CoreWeave Fully Connected 2026 | Adrien Gaidon, Walden Robotics & John Mancuso, CoreWeave

Adrien Gaidon, Walden Robotics & John Mancuso, CoreWeave

This interview explores physical artificial intelligence and general-purpose robotics applied to manufacturing at the CoreWeave Fully Connected 2026 conference. It examines simulation-to-deployment workflows, continuous learning in deployment and human-in-the-loop co-design, and how advances in compute and infrastructure enable robots to learn on the job. Adrien Gaidon of Walden Robotics is co-founder and chief scientific officer. Gaidon is a specialist in machine learning and physical artificial intelligence and discusses Walden Robotics' deployment-first approach to general-purpose robots for manufacturing. They describe continuous learning in deployment as a means to close the automation gap while enabling higher-mix personalization and maintaining safety and quality. John Mancuso of CoreWeave is vice president of field engineering. Mancuso outlines CoreWeave's infrastructure and forward-deployed engineering support and emphasizes the company's role in providing simulation, specialized compute and deployment expertise to close the loop between virtual model training and real-world operation. They highlight how improved compute and edge capabilities accelerate deployment and on-the-job learning for robotic automation. The conversation, hosted by John Furrier and Dave Vellante on theCUBE Research, provides practical guidance for manufacturers and technology partners on co-design with production teams, robust evaluation and trust-building in live operations. These topics are relevant to organizations pursuing robotics deployment, artificial intelligence integration, machine learning driven controls, simulation-based development and edge-to-cloud compute strategies.

Share this session

Clips
More from CoreWeave Fully Connected 2026

Adrien Gaidon

Co-Founder & CSO

Walden Robotics

John Mancuso

VP Customer Experience

CoreWeave

play_circle_outline General-purpose robots capable of multiple tasks and rapid retraining.

play_circle_outline Automation gap: tasks like machine tending, parts kitting, subassembly.

play_circle_outline Scaling Physical AI: Funding, Co-Designed Hardware, Perception, Simulation, Compute Partners, and Trusted Deployment

Info
Transcript

Adrien Gaidon, Walden Robotics & John Mancuso, CoreWeave

Adrien Gaidon

Co-Founder & CSO Walden Robotics

John Mancuso

VP Customer Experience CoreWeave

In this interview from Fully Connected 2026, Adrien Gaidon, co-founder and chief strategy officer at Walden Robotics, joins John Mancuso, vice president of field engineering at CoreWeave, to talk with theCUBE's John Furrier and Dave Vellante about the rise of general-purpose robots in manufacturing. Gaidon explains how Walden's robots move beyond single-task automation, learning multiple jobs on the fly and compounding usefulness after deployment. He details the company's "deployment first" philosophy, pairing robots with workers on tasks like machine tending... Read more

explore Keep Exploring

What is the biggest difference between industrial automation and physical AI today? add

Are you focusing on instrumentation or on-line (production-line) automation, and what kinds of manufacturing tasks are you automating — can you give examples? add

What was the system doing from day one, what has it learned since, and what can it incrementally do? add

bolt Powered by CUBE AI

Adrien Gaidon, Walden Robotics & John Mancuso, CoreWeave

search

(INTRO)

John Furrier

>> Welcome back, everyone, to theCUBE's live coverage here in San Francisco for CoreWeave's first annual conference. In fact, the industry's first annual conference around the new layer of infrastructure around AI clouds, formerly Neo Clouds. Got two great guests here. We're talking about physical AI and again, how this is all coming together as intelligence gets injected into the real world. We got Adrien Gaidon, co-founder and CSO of Walden Robotics. Thanks for coming on.

>> Thanks for having me.

John Furrier

>> Looking good. And John Mancuso, VP of Field Engineering, CoreWeave. Are you forward deployed field engineer?

>> We're—

John Furrier

>> okay, we're deployed.

>> Specialist field engineering, all those.

John Furrier

>> Thanks for coming on, guys. We're getting down to our 2 days, but you're the second to last interview here on theCUBE. But the story is simple. Intelligence is being manufactured. These AI factories are an entire system producing intelligence. Okay, great. That defines AI factory. But when you inject it into the real world, That's physical.

Dave Vellante

>> Yep.

John Furrier

>> Robotics, construction. We saw Caterpillar on stage, our lives, wearables, F1. So you're starting to see a lot more of physical AI. So I want to get your take as robotics is the use case and the other robotics is, more manufacturing and other physical big things. What's the physical AI state of the art now? Compute's getting smaller, faster. I mean, robotics is essentially an AI factory in some sense.

>> Right, right. I think the biggest difference from industrial automation to physical AI now is really general-purpose robots. It's this ability to do two things. One, do multiple things at a time. These robots have capabilities that enable them to do multiple different tasks that couldn't be done before and switch between those tasks. So part of the value proposition is not just you do one thing 24/7, 365, 'cause that's been already automated. You do multiple different things and humans can program those robots on the fly and really maximize the usefulness of those robots across new tasks. But the second half is how quickly can it learn those new tasks? And that's where once you deploy a robot that can do multiple different things, how does it learn the new thing faster and faster? And how do you keep this value proposition growing over time?

Dave Vellante

>> And that's what Walden does, general purpose robot.

>> That's what we do. We do general purpose robots for manufacturing environments and we're really deployment first as a company. Where the goal is to deploy robots really fast that demonstrate usefulness and ROI on day one, but keep compounding and become more and more useful. Typical experience we have is we deploy our robots, they do something useful, and then people go around the factory and say, oh wait, could it do that? Could it do this? Could it do that? And the answer is to say yes as fast as we can.

John Furrier

>> Yeah, and the old answer was, we'll get back to you in a year.

>> Yes, yes.

John Furrier

>> Still, there's a multi-purpose kind of thing going on, very Matrix-like. Teach it to do something, upload some software.

>> kung fu

John Furrier

>> now.Yeah,

>> exactly.

Dave Vellante

>> Yeah.I remember when robots couldn't even climb stairs. Now they're playing tennis.

John Furrier

>> Yeah.That's the Chinese Unitree. But this is real. People see the robots. People love robots in space. So that's one thing we learned.

John Furrier

>> Yeah.The robots equals ratings. And so, can you get us some B-roll?

>> Oh yeah, we can get you tons of B-roll. Our robots are very cute, so people love

John Furrier

>> them.What are the use cases? Because when you say general purpose, I'm envisioning, are you lifting heavy things? Are you doing instrumentation? Are you on the line? Take us through some of the examples.

>> Right, so there's this very large automation gap, things that were not automated before. And what we look at here is tasks that are typically like machine tending or parts kitting or like tool settings or subassembly tasks. So those are tasks in manufacturing environments that people have tried and failed or not even tried to automate before. And amongst those tasks, there's tasks that require the human craft, require real skills. There's half a million open positions for machinists in the US because it's a hard job and it requires skills. But part of that job does not require the craft or the human creativity. So what we're doing with our robots is we're working alongside people to help them as an assistant and do those tasks that people don't want to do, but enables them to really become craftsmen and improve the production processes and all that.

John Furrier

>> Talk about the nuances. As AI gets better, so does the robotics.

>> Yes.

John Furrier

>> So you look at the agent world, deterministic, probabilistic, okay, and they're going to deal with all that stuff. But what we're hearing on the physical side is, let's take a manufacturing plant. It's deterministic by its physical plant, but non-deterministic by the chaos that ensues inside. Something falls on the floor, something doesn't get shipped to the right spot. So there's a lot of so-called deterministic things that should be deterministic that don't become deterministic and you have to adjust. This is where we're seeing a lot of AI advancements. Talk about that piece and why the current AI is starting to solve that problem.

>> Well, John, I was expecting you would say like a Buddhist Zen, the only constant is change. This is what you realize when you deploy in manufacturing environments because why is there an automation gap? the interesting question is not what's already automated, it's what is not and why. And variance is a big part. And you would think what Lean Six Sigma is all about is reducing variance, but actually the next step of mastery for companies like Toyota, Boeing, Samsung, which are partners, which are the best manufacturing companies in the world, is how they leverage their people to improve the processes. And so if you improve things, you're changing things. And if you're changing things and you have machines that cannot change and continuously learn in deployment, then you become a burden. So what we do is we do machines that continuously learn from deployment and adapt so customers can change their mind and improve the process and give them the superpower to change even more and faster.

John Furrier

>> So because you have the plant, because you have the

Dave Vellante

>> video—sorry, go ahead. I'm watching this video of the Toyota factory, which is incredible.

John Furrier

>> Yeah, but this variance thing is interesting because that's actually opposite of what other AIs do, because you actually are trying to get to reduce the defects. Yes. Because you're in kind of a bounded structure.

>> Yes.

John Furrier

>> It's not like you're a Waymo going around San Francisco.

>> It used to be to reduce defects, you have to be like Henry Ford. you can have any color you want as long as it's black. But if you have frontier physical AI like we have, what you can do is you can increase mix, you can increase personalization, you can increase variance while not sacrificing safety first, not sacrificing quality, and not sacrificing delivery. And this acronym in Lean Six Sigma, SQDC, safety, quality, delivery, cost, in that order. How do you supercharge everything at the same time?

Dave Vellante

>> But the trick is the robots have to show up and be immediately useful.

>> That's right.

Dave Vellante

>> And to get value and then learn from there. And then you're compounding that value. Yeah. So how did you guys, work together on that. Maybe, John, you could chime in, talk about the journey that you guys took together.

>> Yeah, well, first of all, the journey is just beginning, but the way that we work with our customers at CoreWeave, the reason, actually at the beginning you talked about forward deployed engineering, and yes, that's part of what we do, but what customers really want are specialists in their field, right? So I know that for Walden, what would be important is that I could bring physical AI kind of specialists into the conversation literally mechanical engineers that understand the physical world and can learn to operate in that world. You know, robotics experts that can again talk about

>> exactly—Richard and

Dave Vellante

>> etc.

>> Exactly.

John Furrier

>> Yeah, yeah, yeah.You got Vera Rubin to boot. You got a lot of— yes, first mover on all the best stuff. Yeah, the infrastructure specialist.

>> It's basically the technology plus the

John Furrier

>> people.And the numbers, by the way, joking aside, being out first is great. But the benefits are immediate. Optimization, token usage, it moves the ball down the field very quickly.

John Furrier

>> Yeah.Yeah. I didn't mean to

>> interrupt.No, it's okay. It's all about this partnership.

>> There's many different things. It's like John was saying, this is an early partnership. So they're an investor in us. They've been an early believer. We're starting our journey on physical AI together. We've been customers of Weights & Biases. I have this claim to fame that I was the first ever customer of Weights & Biases before OpenAI, which is a very— I'm very proud of this. And so Lukas, Shawn, Chris, all these people, we're still working with them very closely because it's very empirical in nature. So the compute is a very big thing in how they scale and partners for scaling, et cetera, but also the services around and the expertise that they have around physical AI that they're building, that we're building together. No, it's a very exciting time.

Dave Vellante

>> We've all seen these industrial robots. They're behind a cage. Right. You can't go near them. You have to obviously design for something that can move around the factory and learn.

>> Yes.

Dave Vellante

>> So how did you get started in that? Take us through how you created that.

>> Yes. So we created Walden Robotics in January, spinning off from TRI, Toyota Research Institute. So Russ Tedrake, our CEO, is an MIT professor of 20 years, led robotics. I led machine learning there. We did a big project called Large Behavior Model, which was just the AI. We put it on the Boston Dynamics Atlas robot. It worked great, all this kind of stuff. But we wanted to make it, to bring it to market. So we created the company in January and we built our first 2 robots in the first 2 months, put them on the production floor on a non-production weekend, then passed all these gates, all these KPIs, etc., etc. And we've been operating, doing real work alongside people in production since May and working full shifts. And so the way to do that is very complicated, but it goes— it's basically co-design, this concept of co-design, of designing the hardware, the software, the AI, the deployment muscle together with your customers, including the business case too. And then if you do this and you listen to your customers and you want to build a product, the technology is so powerful that you can really move very, very quickly. And this acceleration is what we're experiencing now since we came out of stealth, which is very intense, but it's a thrilling journey.

Dave Vellante

>> So you came out of stealth, I think you announced $300 million.

>> That's right.

Dave Vellante

>> Investment. And what was, and take the Toyota example, what was the system doing from day one?

>> Yeah.

Dave Vellante

>> And what has it learned since and what can it incrementally do?

>> So the most important thing is it was doing real production work, right? And we go with customers, we look at their business case, we look at this automation gap that I talked about, we look at what is the most valuable to them, what is the easiest for them also to deploy.

Dave Vellante

>> Here we go.

John Furrier

>> Robots here.

>> And robots doing useful work. And so we've been evaluating a bunch of different tasks with them, selecting a subset, prove that we can deploy, then deploy, then go to the next one. So it's this continuous deployment, continuous iteration, and continuous improvement. But the important thing is it has to be useful. You have to earn the right to be deployed in production, keep earning that right. And if you want to scale, you have to grow the value to your customer.

John Furrier

>> What's been your co-design strategy with Toyota? Obviously, CoreWeave, we had good visibility into that. They're a great partner and your other partners are echoing the same thing. Tier 1 talent, access to great supply chain expertise. On your side, the customer. Yeah, I like that piece because obviously that's a real validation. What's their involvement? What's their feedback? What's their focus?

>> They're a strategic partner. We've worked with Toyota for like, you know, before that for 10 years. We have a deep respect and trust that we built around the Toyota Production System, around all these things. So first, we know the environment, we know the people, we know the task, we know the KPIs. If you look at physical AI, look at who writes the tests, right? We don't write the tests. Toyota or our customers, Samsung, Boeing, other folks, they have this very long list of acronyms and metrics, et cetera, that we have to say, okay, that's the test, we have to pass this. So co-design means how do you define usefulness? How do you measure? What are the constraints? And that feeds back into everything, the data strategy, that feeds back into the task selection process, it feeds back into the hardware. We have some of the best actuator and hardware people in the world, that worked at Amazon, worked at Boston Dynamics. They have been the largest manufacturer of robots in North America. And so we can touch any part that is a bottleneck. And that's kind of the lesson of the Toyota Production System. It's like, if you can imagine you have a tube, if you're improving stuff which is not a bottleneck, you're not increasing your flow. But if you're looking at the bottleneck and you're increasing that, you're increasing the flow. But that bottleneck shifts over time. So that's this process that we do with customers. And that's why we're deploying first.

John Furrier

>> Yeah, I love what you guys do.

Dave Vellante

>> John, how are you guys participating in this continuous improvement loop? Loop, there it is again.

>> Yeah, loop. Yes, exactly. It's all about the loop. It's really exactly what was just said. In the end, before a model is ever going to do something in the physical world, it's going to have a tremendous amount of simulation that happens in the virtual world. So again, we're providing the infrastructure. And a lot of knowledge across many customers to be able to help Walden and all of our customers understand the full loop that they run, to curate, improve, evaluate that whole thing. So yeah, that's really where we come in.

John Furrier

>> What are the blockers? Because I was having a conversation when I was in Paris this July on our robotics event. It was an AI event, the RAISE Summit. The Chinese were flexing with Unitree all their marching robots. And so there's clearly an advancement in the US and North America and Europe. So we're expecting there to be a big surge next year. A lot of cool things are tailwinds on this. You got FPGAs, you got sensors are getting better. Everything on the hardware side seems, and the software stacks are pluggable, you can plug in. What's the big blocker? That's gonna be broken down in your mind? 'Cause it's just a matter of time in our mind. So, 'cause it's going so fast.

>> Right, right. I think there's multiple different layers. There's one which is underestimated, which is trust. Robotics, it's heavy, right? And so how do you build trust with the customers? How do you get deployed and gradually build that trust and accelerate, right? So the bottleneck is, very much a deployment bottleneck because like you said, the capabilities are there, the technology is progressing. So we're very focused on that and earning the trust of our customers, proving it with data, proving it in the field, et cetera. I think it's the trust, the hardware, the software, the AI. That's why it's a deployment bottleneck because everything flows together.

John Furrier

>> Is that deployment just on the customer side? Is it accuracy? 'Cause robotics, you can't be wrong. It's like a financial institution and their AI. You talk to those guys, they're like, we're in the numbers game. You can't be the zero, 100% success.

John Furrier

>> Yes.Because safety is

>> huge.Yes, safety is huge. I think a lot of times people see things as binary, or false dichotomy. It's like either autonomous or it's not, right? But reality is autonomy is a ratio. A lot of these metrics, they're continuous. They have thresholds which unlock capabilities, right? But they're continuous. So the goal is, again, continuous improvement. And so working on that and closing that loop that we're talking about is the way that we basically unlock these steps, these milestones.

John Furrier

>> Computer vision's been a big game changer.

>> For sure.

John Furrier

>> How has that changed robotics?

>> I did my PhD in computer vision. I worked for 20 years in machine learning and computer vision. I did a lot of that. I'm a big fan. It created the deep learning revolution and then, eventually, LLMs and now VLMs, now VLAs and world models. So there's a long history, you know, science stands on the shoulders of giants. And now we have these strong playbooks for how to build datasets, how to build evals. You were talking about another bottleneck we didn't talk about is how do you evaluate before you deploy so that you don't just say, trust me, it's going to be great.

John Furrier

>> Yeah, yeah. Parachute will open. Yeah, exactly.

Dave Vellante

>> Exactly. I have a quote here from our partner Qualitate who provides some data sometimes. And so it says you're not going to just install a robot on your working factory floor and say, hey, go and learn.

>> Yes.

Dave Vellante

>> Okay, so that means you've gotta do some simulation beforehand.

>> Right.

Dave Vellante

>> How are you doing it? Are you using things like Omniverse? Do you write your own system?

>> Yeah, so we use a variety of different solutions. So NVIDIA is also a close partner. Russ Tedrake has the Drake Simulator at MIT, which is a very powerful physics simulator. I think the idea that we have is again, as a third choice, people say either you deploy the thing in the lab when it's 100%, you ship it. But then when you do that, it goes to 10%. There are studies that show this, and I think a lot of roboticists have scarsbased on this.

Dave Vellante

>> So you've gotta have that field experience.

>> Yeah.

>> Or you just say, no, I just only learn in the field. So the reality is always something else. We learn powerful frontier AI models. They're never gonna be perfect, but they're good enough to basically demonstrate value. And then what you do is you co-design a system with a human in the loop. So if you, I don't know if you've taken a Waymo here, but they have remote assistance. And that remote assistant for Waymo has a very strong connection with how Toyota made looms 120 years ago. They made automated looms and had one person overseeing 40 looms, just like we have Waymo overseeing 70 vehicles or something like that. So we have the same thing, which is our robots are autonomous, but they know how to call for help. They know, they're humble robots, I would say. They know when they need some assistance, and that assistance feeds the most valuable data to improve over time. And then eventually you can scale the human oversight, which you always want a little bit of human oversight because I often say we're roboticists on team human. We make robots for people for increasing the productivity of people.

Dave Vellante

>> So Mechanical Turk, team human, self-effacing, right?

>> We have remote assistants essentially that we have many different names for them, but I think remote assistant is kind of like the term. But the key thing is how AI knows how to call for help.

John Furrier

>> Well, Adrien, we have to get you in our robotics series for our NYC Wired program. Pleasure. Congratulations on the continued success. It's just early innings. we were comparing this to the early AWS days where they had EC2, S3, and queuing. Now you got AI factories, power, and optimization.

>> Yeah.And this is systems gate we're putting on

John Furrier

>> top.Those are the basic building blocks. And the more come, the more will come.

John Furrier

>> Yes.Thanks for coming on. Appreciate

Dave Vellante

>> it.Thanks, you guys.

John Furrier

>> I'm John Furrier with Dave Vellante. More coverage As we wrap up day 2 of theCUBE coverage from SiliconANGLE after this short break.