Physical AI in the Real World: Rhoda AI's Andrew Wooten | The AI Rabbit Hole 2026 | DVC

All right, I have so many questions. I hope we will leave some some some time for the questions from the audience but I have a lot to ask and I'm very happy to introduce Andrew to you because well he is working on something very exciting and we can all have a glimpse at what is happening in robotics and in happening where understand what is happening when AI leaves the screen and entering physical world and I'm we are not talking about humanoids here But autonomous systems are running already on real factories, generating real revenue, bringing real value and bringing real labor and as Nick mentioned we are talking about like not just knowledge work but like everything that labor can do on earth and maybe in outer space. The first questions you are a co-founder of a two billion startup no one okay almost no one heard about until this May can you give us like a short version what is road AI what are you doing and what excites you

yeah absolutely so at ROA we're trying to solve the problem of generalist intelligent robots so to frame that really quickly if you look at the world of robotics today there's there's broadly two categories there your traditional factory robots OTS and these are robots that basically make the same movement over and over again all day every day for a decade. And if you go to any automotive factory or you know electronics factory, you'll find a lot of these types of robots and they're very reliable and they're the reason you know we have the goods that we have at the prices that we do. However, these robots are very constrained. As I mentioned, they only make the same movements over and over again. And because of that, most tasks that require any kind of variability if there's a change of goods or if there's a deformable object can't be handled this with this kind of automation.

And essentially there's this you know second category of robots which is intelligent robots. and here the idea is instead of pre-programming a motion use an AI or data strategy in order to adjust to the real world in real time and execute the motion intelligently. and when we started road, what we found was that the second category you know we had all seen videos of Boston Dynamics humanoids hopping around or you know cool robot demos. What we found is that when we talked to customers there are virtually zero of these robots deployed in industry. virtually zero deployed in warehouses or factories. And we realized that this is because these you know robot demos are actually very unreliable. they can work in a very constrained environment where they're perfectly set up for the task, but they can't work in the real world. I'm I'm sure I'll dig deeper into into what we're doing exactly, but what we found is that it's ultimately a data problem.

In the world of language, there's internet scale text pre-training data. in the world of, you know, video or images, you have internet scale pre-training data available. In the world of robotics, you don't really get a lot of robot data. Robot data might be something like teoperation data or simulated robot data. But there's there's huge problems with both of these data sources. For teleoperation, which is when you know a person basically puppeteers a robot, you just don't get enough scale. It would be as if you know OpenAI tried to train models by asking their engineers to write books. It just it just wouldn't work. You need to have that internet scale. with the simulated data, in theory, you can you know exchange compute for simulated data, but the problem is this persistent sim toreal gap. And any data that's simulated is only as good as the simulator that was used to generate it.

And so this is kind of the the core problem in robotics is that there's just no internet scale data to solve the problem. and this is exactly the problem that we ended up attacking. But I don't want to I don't want to spill into your future questions in case

Yeah. Yeah. Just we can continue because like your approach is like through video generative robotics problem not like just general robot problem. How did it click? How how why did you choose this strategy over like existing ones?

Exactly. So when we when we saw this data scarcity, we realized, okay, to solve robotics, we're going to need an internet scale data set that has an understanding of physics. A lot of robotics labs were, you know, going down the VA, vision language action route, which is basically trying to solve this problem with language data. We felt like language will never have, you know, a rich semantic understanding of physics and dynamics. And so we felt that that was the wrong approach. So a year and a half ago we decided why don't we try to solve the robotics problem with internet scale video data and this is a pretty radical approach at the time but specifically what we said is you know let's say let's say the robot needs to pick up this this beautiful teapot here our strategy was to say how about the robot in real time generates a video of itself picking up this teapot and then in real time converts that video into robot actions and the the benefit of this approach is that now you're not solving how to do the task with robot data.

You're solving it with video data and video data you have virtually unlimited of. So you have you have this incredible benefit of you can now use an internet scale virtually infinite data source to solve the core problem which is how does the robot do the task. now when we first you know went out and pitched this idea people said it was impossible and the reason they said it was impossible they said there's two reasons. One is video models are not accurate enough. the physics, you know, they look good. They're they're visually appealing, but the core underlying physics mechanics, if you're dealing with, you know, precise objects, it'll never be accurate enough. And the second reason was they said, well, you just can't run it fast enough. Even if, you know, if you generate a video, often it takes, you know, half a minute or a minute to generate a video.

If you're controlling a robot in real time, you want to be generating, you know, what we call closed loop, like a couple hundred milliseconds every couple hundred milliseconds over and over and over again because that's what you need to operate in the real world. And so what we found is that we we felt that you know this is this is the only approach that's going to scale. robotics is not an easier prog problem than language. If language required internet scale data, robotics will also. So we felt that this is the only approach that was going to work. And we instead focused on solving those two problems. How do you build a video model that is incredibly physically accurate and also incredibly fast in its inference? And that's exactly what we spent the first year and a half building. and we actually built all of that before coming out of stealth because we wanted something to show the world.

but essentially we now have a model that runs in real time. It takes about you know 200 milliseconds to run inference and send the instructions to the robot end to end. And what this has resulted in is a robot policy that can do tasks that were previously impossible to automate. and I think I think there's a video that we can show and I'd love to share it. So before we play this or actually or while we're playing it or before we play this after all the this task it looks like a simple task to a human. Basically what's happening here is this is an automotive factory. They make 3,000 cars a day. Every car has 10,000 parts, 30 million parts. All the parts that are produced locally might be like brought in on a tray, but the parts that come from overseas, say from China, they're brought in on boxes.

And the boxes arrive at these factories. And what I was shocked to find after speaking to you know a hundred large enterprises be it in manufacturing like automotives or logistics like Amazon no one's invented technology to automatically open a box. Boxes are always open by humans which is kind of ridiculous. but the reason is boxes show up. They're often deformed. This one might have a strap. A different one might have a hood or a lid or something you need to cut open. And then when you open the box everything inside is jumbled together. And then you know there's paper plastic dunage because you can't just put the items inside. They'll get damaged. So you have to pack it in with paper and plastic and suppliers don't want to leave a lot of empty room. They want to package it very tightly and because of that every previous approach of automating this through like rule-based automation has failed.

and this is you know a demo we did at a real car factory in Germany and and now let's go ahead and play the video. So what you see here is that every time the box is in a different state the straps are in a different state there's this deformable material. clear plastic is, you know, traditionally considered impossible to deal with with robotics, but here it's actually able to deal with every single one of these boxes. And it can keep playing while we while we continue our discussion.

Oh, yeah. I I know that like two hours of videos like doing the same task, the same job. which is very different from demos that you mentioned. Most of the demos are still theater. And recently we've been on TED and we've seen how humanoids were pulling back from the TED stage and given a literally trans room like a kid would do. but yeah here we we can see absolutely different results. This is a real factory real deployment hours of perfect video to make everyone asleep. probably we can stop here. You train robots by showing them videos how people do things. Nick, sorry I I I have to bring you back again. my husband watches a lot of YouTube videos how to fix things at home, but he can actually can't actually do this. How your robots are different? Why they can do this?

I

you're not prepared.

I I I think I think there's hope for Nick. I think I think Nick can can can can can figure out how to fix the items from videos.

maybe he need more data to train.

Yeah. Well, in fairness in fairness, this robot watched 200 years of videos.

so maybe it's just a it might be it might be a volume question.

Can we do something like more like focus on post training?

That that that that is a smarter approach. But actually

reinforcement learning anything please. you're you're on you're on to you know something quite interesting here which is how much kind of post- training you need is purely a function of the power of your pre-training model

and I think the the best example of this is self-driving traditionally you know self-driving has been solved by collecting lots and lots of driving data and so there wasn't you know an internet scale pre-training in self-driving now now that's starting to change but this is kind of the traditional approach of solving self-driving and if you don't have a strong prior to learn a skill like driving you need to collect a hundred you know, million miles of data. humans, on the other hand, can learn how to drive in a couple of hours. you know, my wife says, "I still haven't learned, but most most people learn, you know, well enough after a little bit of time." And the reason we're able to do that is because, we have this really powerful pre-training. We've seen so much things, so many things move in the world. We've seen so many physical processes that when we get behind a wheel, it's an unintuitive motion at first, but we learn very, very quickly.

our approach to training robots is kind of similar to that. essentially we pre-train on this internet scale video data set you know hundreds of years of video data but then we do some post- training so for this specific task that you saw we had a pilot you know pilot this robot for you know on the order of around a dozen hours and then we use that as you know fine-tuning data to kind of teach the model exactly you know what motions and what strategy is best for general box opening of course you can't pre-program every trajectory if you tried to do that we would need to spend you know many lifetimes s collecting data just for this task alone. But the power of having a really strong pre-training prior is that now you can learn a task like this, a task that thousands of people, you know, are unfortunately, you know, spending a lot of hours doing with just 10 hours of Telop data.

And if you can do that, you actually have a really powerful tool for industry because there's a lot of repetitive enough where this approach works very well, but not repetitive enough to where you can use traditional automation tasks in industry. And if you can automate a task that companies, you know, are spending tens of millions of dollars on with 10 hours of post training, it completely flips the script of what can be automated.

You are doing everything. You're doing hardware, you're doing software, you're doing AI, and go to market everything at the same time. is like working on full stack is your mode or are you planning to focus on like specific layer later? Yeah, we didn't we didn't when we started the company, we didn't necessarily believe that we had to do everything full stack. I think we initially thought that there's, you know, a lot of interesting hardware out there. We can probably buy something off the shelf. maybe we focus on the tactile sensing or the hands, but really we've we've tested a lot of options. We actually went out and bought a couple humanoids. We ran them through like our rigorous testing process and we found that nothing's quite there yet. and our philosophy at ROA is we don't want to use you know our limited resources and time building something that's already invented but if something's not been invented yet and we need it to achieve our mission we're not afraid to build it.

and so by that logic we've started this deployment that you see on the video this is with off-the-shelf hardware but there are you know inherent limitations of the use cases we can tackle with off-the-shelf hardware. So, we're also developing our own full stack and with go to market ultimately I'm sure we'll have channel partners but I think the only way to you know build a real business is to speak with your customers understand their needs really deeply and make sure that you're solving 100% of their problem instead of solving what your impression of their problem is and for that reason you know it's it's really critical that we're working with customers very directly very closely walking factory floors and the only way to do that is to have a go to market function so I think at this point we are very full stack but you know not not entirely full stack.

We use you know a compute cluster that we didn't build and we use a lot of software library that we didn't build. So but but the things that we think are necessary to build we're not afraid to.

Yeah you're not afraid of challenges. that's actually good. Can you share what is your business and pricing model right now?

Yeah so I'm sure this will evolve over time. at this point we feel that you know the technology requires it it's not necessarily it's at a development stage where it requires a lot of expertise to operate and we want to take kind of all of those problems off of the hands of the customer and we're starting with a robot as a service model and what that means is if the customer has a use case like maybe they have a loading station where there's a lot of boxes need to be unpacked as one example or they have a returns processing workflow where there's a lot of items need to get returned and reviewed. we'll just essentially charge them an annual fee to take care of that to take care of that you know essentially scope of work and it's a subscription model where we both have a predictable stream of revenue and the customer can pay us a predictable amount and it's it's a lot easier for them to justify you know the ROI of switching to automation if it's just going from a variable cost to a more predictable variable cost as opposed to you know investing in heavy capital expenditures and then having to do maintenance which is very difficult for them to predict so essentially we're taking the work off the customer's plate and charging them an annual subscription fee.

I think the last question for me Yan Lean said that LMS are not going to work. A lot of people said that VAS are going to work. You mentioned they won't but video generative models are going to work. Yan Lehan and Fily are working on the world models. So my question is, has anyone asked robots what do they think? What do they prefer?

I we're we'll we'll make sure to ask them when we when we have them start speaking. but I I think yeah, ultimately in the world of robotics, it's it's been a for folks that have been in the industry for a decade, it's been kind of frustrating because there's been lots of waves of, oh, the robots are going to start working now. I think the the real Litmus test is real customer deployments. I think, you know, we're working in a lab is is easy and we've had incredible demos of robots doing cool things in a lab for more than 20 years now with very little real value in the real world. And I think that, you know, this year and next year are going to be really the inflection point years of intelligent robots getting deployed to factories and starting to create real economic value. And I think that's going to be the point at which people stop debating about the right AI, you know, technical approach and start focusing on, you know, the the metrics and and deployments and evidence that that comes as a result of the technical work.

And congrats on again with unicorn status and we are very happy to be the part of your journey and thank you so much for the incredible talk.

Thanks. We're grateful to have you on board. Thank you.