Can robots learn from the internet the same way ChatGPT learned from text? In this episode, Andrew Wooten, co-founder of Rhoda AI, explains why his company believes the future of robotics isn’t collecting millions of hours of robot data … it’s learning from internet-scale video. Instead of relying on traditional vision-language-action (VLA) models that require enormous training datasets, Rhoda’s approach teaches robots physical intuition by predicting the future through video.
We also explore why, in Andrew’s opinion, warehouses and factories will likely be the first major market for humanoid robots (not homes!), why Rhoda chose a wheel-based humanoid design, how language models fit into physical AI, and how the company’s robots can learn complex tasks with just 8–10 hours of training data instead of 10,000+ hours.
Check it out:
In this episode:
- Why warehouses beat homes as the first market for humanoid robots
- Why Rhoda chose wheels instead of legs
- The biggest limitation of today’s robot AI models
- How internet-scale video teaches robots physics
- Why predicting the future helps robots manipulate the real world
- Edge AI vs. cloud robotics
- The role of LLMs in controlling robots
- How Rhoda cut robot training from 10,000+ hours to just 8–10 hours
- When zero-shot robot learning could become reality
Transcript: 1000X less training data required?
Andrew Wooten: You know, someone asked, “Well, why don’t humans have wheels?” And we thought about this deeply, and we actually found out that wheels have never evolved organically in biology.
John Koetsier: Can you train humanoid robots just on internet video? Maybe. Super pumped to share this conversation with Andrew Wooten, who’s the co-founder of Rhoda AI. They train their robots on internet-scale video, and Wooten says that this methodology lets its robots learn complex tasks with just eight to 10 hours of training data versus 10,000-plus hours for normal VLAs, vision-language-action models.
Check out our conversation here.
Andrew, let’s start from the top. Let’s ground this in reality. What are the use cases for humanoid robots?
Andrew Wooten: Yeah. So I think, when thinking about use cases, you want to start with: What are the categories where the robots are going to be deployed?
And there are a lot of conversations now around: Do you target the home with some kind of direct-to-consumer model? Do you target more commercial settings like hospitals and hotels, where you’re constantly interacting with humans? Or do you go after a more structured environment like a warehouse or a factory?
We thought about this early on, and we decided to go toward warehouses and factories. So the criteria that moved us in that direction were a couplefold. One is that there’s a really strong economic case there. If you’re talking about selling a robot into a home, there’s kind of a willingness-to-pay question there.
Is it comparable to a luxury vehicle, a car, an appliance, or a fridge? What’s this category? And people debate whether it’s $20,000 or $50,000, but it’s typically not on the order of $100,000 a year. When you’re talking about a factory or a warehouse, the willingness to pay is much higher because a lot of these places run three shifts a day.
Handling a workstation can cost upwards of $200,000 or $300,000 a year. So there’s an opportunity to create substantial economic value while still creating compelling economics on the humanoid hardware-software model supply. So that’s one notch in favor of logistics and manufacturing: the unit economics.
The other is that I think there’s a real technological maturity that’s expected when you sell into consumer segments. If you had to call a maintenance person to repair your TV on a weekly basis, I think you’d be very frustrated. And we take this for granted.
There are lots of technologies out there that have various maintenance cycles despite being very impressive technologies. An example is a tank or a fighter jet. This is a very mature, very impressive technology. It needs to be maintained every day.
John Koetsier: Yes.
Andrew Wooten: No one’s selling—I mean, there are other reasons you wouldn’t sell a fighter jet to a consumer, but perhaps one of the less important ones is that you can’t maintain it every day.
And I think that we should assume that humanoids early on are going to really benefit from the availability of maintenance. If you go into a warehouse or a factory, there’s already maintenance staff there, so it helps smooth that out. Practically, we’re building something that we’re targeting to be able to operate without maintenance for prolonged periods of time, but having that maintenance availability is going to create much better experiences early on. So, yeah: unit economics, maintenance, and also the structure of the tasks, I think, is helpful.
In the long run, we’re hoping for humanoids to take on very complex workflows with lots of steps, lots of decision-making, and lots of analysis. But early on, having a use case that’s bounded—some variability, but bounded in terms of potential branches of what you may need to do to succeed in the use case—is helpful.
In other words, something that can be learned pretty quickly is going to be really valuable. And then, finally, when you look at where the need is, there are 500,000 unfilled vacancies in manufacturing and logistics in the U.S. There’s a huge pain point there—a hair-on-fire kind of pain point for managers of factories and warehouses.
So I think there’s a lot of value add there. Kind of a separate point, but I actually personally get very excited about the impact that can happen in manufacturing and logistics. If you look at the last Industrial Revolution, it obviously affected everything: everyone’s quality of life, public health, art.
Everything we have in the 21st century is, in some ways, a product of the Industrial Revolution in the 19th century. And if you look at GDP scales, long-term GDP growth went from 0.2% before the Industrial Revolution to 2% after. And if you look at what the Industrial Revolution was, it really just changed how things were made and how things were moved, and that’s exactly what manufacturing and logistics is.
So I think sometimes people will say, “Well, home robots feel like a bigger mission.” I actually disagree. I think what gets us to the next step change in quality of life and advancement in civilization is a step-change upgrade in how things are made and moved.
So those are both my pragmatic and personal-excitement reasons why we decided to go after manufacturing and logistics.
John Koetsier: It makes sense. I totally see that, and you have a lot of company. There are a lot of companies as well that are going for the home market, and of course, that’s going to be niche, right?
Because there isn’t an economic imperative. I know somebody who’s ordered a NEO from 1X. It’s a $20,000 cost if they’re not going with a subscription, which I think is $200 a month or something like that, right? I’m guessing 1X is selling that at a loss, at least initially, and we’ll see how much it can do.
Hopefully, it doesn’t have maintenance requirements like fighter jets, which I think may require two hours of maintenance for every one hour of use. But we’ll see how that goes. When we were chatting about this before, you were also talking about how it’s not just humanoid robots; it’s a generalist robot frame as well, because you can do a lot of things.
We’ve seen some wheeled robots come out that are really, really quite cool. In fact, we’ve seen—I forget if it was Agibot or Unitree—that put their humanoid robot on wheels, and it’s balancing, it’s getting around. That’s a very efficient method of locomotion, by the way. If you need to save power because you can’t take that much energy with you, that’s great.
We also see, in the logistics, warehousing, and factory space, wheeled robots. They’re humanoid in a sense. There’s sort of a body, there are arms with lots of degrees of freedom, there are grippers or fingers or hands, but there’s a wheeled platform, which makes a ton of sense for the flat concrete world that they live in.
Andrew Wooten: Yep. Yeah, absolutely. I think what we’re thinking about at Rhoda is that you want to go to the customer, observe the environment, and have that inform your design while still preserving generality. So those are the two balances. If you look at a very specific use case and say, “What’s the perfect robot for this?” you might end up with a machine that can only do that use case.
On the other hand, I think saying, “Let’s make it look like a human because that’s the most general form factor,” is also taking it a step too far. Our approach at Rhoda has been to talk to dozens of customers, evaluate hundreds of use cases, and understand what embodiment is going to enable these use cases in an efficient, reliable, safe way.
John Koetsier: Yeah.
Andrew Wooten: And we actually landed on—we call it a humanoid, but effectively, it’s a humanoid on wheels. We think that a wheelbase is much more practical for many reasons. One is that it’s much safer if you need to E-stop the robot, and this is a certification requirement. You have to be able to E-stop any robot on a factory or warehouse floor.
So you need this E-stop capability, but if you E-stop something on legs, it’ll tip over. It might fall on a human, break—
John Koetsier: Yeah.
Andrew Wooten: —or hurt someone. So, just from a safety perspective alone, you want wheels. But also, wheels are faster, more reliable, can carry more payload, and can carry more battery capacity.
There are actually a lot of benefits to wheels. Someone asked, “Well, why don’t humans have wheels?” And we thought about this deeply, and we actually found out that wheels have never evolved organically in biology. There are some animals that kind of roll into a wheel and roll, but nothing has ever evolved wheels.
So maybe wheels are the ultimate mode of transportation, and one that nature just wasn’t able to evolve.
John Koetsier: One of the first technologies, right? Absolutely. I mean, wheels require gears and things that are hard to build in a biological system. So that makes a ton of sense. I want to go lots of different places with you.
I want to talk about how you build physical AI, which is critical, and the models we see today. I want to talk about LLMs, where they fit with physical AI models, and how you train. But we’re here right now, talking about your humanoid, your robot that you’re building. And I mentioned in our prep, you haven’t really come out there and said, “Hey, we’ve got a new robot. Here’s its name, here are its capabilities, here’s what it can do. We’re really excited about it, and here’s all the information.”
You’ve just sort of put something on your website that looks humanoid-ish. We see the torso and maybe a bit of the head, and you said, “Hey, this is what we’re working on.” Tell us more. Do you have a name for this robot?
And obviously, it has wheels. You mentioned that. What else can you tell us?
Andrew Wooten: Yeah. So I think Rhoda is fundamentally a physical AI company. Our co-founders are from the AI research space, so we really started with the AI model because we see that as the bottleneck.
Practically speaking, a bird with a single gripper, which is a beak, can weave a whole nest.
John Koetsier: Yes.
Andrew Wooten: There’s really a lot you can do if you have a very smart brain, even behind a very simple embodiment. So we felt that the bottleneck was really in the model and not necessarily in the hardware embodiment, and that’s where we started.
Now, that said, now that we’ve worked with lots of customers, identified use cases, and surveyed over 120 hardware options out there, we do think there’s an opportunity to make hardware that is more performant, safer, has higher payloads, and is just better suited to the applications we’re going after. And that’s why we chose to build our hardware.
But our roots are very much in the physical AI and model space.
John Koetsier: Yeah. It’s really interesting that you are building hardware because there was a report that came out from Bessemer just recently—I covered it in my Forbes column—and they said full-stack vendors have the advantage right now.
And there are a bunch of reasons for that, right? There’s a lot of vertical integration. It’s a new industry. It’s not super mature, right? There’s also a lot of data that you get—training data that you get—as you have your physical embodiment of your software, essentially, your AI, right? So there’s an advantage there, and it probably helps you on both sides of the equation: your physical AI as well as the hardware piece.
So I look forward to learning more about that when you guys are ready to release it. Let’s talk about the physical AI.
Andrew Wooten: Mm-hmm.
John Koetsier: There are lots of different ways that people have built that, right? Starting from way back in the day, coding every possible scenario that might happen, to: We have LLMs, we can just feed them into our robots, and our robots instantly know many, many things. They know, okay, I’m putting away the groceries, and the milk goes in the fridge, and the cereal goes in the cupboard, or something like that, right?
To: Well, LLMs aren’t really enough. We really need to understand physical reality. Where are you right now?
Andrew Wooten: Yeah, exactly. I think that’s a great way to illustrate the history of physical AI. When we started Rhoda, we identified that this was the real bottleneck: There’s a gap in manipulation.
There are a lot of robots out there that can do tasks repeatedly and reliably, like your industrial machines—your Fanucs, your KUKAs. They’re pre-programmed machines, so they do the same task all day, every day. And these are actually really the only robots deployed out there, at least in factories at scale, creating economic value.
Then there’s a second category of robots that is referred to as intelligent robots, and here the problem is that they work well in a lab, but virtually none of these have translated into real-world deployments. And the problem is that they’re learning from data, but the dataset is grossly insufficient.
So if you look at—I’ll just take a step back. If you look at how successful paradigms in AI have been trained, like language models or video models, it’s always the same structure. You start with internet-scale pre-training. So in the case of ChatGPT, OpenAI downloaded effectively the entire internet of text, used that for its pre-training, and trained a transformer model using internet text.
And this is a model that emerged with capabilities for understanding logic, reasoning, and outputting text with good grammar. So there’s something very interesting going on there. And then they fine-tuned it with about 10,000 examples of questions and answers. And these examples kind of taught the model how to be a helpful assistant.
So this is the paradigm of how we’ve seen successful AI models be trained. This is how ChatGPT, Claude, and video models, in their own way, are trained: You start with internet-scale pre-training and then fine-tuning. The problem in robotics is that there is no internet-scale robot data.
So robotics is fundamentally a data problem. And how a lot of labs have approached this—I’d say the big paradigm of 2025 was the VLA, the vision-language-action model. The idea there was, “Well, we don’t have internet-scale robot data, so let’s instead take a backbone of this smart model—an LLM or a vision-language model.” This is a model that can look at a picture, reason about the picture, identify items, and describe the steps to accomplish something.
So they said, “Let’s use this as a backbone, and then let’s post-train that with some robot data.” What you get from that is a model that can reason in language but is actually really bad at executing the tasks without—
John Koetsier: Mm.
Andrew Wooten: —post-training data. So it might be able to realize, “This is a bottle,” but when it goes out to pick up the bottle, that action isn’t reliable.
And the reason is that the action is solely relying on that robot-data post-training. So you’re getting language-level reasoning from your internet-scale pre-training. You’re not getting any physical reasoning from your pre-training. And this is the fundamental flaw with VLAs. How folks were addressing this was by trying to collect more and more robot data.
So if it doesn’t work with 10 hours of robot data, let’s collect 100,000 or 70,000—or someone announced 700,000 hours—of robot data and put that on top of our vision-language model. And that started to kind of work for very specific tasks, but not to a level of reliability that made it useful in production, which is why all these demos were effectively stuck in laboratories.
And reliability is really just a special case of generalization, if you think about it, because reliability means being able to solve the task in any kind of permutation that can hit you, which requires generalization. So our approach is fundamentally different. We said we don’t believe that you can solve the robotics problem by trying to collect an internet-scale amount of robot data.
We just think that’s never going to scale. And it’s not even a path necessarily worth going down. What we said instead is that we wanted to leverage an internet-scale dataset to teach the robot about physics and manipulation, and actually teach it this action modality by learning from an internet-scale dataset.
And the background of how our team came up with that is that my co-founder and Rhoda’s chief scientist, Eric Chan, did his PhD in modeling 3D spaces. He was one of the early people who trained a generative model on 3D spaces. And if you think of 3D spaces, it’s a problem that’s kind of similar to robotics.
The way a lot of folks were approaching it was that you could either create the 3D space in simulation, and that way create a dataset and train on it, or you could go out and collect a lot of 3D data with a special camera and train a dataset on that. And Eric’s insight was that neither of these approaches was going to yield the scale and diversity we need.
Instead, he came up with the idea of leveraging images and refactoring those images as a dataset that enables us to build 3D spaces. And because you have virtually infinite image data, he was able to train a generative model that is a lot more capable than one trained on these specific datasets.
So, bringing that back to robotics, Eric’s insight was: Instead of trying to collect these datasets, let’s use a dataset already out there, which is internet video data. And if you look at internet video and the video models trained on this data, there’s a lot of physics knowledge in there.
If you’ve tried Sora, which was OpenAI’s video model, you can ask it, “Generate a video of a horse walking up a hill,” and it’ll generate that video, and it’ll look very realistic. The physics will be quite plausible. Not perfect, but enough to where you realize, okay, this model is starting to understand physics. And that was really our insight: If you watch enough videos, you can start to get an understanding of physics that’s not coming from robot data, not coming from simulation, but actually coming from the real world.
And this is the core of Rhoda’s model. Specifically, how our model architecture works is, if you’re trying to do a task like having the robot grab this cup, the robot in real time will first generate a video of itself grabbing the cup, and then it’s going to convert that video into actions in real time.
So that’s essentially how we’re approaching physical AI. And the difficulty—why hasn’t everyone done this already?—is that it’s actually very hard to build a video model that meets the right specifications for this because most video models don’t run in real time. They take 30 seconds to generate a video.
And even if it does run in real time, the physics have to be accurate enough that you can control a robot. So what we spent the first year and a half at Rhoda doing, before we came out of stealth, was really focusing on building a video model that’s fast enough—it runs twice as fast as real time—and accurate enough that it has real-world accuracy of physics such that you could use a video model to control a robot in real time directly.
So that’s essentially the approach that we’re using at Rhoda.
John Koetsier: This is what I love about startups. You have an idea: We’re going to use video. It’s already out there; we’ve just got to grab it. Now we need to use it, ingest it, use it way faster, analyze it way faster, and get way more data out of it than anybody’s ever done before.
It’s going to take us a year and a half to do that. You don’t know that in advance, but it’s going to take a year and a half to do that, and there’s no guarantee of success. Maybe it’s just not good enough. Maybe it just won’t work. Maybe you’re not smart enough. Who knows, right? And at the end, you’re either kings or paupers, right?
You win or you lose. That is so crazy. That is really crazy. So there’s a lot of edge AI going on here. That’s a question I was going to talk about: edge versus cloud, right? The robot is calculating that, making that vision model in real time. You’ve got some compute on this thing.
Andrew Wooten: Yeah. So the way we think about it is that there’s the higher-level video-prediction task—
John Koetsier: Yeah.
Andrew Wooten: —and then you can have a lower-level control task. The video prediction you can run—I don’t want to say exactly—but at a frequency that is a couple of times a second, or multiple times per second. And that’s actually a frequency at which you can choose to either run that through the cloud or on the edge.
The lower-level control, you definitely want to be running on the edge. We have various solutions where the compute is either local or in the cloud, but because we’re making a video prediction that’s about half a second into the future, we don’t necessarily need that video prediction immediately before the action takes place.
So if you make a video prediction and then spend 100 milliseconds getting it to the robot and processing it, you still have enough time left in that prediction for it to be useful. So our architecture is actually kind of unique in that you could run it through the cloud if you chose to.
John Koetsier: It’s really interesting. I’m trying to compare it to a human model, and of course, we have no idea of the complex physics that we are doing almost every second of our lives, right? We catch a ball, we throw something, we walk, we sit down somewhere, we lift a weight. Whatever we do—you lift your cup there.
You didn’t even think about it. Your conscious mind wasn’t even present there, and you take a drink. But there’s an immense amount of math going on somewhere at some low-level processor inside here to make that happen. Do you see analogies between how we do things and how Rhoda AI is working?
Andrew Wooten: Yeah, I actually think that’s very well put. I would posit that there’s no math, in the traditional definition of math, going on. I think it’s almost an intuition. It’s like a physical intuition that happens in the back of your mind.
One way to solve robotics is to try to define everything you can see, convert it into numbers, convert it into exact points, and calculate physical formulas. And we’re doing almost the opposite of that, where we’re watching a lot of videos, and from those videos, our model learns an intuition for how things move.
And I actually think that is much more human because kids can catch an apple way before they learn about the laws of gravity. So I think, with our approach, because you’re learning from lived experience, essentially a human brain can calculate physics through lived experiences.
Not necessarily calculate, but you’ve seen enough apples drop that you have an intuition for how it’s going to drop. And I think that’s exactly how you could describe our physical AI reasoning. It’s based on an intuition of physics that comes from real-world video data.
John Koetsier: What’s fascinating about your model—I’m just saying this hearing it for the first time, maybe I’m 100% wrong—but what’s fascinating about your model is that it works on predictions of the very near future.
Where is this cup? Where will that person be if they continue on this path? And you can refine that every couple hundred milliseconds—five times, 10 times a second, something like that, right? That’s really interesting because then you don’t need to calculate three access points of where you need to move your arm.
You need to move your arm closer, and then you can get closer, and eventually, you’re right there. You don’t have to have the exact path. That’s really quite interesting.
Andrew Wooten: Yeah. And another cool thing about it is that because we’re predicting in video, we can actually run that autoregressively. So if we do want to predict 30 seconds in advance, we can make one prediction and then feed that into the model and make another prediction. And it’s actually very useful in specific instances.
If you’re trying to solve a puzzle, you could make a tree of predictions of how you are going to solve that puzzle in video—
John Koetsier: Ah.
Andrew Wooten: —and then choose the one that successfully solves it and actually physically map that one out. I think trying to predict 30 seconds into the future how a human operator is going to act probably isn’t useful. But trying to predict different strategies for how you could assemble something, and what sequence of events you should choose, is actually very useful when you’re able to layer on predictions.
So with a video model, you can actually make a farther-out prediction. But yes, essentially, you also have the power of recalibrating your prediction with inputs that you’re receiving every couple hundred milliseconds.
John Koetsier: Your model for AI and robot control—does it have to do the predictions every single time, or does it build up a certain level of knowledge base so that if the robot is doing the same thing, it doesn’t have to? Like a subconscious, like when you grabbed your cup.
Andrew Wooten: Mm-hmm.
John Koetsier: You’ve done that a million times. You didn’t have to think about it. You didn’t have to make too many predictions. You probably didn’t even have to look down except for one glance to make sure it was right there. Is there anything like that level of learning?
Andrew Wooten: Yeah, I’d say the way to think about it is that making one prediction at a time into the future is kind of like the baseline.
So you could compare this to a subconscious way of doing the task. The model can actually slow down, make multiple predictions, and branch out those predictions, and that would be more comparable to human reasoning. So I think the baseline is the video-prediction-to-action-conversion loop through our video-action model.
There are modalities where you could apply higher levels of reasoning to it.
John Koetsier: I keep thinking of the movie Next with Nicolas Cage from, like, a decade ago, where he can predict the future a little bit. You see him sort of mapping out where he would go and everything. So that’s kind of in my brain at the moment.
Hopefully, I have the right name for that movie.
Andrew Wooten: Yeah. That is what it’s called. It’s a good one.
John Koetsier: Yeah, exactly. Okay, super interesting. Now, not all robots, especially if they’re going to just operate on a production line or in a warehouse or logistics center or something like that, need to communicate.
On the other hand, it’s nice if you have a humanoid model, vaguely humanoid or fully humanoid, and you can say, “Hey, get that thing,” or, “Hey, do this,” or whatever. Are you going to build that functionality in? And if so, what AI are you going to use for that?
Andrew Wooten: Yeah, absolutely. First of all, we definitely want to be able to enable operators and humans to interact with the robot and give the robot instructions.
There are a couple of layers to this. I think the higher-level, back-and-forth communication layer you would want to do with a language model. So you could use an open-source one. I don’t think you necessarily have to train a language model from scratch. But yeah, the higher-level language you would do with a language model, and then my guess is that language model would break down the high-level language communication into specific instructions.
And then you need to have a bridge that takes those instructions and applies them to the physical AI model. What we would likely do here is text-conditioned video generation. Essentially, our model operates by generating videos of what the robot should do.
We can also condition these videos on text. So if you have a high-level conversation where you go back and forth with the robot, describe what to do, and the robot realizes, “Okay, I need to move left and pick up this tennis ball,” then, as the videos are being generated, it would be given the text condition, “Move left. Pick up the tennis ball.”
And then that text condition would generate the videos, or it would influence the videos that are being generated, and then those videos would be converted to actions on the robot. So you still have that full language pipeline, but you would probably use a language model for the high-level reasoning and then text-conditioned video generation for the actual execution layer.
John Koetsier: I’m going so neuromorphic on all this stuff. Maybe that’s just a human-centered viewpoint; I’m not totally sure. But you’ve got the prefrontal cortex, right? Executive function, communication, and all the stuff like that. You’ve got the stuff that’s closer to the spine, the base of the skull—you know how to move, proprioception, all these other things.
So fascinating how you’re doing this. Is the robot that you’re building that will embody these things—and I’m assuming you offer your AI to others as well if they want it—the robot that you’re building, do you have an idea of when you want to release it? Is it going to be mostly an internal test thing, or what’s your plan?
Andrew Wooten: Yeah. So we started with the model because we felt that was the bottleneck. And now that we have a model working that we think can deliver real value in the real world, in industry, we’re currently using off-the-shelf hardware to deliver our model into real facilities.
So we’ve partnered with a specific company. We buy their off-the-shelf hardware, integrate it with our model, and then that’s what we’re planning to ship to customers for production. That’s what we’ve used for the proofs of concept we’ve done so far, and it’s what we’re going to be shipping into production this year.
John Koetsier: Okay.
Andrew Wooten: But what we found when we went out and surveyed the hardware—we started with humanoid robots because we thought all the degrees of freedom we might need were there—was that there were a lot of critical flaws with many of the humanoid robots out there. I’d say the three main ones were reliability, safety, and payload.
If you want to be operating in a factory or warehouse, you need to be reliable for three years. You can’t have an actuator burn out every week. The second is safety. Virtually none of the humanoids out there are certified to operate in factories, warehouses, or really anywhere around consumers. You need a CE mark or a UL standard in the U.S.
Effectively, almost no companies have these certifications. And then the third is payload. We actually found that, surprisingly, a lot of robots out there are spec’d for payloads of maybe three kilograms, and that’s just not very helpful. They might say 10 kilograms, but they mean 10 kilograms peak, not 10 kilograms rated.
What you need is a robot that, if a box needs to be picked up, can pick up the box and doesn’t over-torque or break. So there are a lot of these really core issues, and what we found is that the best hardware to start with is actually industrial hardware.
We’re using a seven-degree-of-freedom arm. You can see this in our launch video, but it’s effectively an arm that would otherwise be running algorithmic control in a factory. It is rated and certified to be in a factory or warehouse, doing real industrial applications. It has the right payload and all the right communication bandwidth.
It’s safety-certified. And then we just put our model on that hardware. What we’re building is a robot that’s actually going to solve those problems because, with the industrial hardware, you get the safety certification and the payload. What you don’t get is reachability. You’re usually installed on some fixed platform, and there are a lot of fixed use cases you can tackle out there, which is what we’re starting with.
But you want to have that mobility. You want to have that larger workspace. And if you’re trying to get a robot that is safety-certified, has a high payload, has high reachability, and is reliable, there’s just nothing out there. And if there were, we would just buy it. But there isn’t, so we have to build it.
And our philosophy really is: If something exists, let’s partner or buy it. But if something doesn’t exist, we don’t want to be afraid to build it. So that’s been the outcome of our reasoning on hardware.
John Koetsier: I love it. There’s got to be a Steve Jobs quote in there, like people who really care about customer experience want to build the whole thing.
So that’s interesting. Fascinating. When do you think we’ll have truly learning robots? For many of the robots—and I was just in a factory in Boston last week—there are literally 100 humanoid robots in this factory, and they’re all learning. They’re all training.
They’re all actually trying to do stuff: lift stuff, move stuff, put it in a box, take it out of a box, stack it, all that stuff. And they are gathering training data. Also, when they go into a customer’s actual facility, they’re going to have an onboarding phase, a training phase. It’s normal for humans.
We have a training phase ourselves when we get a new job. We’re pretty quick at a lot of stuff, right? I could probably go into your kitchen and make a cup of coffee. Probably no humanoid robot can do the same. But there is a training phase. When do you think we’ll have a robot that you could just ship, somebody turns it on, and somebody says, “Hey, your job is to stand by that conveyor, lift things off, and put them over there, and I’ll tell you other things as we go along”?
When do you think we’ll have that?
Andrew Wooten: Yeah. I think that’s a great way to frame it. What you’re talking about is how efficient the learning protocol is. And I think if you look at a VLA, like a vision-language-action model, some of the examples out there have been that you need to collect on the order of 10,000 hours of data to learn a pretty simple task.
And that’s obviously a huge operational problem, especially if, after those tens of thousands of hours, it doesn’t generalize, so it can literally only do the configurations that were in the dataset. I’m very excited about where we are right now at Rhoda. What we’ve demonstrated is that we can learn pretty complicated tasks with on the order of eight to 10 hours of robot training data. And that actually opens a lot of possibilities because, if you’re talking about automating a task in perpetuity, spending eight to 10 hours training it is actually very cheap.
Now, it’s still very valuable to continue to work on making that number go down further. An example is that you might be in a warehouse, and your job might be packaging items or kitting items, and you’re receiving a new type of item every couple of hours. If you need to learn how to kit different items in a couple of hours, then that practically doesn’t work.
So then you need to have an even faster training modality where someone can explain it to you with text or show it to you with a human demonstration, and then you can use that human demonstration and text to immediately execute that task. And that’s something that we’re currently working on. We showed a couple of human-following demos, but this is something that we’re actively doing a lot of research on.
So, as of right now, where we are—and I think where the physical AI world is in terms of efficient learning, and I believe Rhoda is at the frontier of that—is that we can learn a somewhat complex task, like unpackaging something and recycling all the materials, with on the order of eight to 10 hours of training data.
But the vision is to get to a point where you can do tasks with text or visual instruction zero-shot. And I think it’s hard to make predictions, but it feels like that would be on the order of one to two years away.
John Koetsier: Cool. I want to thank you so much for taking this time. It’s been fascinating to learn about what you’re building, how you’re building it, and how you see the world and the ecosystem.
Thank you so much.
Andrew Wooten: Yeah, great speaking with you. Thank you.