new machines currently, at least you're retrofitting existing construction equipment with autonomous capability, is that correct?
Boris Sofman (25:26):
That's right. So we're upfitting existing machines with a suite of sensors and compute that taps into the machine.
Cody Simms (25:32):
Upfitting not retrofitting, you're taking them forward. I see,
Boris Sofman (25:36):
I
Cody Simms (25:36):
Like it,
Boris Sofman (25:38):
I'll take it. But the point is that we're doing it in a way that doesn't even permanently modify the machine. It's a hundred percent reversible and we could do this in less than three hours. That's pretty powerful. So now you have a machine that has a variety of sensors. We have cameras, a lidar GPSs, dual GPSs, IMU LT modules, and it has all of this signal from around the machine. We also see the signals from the machine itself and then we can turn an existing machine into one that can be fully autonomous for the scope of work that we approve, that surfacer is going to increase over time with software updates and you can still operate it manually anytime you want to and that's a pretty powerful value proposition for a customer versus having to purchase a brand new half a million dollar machine.
Cody Simms (26:18):
Excavator, backhoe, dump truck. Is there a specific set of vehicles you're starting with or specific set of use cases you're starting with?
Boris Sofman (26:25):
We're starting with excavators and that's very intentional. So an excavator is on average, the most prevalent machine is probably about a quarter of all machines in the construction industry. It's usually the highest utilized and it's also one of the hardest to learn because you have six degrees of freedom and this isn't like driving where you pick it up, but if you want to get a trucker CDO license, you can get pretty competent in three, four months at least to be able to work. These machines are six degrees of freedom and they're really hard and a little bit unnatural for people and they can oftentimes take many, many years to become really competent at these machines. And so you have a lot of business pull if you can solve this problem. And so we're starting with excavators, but very much thinking about that roadmap where you have, like you said, wheel loaders, bulldozers, backhoes, motor graders, compactors, scrapers, then you have this long tail of machines and to solve that problem, you have to use these approaches that we're using today. If you did this five years ago, you would've probably anchored on the more traditional approaches and you can slice off small segments of problems, but you had no leverage and you end up plateauing. And so I think that's the really interesting unlock that we really saw in the parallels to Waymo all over the place on how do you think about safety, how do you evaluate, how do you do vehicle programs and hardware outfits and simulation stacks. So we had the gas and ran with it.
Cody Simms (27:40):
It strikes me that with an excavator and really with construction generally you're dealing with a physical environment that by the nature of construction is changing around it. We have a portfolio company that is doing robotic solar panel installation that is a repeatable task which a robot can learn and repeat and they're not necessarily digging and creating different geology around the machine itself, but in your case, you're actually shifting the topography in which the machine operates. How do you navigate that from a training perspective, from a learning perspective that feels different than the problem you were working on at Waymo? I suppose
Boris Sofman (28:20):
You're right. That's one of the places that's different and it does introduce challenges. So Waymo, you're just a participant on the city streets. You don't physically manipulate them or modify them, and you can take advantage of that fact here you're terraforming, demolishing, putting holes in the ground, digging and loading that has complexities. For example, in simulation, you want to simulate now you've got to deal with earth physics and all these interesting things. When it comes to the actual work itself, we frame this as just part of the problem itself. We use a large scale kind of end NML model. We're taking input all the sensor data, the information from the machine, the position and part of the input is actually the goal of the design of what you're digging to. And so let's say you're digging a foundation, you can have a 3D model of transitions and these systems, what we're actually learning is how to mold the environment from a starting point to a new point. You are learning these patterns across thousands and thousands of hours of data, tens of thousands of hours and hundreds of thousands of hours of data, and eventually those patterns generalize whatever environment you're in, you start to be able to apply that learning on how do you manipulate it to get to a goal state. And so the fact that you're manipulating the earth just becomes inherent in the most core aspect of what the autonomy system is actually doing. As you get better and better, you end up generalizing to broader use cases in this way.
Cody Simms (29:43):
Why full autonomy for this? Why not teleoperation as an initial path into the autonomous market?
Boris Sofman (29:51):
It falls in the same bucket is why not level two then level four where you do an autopilot and then a full driverless? Fundamentally, first of all, the value proposition is just astronomically different. You have a very different type of company and product. If you're fully driverless versus you're an assist or you're changing some properties or you're doing arbitrage on labor costs, something like teop, it's a tool, but if you want to actually do this as a stepping stone and as a product is definitively complex, now you got to deal with latency and dropouts and safety of a different kind and at the end of the day you might be able to capture some savings, but you still have a person in the loop. And I personally feel like if you're going to go and do that, this isn't like a three months detour certainly doesn't come for free. This is many, many, many years to build an actual company that goes and starts to build a business and that deviates away from the actual path to autonomy. We felt, especially given what we knew from Waymo that leveraging all that knowledge and this particular problem, we could go straight to a fully driverless system, take advantage of the fact that in construction you don't have to solve a hundred percent of everything. You can create competencies that become still very, very useful.
Cody Simms (30:54):
I suppose all the complexities we talked about about terra formation, on the other hand, you don't have the complexity of a completely wild environment around you where dogs are running in front of the vehicle and a police siren is shooting in front of you on the street.
Boris Sofman (31:07):
Kids are jumping out from behind cars and I can tell you stories about the horrors that millions of miles of San Francisco will throw at you, people jumping on your hood of your car on purpose with enough frequency that we actually had a rate for it. You have everything. Here, you have a much more controlled environment and you also have the ability to take advantage of the speeds and so forth. But when you ship a Robo Taxi in San Francisco, you have no product flexibility. You got to solve all of San Francisco and everything it throws at you because your long tail is the extremely rare safety challenges that'll exist here. We actually have a very, very distinct way to go and sequence capabilities and there's machines that for 80 hours a week for a year will go and do mass excavation for a factory and it's okay that they don't yet carry a pipe and so forth or do demotions, so there's these sequences that you can take advantage of and build it up. We felt like we had the right team and the right understanding to go straight to the final solution. Of course you can take advantage of remote assistance and these things, but it becomes more of a tool like Waymo uses it as a tool versus the product itself, which is a completely different type of company.
Cody Simms (32:10):
How much of the onboard controls are on local computers on the machines and how much are you needing data connectivity and particularly if you're doing construction in a remote environment like a data center in the middle of a desert in Arizona for example. I guess there's Starlink now you could connect into, but you may not have 5G connectivity from a data pipe perspective. How does that work?
Boris Sofman (32:32):
This is a pretty important architectural decision and we certainly I think learned a lot about this in the challenges that Waymo carried over a lot of the philosophies. So we're choosing to keep all of the time critical real-time functionality local on the machine. The computer that gets placed inside of a excavator for example, is responsible for all safety, all real-time functionality behavior, and on any local timeframe can go and operate. And it's exactly because of the reasons you said. So we may have latencies, we may have connectivity issues, and the way you qualify a system, if you can't guarantee these elements, it completely changes the challenge that you have in front of you. You still use interfaces whether they're local or through the cloud to do less time sensitive things. So for example, coordinating a fleet of machines on a site or developer or a general contractor checking on the state of the machines or giving an update to a goal or us checking the functionality of a system, those are fine just like each Waymo's completely independent, but there's a central server that coordinates the hundreds of machines in a city.
(33:35):
That's a nice abstraction where as you scale, you don't stress extra complexity surface area. That becomes safety critical and actually makes a lot of sense because going back to our example from Anki where we were leveraging the ecosystem of smartphones today, we're able to leverage this ecosystem of automotive industry where more and more cars are starting to get level two, level three functionality. I think every single OEM by probably 27, latest 28 is going to have to have a autopilot style solution. Doesn't mean internally it could be partnering with a lot of the companies that are doing level two systems. What that means is that ruggedized cameras that are meant to last 10 years are plummeting in cost and are already incredibly robust and reliable. Fantastic. We don't need to reinvent that. Same thing with compute where we can use chips that Nvidia is making and others for these sort of applications very, very valuable. And so there's these trends that naturally just accelerate you and make it natural to just embrace that functionality in a way that has all sorts of advantages.
Cody Simms (34:37):
Talk about some of your early customers. You've mentioned data centers a few times. Do you think that is your initial beachhead market that customers are going to be buying your systems to go out and tackle?
Boris Sofman (34:49):
We partner very closely with general contractors and subcontractors that in the end are very representative of who would be our customers. So these are the companies that already buy and manage fleets of machines, general contractors that self perform and do earth work or demotion, whatever the case might be for heavy civil or for industrial applications. And so we have a number of partners, they're wonderful. So a few examples are Sun, which is a general contractor based in Phoenix that has incredible innovative spirit. We have a number in Texas, Zachary Construction, Austin Bridge Road, and so they're all very excited about how this can unlock far deeper scalability and productivity, better safety and just reinvent the operations of a general contractor in the future. So we're partnering very closely with them and the types of work that excite us are the ones that have a few properties and so we think of this as like a beachhead where you get to market and you can start the scale of business while you're expanding your surface area and your scope and that has to be narrow enough to where we're not building for years and years and years, but broad enough to where it's very valuable and very applicable.
(35:52):
And so the sort of sectors that we really like as potential starting points are heavy industrial sectors, building factories, data centers, warehouses, municipal plants, giant projects where you're moving hundreds of thousands of cubic yards of earth and so some of these machines are operating 12 to 24 hours a day for the better part of a year. That's a fantastic property for machine learning type systems where you want to get something that's consistent and learnable heavy civil has the same property as well, a lot of consistency and reuse and so what we might put off to later for example, is much more one-off type work like residential projects and things like that where you also have interesting challenges around interfaces where it's a lot easier to define what's the foundation you want for a new factory or a warehouse than it is to define some incredibly complex project that needs to be done for one particular type of house. And so we think about it that way where everything's going to expand and everything's in scope in the long term.
Cody Simms (36:48):
Even a housing development has complex landscaping because you have multiple backyards and stuff
Boris Sofman (36:53):
And maybe if it's like a planned community where you're doing a swath of land for 200 houses, that's different, but you go into urban construction in downtown Manhattan for example. That's a very, very different type of work. We can be very clever about how we enter the market and we look for these huge volume high utilization projects where not only is that very conducive to really deploying something valuable, but it also is very valuable for the customer because it's so high volume and creates bottlenecks on the whole project.
Cody Simms (37:20):
Where are you today in terms of machines actually on the ground doing work?
Boris Sofman (37:26):
We've been doing autonomy testing since late last year. We're progressing quickly. We're simultaneously collecting a lot of data, improving our autonomy, maturing our capabilities, also maturing the safety system that will allow us to go driverless and then next year we're planning to go driverless in our first deployments and that'll be the beginning of our commercial scale and context. It's really, really different than what Waymo experienced. Waymo is a 15-year-old company at this point. Most people don't realize or 16-year-old company really started commercializing very, very initially in 2019 and Phoenix and then really aggressively a few years later coming out of San Francisco. The fact that we can go from zero to first deployments in a few years, that's a sign of how powerful some of these evolutions and learnings in the ecosystem are.
Cody Simms (38:11):
Those driver assisted Waymo's were on the street for a very long time before they started going full level.
Boris Sofman (38:17):
That's right. To be fair, that is a astronomically more adversarial path to market to go zero to one because you do not have a path to release something other than level two release something that's really, really good but not perfect and you have to really solve all of San Francisco if you want to ship a product. I actually don't think that's a startup game. I think you actually have to have the conviction and muscle of an alphabet or an Amazon to be able to push this through because you're just too vulnerable to the physics of external markets and fluctuations and there's not too many examples where startups have been able to spend billions of dollars before getting to revenue and that actually works out well. We're able to take advantage of a lot of the properties of this space and start to build the business along the way.
Cody Simms (38:57):
A good transition there in that you just fairly recently announced your series A funding. Maybe share a little bit about how you've capitalized the company so far.
Boris Sofman (39:06):
We've raised two major rounds of funding seed round and series A. Really love the investors that we've had the chance to work with and we were very oversubscribed and so we were fortunate to have the chance to work with a lot of folks that we go to know very well and respect. So it clips venture led our seed round, A, B, C, led our series A. We have other amazing investors that have participated, Valor, Tishman Speyer, Nvidia, Two Sigma Ventures, a number of others, some really great individuals and old friends, old CEO for Waymo, ex-CEO for Waymo, John Krafcik and old friends from the industry. And so we kind of had the fortune of being able to not just get amazing lead investors but also try to pull together really great partners across the ecosystem that help us. We've raised over $80 million at this point and are starting to get close to maturing this into a fully autonomous product.
Cody Simms (39:57):
Congratulations. With Waymo, it was kind of Alphabet doing alphabet's thing. You built this as a standalone service, you had to do a lot of work with regulatory approval and all of that, but at the end of the day it's a consumer picking up their phone, hitting a button, car comes and picks 'em up, they get in the car, they go where they're going, and you don't have to interact with the existing system all that much. In your case, you're interacting with a construction culture, you're bringing your tools into the workflow of a pretty traditional industry that's been around for a long time of people who might be thinking, oh, is this coming after my job next? How do you deal with that and how are you learning to deal with that as the CEO?
Boris Sofman (40:46):
You have to really respect these types of industries because the learnings are just magical with these people that have been doing this for 40 years and how they're able to deal with the randomness and chaos of these spaces to solve problems. So you're right that we will had the luxury of fully controlling the experience interaction with customers. At the same time you're driving on public roads, you have all this extra complexity with pedestrians, the existing traffic and challenges that are out there, the regulatory exposure that you now have on every level, local, state, federal. There is an advantage to being on these controlled closed environments that block off some of that surface area where you still have to be fully capable of handling whatever the world throws at you, but there's things that we saw way more that we will never see on a construction site.
Cody Simms (41:29):
You arguably have a simpler technical environment, but I would argue culturally is very different.
Boris Sofman (41:35):
The mistake that AI robotics companies can make is that they come into a space and they say, well, I'm going to reinvent everything. You're going to use my interfaces and now you're going to change your workflow and do this and that. And that's an easy way to be ignored because people don't have time to change your entire workflow. So an example of that is you have to cordon off this whole area. Nobody can walk by and you got to do all these things and here's a bunch of constraints. That's a very quick path to just being kicked off of a project.
Cody Simms (41:59):
I've seen it a lot in agriculture, we're going to make harvesting better or whatever, and it's a tough space because you got to change complete workflows and adapt your way into a culture there.
Boris Sofman (42:09):
One of our philosophies is we try to be incredibly respectful of the way people operate today, and we are expecting to splice into today's workflow with a minimal amount of changes required mainly around the interface of how do you just communicate to what you're trying to do, which is going to be very natural and use a lot of the same types of representations that already exist, but otherwise, this thing is expected to operate around people, around machines, around anything might happen. It'll be safe, it'll be conservative. You don't have to sit there and put up yellow tape and keep people out in order to have this work. That's critical and a lot of the reasons why people have done other approaches because it's hard to solve these perception problems and city problems, but you have to do that in order to absorb in a successful way because we're not solving the whole construction project, we're solving a piece of it, then we're going to grow it and we're going to add new capabilities and add new machines.
(42:57):
And only then after you started to capture more and more surface area, you can go and start to have the conversation of, Hey, we're going to start to think more as an operating system for these machines, but it's worth it because we're going to unlock all this exponential value and flexibility and control and visibility and predictability and then it makes sense. But on day one, when you're splicing it as an excavator, as part of a project that has 50 machines on it, you just can't be arrogant enough to think that you can change everybody's workflow. And I think that's been very core to us and that's why we spent a huge amount of time with our partners to really think through this in a way that feels acceptable to them.
Cody Simms (43:32):
I suppose the promise is you can have 30 excavators operating at a time and they can operate 24 7 and you can do construction at three in the morning and all of that, but it's going to take change on behalf of the construction site operators and everybody to start to adopt that style of work that manages these autonomous or semi-autonomous construction sites. I would think
Boris Sofman (43:55):
You're right, and even though the layers of potential value go so deep and you can completely change the entire P&L of a general contractor, which is very aligned with the interest of the country of just increasing productivity and utilization of machines in the meantime, there's such a shortage of skilled labor that just being able to plug into today's workflow and solve these problems is already incredibly valuable, and that is a great place to be because this company's going to have a huge amount of chapters and even though we're trying to build this engine that can go and cascade in the meantime, we want to have an incredible entry point that doesn't require a huge amount of thought or manipulation of your entire operations to get it.
Cody Simms (44:31):
That's the painkiller versus the vitamin argument. The vitamin is, oh, you can imagine these 24 by seven work sites and all that. The painkiller is, I can't even hire someone to run the excavator today and you're going to help me solve that problem. That's what I'm hearing from you,
Boris Sofman (44:44):
And I'm having all this safety costs, like I'm damaging machines left and right and all this trucking companies who go and take away dirt. Oftentimes they don't trust the estimates, the volumes that they provide, so they have people sitting there manually counting truck loads and bringing in people to survey volumes of earth that removed, and so there's all these things that almost come for free that address pain, but the fundamentals of just labor challenges already are a very natural foot in the door, but when you fast forward three, five years, it'll be unrecognizable because then when you actually start to cover more surface area, the exponential value, when you combine this as a system, it becomes like a digitized manufacturing process in the real world and a GC, instead of micromanaging your wheel loader in your excavator, you can set a goal and these systems go and build it, and then you can scale that to multiple sites where across the 15 projects in the metropolitan area, the system can tell you optimizing for your constraints. You can send a wheel loader on Thursday from that site to that site. All that is exciting and we want that on the roadmap, but step one is do really great work that's valuable and build trust by not creating a bunch of headaches along the way.
Cody Simms (45:47):
Do you think the form factors of the machines eventually change in the driverless world, you have Waymo's, which look like cars with fancy stuff on the top and the sides, and then you have Zoox is which are like this totally different form factor that Amazon is now piloting Right now you're retrofitting human driven vehicles. Do you see yourselves getting into building full construction equipment that is optimized for autonomy from the start?
Boris Sofman (46:09):
It's a huge advantage not to reinvent the machine. Zoox spent so much resources on the hardware program that actually took away from the autonomy investment that ended up being much more complicated. Waymo went a little bit down that path, but corrected itself. It's actually really, really hard to manufacture vehicles. It's actually a huge advantage to partner with the Jaguars and the [inaudible]. For us, that's same. It's a huge advantage in these machines or marbles, Caterpillar machines that we use. They're incredibly well designed. What starts to happen when you fast forward over a long period is you realize that most machines were designed to optimize the constraints of what one person can physically do and learn in terms of operation. That's why you have so much fragmentation and sizes and everything else and the very, very long term. I think there's an opportunity to actually rethink what this looks like and how do you embrace the fact that you're no longer constrained by units of human labor or the volume of experience that a person can pick up and a specialization that they need to build.
Cody Simms (47:02):
Most workers are a specialized vehicle driver, like I drive an excavator or I drive a steamroll. Is that correct?
Boris Sofman (47:09):
For the hardest machines, yes. There's some machines that are relatively easy driving a dump truck's not too hard a compactor. Scrapers are easier. Machines like excavators are very unique. Bulldozers are hard to drive. Motor graders are surprisingly hard to drive, and so these are some of the hardest jobs to fill, and so you'll see specialization where you have the 20 year veteran of excavators when you have the hardest jobs you want to get them, and then maybe they can go and pinch it in other things, but they really do specialize. And so that is on its own already an interesting element of this space. And so when you think of what we're making, we think of the product as the Bedrock operator, and the metaphor of this implies an operator that gets better over time, but also warns them more and more competencies. It spans across machines. Today the industry is optimizing the specialization in humans and what they can and can't do, and the fact that you are trying to push the limits of what large scale numbers of people can reasonably pick up and actually perform work in
Cody Simms (48:02):
Boris. Anything else we should have covered today?
Boris Sofman (48:04):
In general, just a really fun time where the next big wave is going to be the physical manifestations of ai, and I think for all the incredible things that have happened digitally, it's not going to be all at once. I think it's going to be waves of these progressions over the next few decades, but I'm incredibly excited. This is the merger of the digital world and the physical productivity that is in front of us, but we got to solve some of these interesting site quests like energy and everything else in front of us.
Cody Simms (48:30):
Thanks for taking the time. It's amazing to hear what you're working on. It's awesome to see the career arc you've had since you and I met over a decade ago when you were building something kind of in the same realm, but very, very different form factor and market.
Boris Sofman (48:43):
Yeah. These are 80,000 pound machines now.
Cody Simms (48:47):
I can't wait to see the company continue to grow and expand and see where the future goes.
Boris Sofman (48:52):
Cody, thanks so much and a pleasure to spend the time together and thank you for having me on.
Cody Simms (48:56):
Inevitable is an MCJ podcast at MCJ. We back founders driving the transition of energy and industry and solving the inevitable impacts of climate change. If you'd like to learn more about mcj, visit us@ MCJ.vc and subscribe to our weekly newsletter at newsletter.mcj.vc. Thanks and see you next episode.
