theCUBE + NYSE Wired: Physical AI & Robotics Leaders | Benji Barash, Roboto AI

Clips
More from theCUBE + NYSE Wired: Physical AI & Robotics Leaders

Speakers

Benji Barash

CEO

Roboto AI

play_circle_outline Advancing Robotics: Importance of Computer Vision Data in Building Reliable Fleet

play_circle_outline Real-time decision-making on robots and offline data analysis

Info
Transcript

Benji Barash, Roboto AI

Benji Barash

CEO Roboto AI

Exploring Robotics Data Solutions with Roboto AI at NYSE Wired AI Media Week

Benji Barash, founder of Roboto AI, joins theCUBE at NYSE Wired Robotics & AI Media Week to discuss navigating the complexities of data in robotics. In this insightful conversation, Barash explains how Roboto AI revolutionizes the management of large-scale robotic data analysis.

Barash, an expert in robotics data management and a former team leader at Amazon Robotics, shares their experiences in creating data platforms that enhance efficiency and reliability in robo... Read more

explore Keep Exploring

What challenges did the team face with deploying robots and why was analyzing log data so difficult in the robotics industry? add

What are some challenges faced when operating robots in real-time and collecting and analyzing data afterwards? add

bolt Powered by CUBE AI

Benji Barash, Roboto AI

search

>> Welcome back everyone. I'm John Furrier, host of theCUBE. We are here at the NYSE Active. Trading floor going on behind that, so you can hear the wave there. We are here for the robotics and AI leaders. We're psyched to have Benji here, he's going to talk about robotics. Benji, great to see you. We saw each other at GTC.

Benji Barash

>> We did, yeah.

>> Thanks for coming on theCUBE.

Benji Barash

>> Thanks so much for having me.

>> We almost had a little segment before we came on about how cool robotics is. Talk about what you're working on first and we'll get into some of the cool things.

Benji Barash

>> Yeah, sure. So we started Roboto. We were working at Amazon several years ago, and we were generating a lot of log data from all the robots that we've deployed at Amazon. And we came across this problem, that really all of that log data needs to be analyzed. And building those tools to analyze a lot of robotic log data is very challenging. And so, ultimately what we're doing is building a data platform to make it easy to search and analyze large volumes of robotic data.

>> And what was the problem that you guys saw out of the gate that was key for the robots? Because if you look at Roboto AI, first of all, I love the URL roboto.ai. What was the problem statement that you guys were solving? What was the catalyst for it? Just problem solving, reinforce data back to make it more efficient? What was some of the things that made this go?

Benji Barash

>> Yeah, it's a good question. We really found that in the real world, when you deploy a lot of robots, they just end up having a lot of edge cases. So things don't ever work quite as well as you're expecting. And you always have to go back to the log data to find out, why did this problem happen? Why did our drone crash? Why did our autonomous vehicle hit an obstacle when we didn't expect it to? Why did the navigation algorithm not work? Why did the controls algorithm not work? And so, you always have to come down to the log data, but the problem in robotics is that the log data is really difficult to work with because they're so big, the log site. In 30 minutes, a drone can end up producing a log that's 100 gigabytes just in 30 minutes. And all that data is point clouds, images. Think of all the sensors on these robots these days and all that data that's being generated. If you need to figure out what the hell went wrong, you ought to then go through all of that data afterwards and analyze it.

>> Yeah, it's interesting. Big data, when we started theCUBE 15 years ago was about Hadoop and storing Hadoop clusters, obviously when the data lakes came around. But all the data was, I won't say pedestrian data, because it's corporate data and other stuff, but computer vision is a data machine, it just sucks everything in. You mentioned it. So a lot of the robots are using vision, computer vision.

Benji Barash

>> Absolutely. Yeah, yeah.

>> And that's a major data. That's why the storage companies are continuing to kick ass, and in vector embeds, it's more data with AI. Jensen said that at GPC, so okay, storage is changing. So computer vision is a key part of this. How much of that does that factor into some of the decisions around what tools you build, how you look at data management?

Benji Barash

>> Yeah, massively. You're exactly right. Yeah, you're exactly right. I think the lion's share of this data that I'm talking about is often visual data. That might be from a camera, so it might be normal looking imagery, because often a lot of robots have multiple cameras on. So it could be from multiple camera feeds, but it could also be from LIDARs or radars that are also visually seeing the world, but just capturing a different type of data, ultimately. And you're exactly right, some of those cameras and some of those types of LIDARs might collect 10, 20, 30, 50 images a second, and if you have multiple cameras or multiple LIDARs, it just kind of stops spiraling. So as you manage that data, you have to be really smart about the kind of decisions you're making, how much processing you're doing, because otherwise you're just going to end up with runaway costs as a robotics company.

>> Benji, talk about the complexity around the data platform. Because one of the things that we've done over the past couple of years with theCUBE, is a lot of end users have come to us with problems, but the share they want to do media around there is Uber. Uber has come to us and said, "Hey, we want to get on theCUBE and explain what we're doing." Because they want to recruit engineers. And so, we'll ask any questions. So we ask them to look at their stuff. Of course, they've got cars, they've got riders, they've got people, places, and things. And so, they have a lot of diverse kind of systems that have to operate in real-time, okay? So robots or you can just abstract that and say, well, a robot is like a car. That's the main thing. But there's a lot of other things going on to power the robot, so harmonizing data. So the data platform challenge is not trivial. So could you scope the complexity involved in... Because I think this is a proxy for all enterprises. The physical AI will have this problem.

Benji Barash

>> That's right.

>> Okay. What is scope, that complexity? Because real-time, you can't be near real-time in flying around drones or having mechanical arms doing real precision work or even craft-type assignments or versatility, for instance. What is the complexity equation? Or can you share what goes on, on the back end to make the front end, the robot work well?

Benji Barash

>> Yeah, totally. So I think as you touched on, when you have your robot operating, you need to make a lot of decisions on the robot in real-time, and often in strict real-time, because maybe if you don't make a decision quickly, someone could die. If you can't swerve out the way of a pedestrian fast enough or you can't find an alternate flight path quickly, you might crash into something. So on the drone and on the vehicle and on the robots, everything has to be very, very fast. But when all that data is getting collected and then ultimately offloaded, perhaps at the end of a flight, at the end of a drive, you can spend more time looking at it. So that's going into big batch processes offline. That's really where we come in. And really, that data afterwards, you start with these big binary log files, you need to split it all up. So you need to be like, that's the image data from the cameras, that's the point cloud data from the LIDARs. This is the time series data from the inertial measurement units and whatnot. And then afterwards, you might have to just go through a lot of analytics kind of processes, because maybe you need to find the patterns. You need to understand, "We had this issue on one of our vehicles, is it going to happen again? Or is it affecting our whole fleet and it's a bit of a ticking time bomb right now?" So it's a huge problem. Physical AI is just going to be a data beast.

>> Yeah, I think that what you're working on is so cool. And I want to just go, if you don't mind, go down the rabbit hole a little bit on this, because we've seen Splunk and companies like Splunk do log analysis to get insights. "Hey, we can do better on some configuration." I get that. That's some nice prescriptive and insights. But you're really getting into what I would call the core dump. Operating systems, when shit fails, you've got to go into the weeds.

Benji Barash

>> Absolutely.

>> And so, because it's operational, this is feedback directly into the system. This is not like, "Hey, how do we get better leverage out of the existing infrastructure?" Which is not bad either, all good, but you're in more of a mission-critical role. What are some of the things that you guys do? Can you share the use cases and some examples, some stories where, is it time to value? Is it more of problem recognition? I can't even visualize what the KPIs would be other than get the system fall tolerant.

Benji Barash

>> Yeah, sure. Yeah. Our ultimate value proposition to these robotics companies that we work with is that we help them accelerate their time to market. If they're still in R&D, get their robots out into production environments. But once they're in production, we help them stay there. Because as they start scaling up their fleet, as you can imagine, the number of logs just also grows exponentially. And then the amount of data they're collecting grows exponentially. So they run into these problems where they have issues in production. And to save face with their customers, and also to have a successful product, they've got to fix any problems as soon as possible.

>> And have an explanation.

Benji Barash

>> And have an explanation for it. And these are complex beasts, as you said. For robotics engineers right now, to sometimes debug these issues, they have to have such deep multidisciplinary expertise across so many different subsystems on these vehicles, that the debug situation is complicated, fixing is complicated, and then building new features and new algorithms and new functionality is also really complicated. Because as you touched on before, the way physical AI works is you take a subset of all the data you've actually collected, and then you train new models, new algorithms with that as well. So we really help companies figure out what's working and what's not working on their robots, and then use all that data to curate and then build new functionality as well.

>> Talk about the market. We're riffing here on this, but I'm kind of walking away with an insight that says, "You're in an expansionary build out phase." So mission-critical is to get the operational safety first model going. That's product market fit, home run. You've got to hit that minimum table stake. But there's also a trend in a lot of these conversations, Benji, where it's like there's a lot of value in that data. So you're seeing NVIDIA actually said in their omniverse that they actually use their synthetic data models. In this case you have real data. So have you guys look at that? Is that on the plan? I'm sure it's on the back burner or maybe it's a couple of dots to connect away, but I'd imagine that you have a lot of data coming in that could be leveraged either for digital twins, other things. Is that on your radar at all? Is that part of how people are thinking? Is it more triage, critical path mode thinking?

Benji Barash

>> You're definitely right. Actually, what I was doing when I was at Amazon Robotics was almost entirely simulation. So I have a big sweet spot, I guess with simulation or soft , I should say. But the nuance with simulation is that you end up producing even more data, right? Because you can suddenly simulate many possible realities. And even for example, at Amazon, what we were doing was simulating millions of drone flights per month to make sure our drones were working in a kind of digital twin sense. If we flew this flight path, would it be safe? If we deliver in this kind of regime, would that be safe? So simulation is critical, and digital twins in this context are critical. And we think in order to build better simulations and better digital twins, you have to be able to use data from the real world to see-

>> So there's a lot of leverage in the data, you're saying?

Benji Barash

>> Yeah.

>> And you guys are factoring that in.

Benji Barash

>> Absolutely. Even managing the simulation data is a big problem, but actually we think we can help manage and leverage the synthetic data to then build better models as well for-

>> Yeah, so you're in the data business for robotics.

Benji Barash

>> Exactly.

>> That's basically your business.

Benji Barash

>> Exactly.

>> And so, right now the critical path is get products to market. Check. Then follow on there is innovations, upgrades, things of that nature, product-

Benji Barash

>> Insights.

>> Insights, that comes in. Yeah. And I am glad I can tie that Splunk example, because they already had a steady state kind of business at scale. Let's talk about Amazon for a minute if you don't mind, because I was at re:MARS. For the folks watching, Amazon and Amazon Web Services jointly put an event on, which I thought was probably one of their best events they had. Intersect was another one that Andy Jassy did that has more of an LA culture with Reinvent, but that's a separate kind of conversation. They always experiment at Amazon, which I love. But this event was probably the most fun event because... And it never happened again, because I think it was probably too hard to pull off because you had Amazon proper, which has all the factories, so you had an industrial IoT vibe. You had aerospace, you had space force, you had satellites. That's just on one side of Amazon and Amazon Web Services. Then you had the entire software. That's where I met some of the first machine learning guys, pre-gentic, like Luis from was coming to Nvidia. Luis says-

Benji Barash

>> Oh, I know him. In Seattle?

>> Seattle. He was a Washington professor and he just-

Benji Barash

>> I was employed with his brother at Amazon as well, who was on the same team as me.

>> Yeah. I mean, so you have pedigree and you have been talking about... It's attracting all the talent. So I'm like, okay, first of all, I love the show. Then I see James Hamilton there. James Hamilton, everyone knows James Hamilton. When he shows up you know it's a good event because... He wasn't speaking, he was a spectator, he was interested, but that event really kind of was ahead of its time because it really brought the confluence of physical AI together, but they didn't even know it. I mean, many stakeholders, there were five keynotes going on at the same time. Pick your area. So it was kind of a really good tech party. That's not reality. So re:MARS had that vibe early. Is that a challenge? You mentioned multidisciplinary. My point is that you started to get into, okay, who do you attract for talent? You're on the data side, so you can see, probably a clear line of sight of that. But you've got hardware, hardware and software in this AI robotics vertical recovering. It is absolutely the number one thing. Where the value is, is people who architect the hardware-software relationship, become super valuable. That's mainstream. So what kind of people do robotics companies hire? I mean, mechanical engineers, aerospace engineers, software developers, data scientists, or just people who can solve big problems? I mean-

Benji Barash

>> Honestly, all of the above. That was my favorite thing I always tell people about working on the drone project at Amazon, was that it was such a multidisciplinary team. I didn't feel like I was on just a team of software engineers, as maybe other companies may force you to be on. But you're exactly right. I was working with aerospace engineers, hardware engineers, and even hardware engineers that worked in consumer electronics or maybe in aerospace, and you had totally different products cycle times. And that's really kind of where the magic happens. If you bring all these super talented engineers together that have different backgrounds, different expertise, you build some cool stuff. And that's exactly like the products you're talking about at Amazon. Whether it's drones, whether it's satellites, the company is on a route at the moment.

>> All right, Roboto AI, give us an update on this company. What's the stats? Share some data, numbers. How big are you guys? What's your focus? Give us a quick update.

Benji Barash

>> We're still pretty early stage. We are only about two and a bit years old. We've got about 10 people. We're working with about a dozen robotics companies very closely. We started working with them right at the beginning of our journey, actually. As soon as we left Amazon and we started the company, we found a bunch of robotics companies that were very motivated to solve this problem in their business and they were very glad to work.

>> Because just time and just-

Benji Barash

>> Yeah, they just didn't want to build it themselves. We think that actually a lot of robotics companies over the last decade kind of failed, didn't really go anywhere because they had to do too much themselves. They had to build an autonomous robot, which is its own beast of a problem, and then figure out how to manage all the data, figure out fleet management, figure out all these other aspects that would be enough for individual companies to solve.

>> I mean, look at Google. How much money do they keep pouring into Waymo? I mean, they've got it working at RoboTaxi.

Benji Barash

>> It's an amazing service though. Yeah. Have you been in one?

>> Not yet. Not yet.

Benji Barash

>> You've got to do it. My wife is in love with it. She said she doesn't want to ever ride in a regular taxi.

>> Well, it's safer for people. No, but this is the point, to get there is a huge task. So the componentization, we had a guest on yesterday, former SpaceX, and he talked about how Falcon X was a system that could do multiple missions. And so, the design around robotics is shifting to more, get a core platform, get off the shelf. So I think on the physical side we see some innovation. That might speed things up.

Benji Barash

>> Yeah, I think so as well. Yeah, I mean there's so many tailwinds now in robotics that didn't exist 10 years ago. We were at GCC, the GPUs are so much more performant now. When I was first using GPUs at Amazon about 10, 12 years ago, they were really bad, the embedded GPUs. They were very unreliable and very low power, and now they're amazing. We had all the AI wave come along, incredible for autonomy capabilities. And also, the hardware has just gotten better. It was really difficult to get good sensors that were also reliable for a kind of autonomous vehicle platform 10, 12 years ago, and now there's loads of them. So all the tailwinds are moving in the right direction and now robotics companies can really focus on building the robot. And actually, it's pretty easy now to build a pretty good robot or two, but it's still pretty difficult to scale up to a fleet of reliable robots. That we think is the biggest challenge now and where we can obviously-

>> And what's the roadblock? Obviously you provide a great service, your one layer, what are other blockers that will be removed over as accelerated innovation comes in? What do you see?

Benji Barash

>> Yeah, so I think the data is for sure a big blocker. And it's a funny thing with the data, because there's almost too much of it, but there's also not enough of it. So there's too much of it to quickly search and analyze, but there's also not enough of it to train better simulations or to train better algorithms that are more robust to the real world. So it's a weird space, robotics, because it's sort of just like, is there too much? Is there not enough? We know we need more data, but we also need more diverse data, more of the right data.

>> Well, Benji, great to have you on theCUBE, good riffing session. Congratulations on the venture, early stage. Funding in the future, we have a lot of investors watching.

Benji Barash

>> Yeah, sure.

>> You looking for funding right now?

Benji Barash

>> Not right now, but we are certainly thinking about it. The space is heating up and we're doing well, so...

>> You guys have capital backing now?

Benji Barash

>> Yeah, we've got venture capital from some investors in the Bay Area and also up in Seattle.

>> All right. So can you name names?

Benji Barash

>> Sure. Unusual Ventures, Fuse. And we were incubated at the Allen Institute for AI in Seattle.

>> Awesome. Cool. All right. Well, great to have you on theCUBE.

Benji Barash

>> Yeah, thanks so much.

>> Good to see you in person.

Benji Barash

>> Pleasure.

>> Welcome to the Trust Network, the NYSE Wired CUBE Network. Appreciate you coming on. Again, theCUBE and the NYSE Wired forming a network of trust among experts sharing their knowledge. Again, robotics is hitting an acceleration point, escape velocity, and you're starting to going to see more action as the systems and the components come together, the hardware. Again, AI infrastructure continues to accelerate and enable this next wave coming in. Of course, theCUBE has got it covered. I'm John Furrier, your host. Thanks for watching.