Saturn Dynamics says its world model beats Nvidia's Cosmos by 73% ... and runs 38X faster

Arc vs NVIDIA Cosmos

London-based Saturn Dynamics says its Arc world model predicted robot trajectories 73% more accurately than an Nvidia Cosmos model in benchmark testing. And on Nvidia's own Jetson Thor hardware, using Nvidia's publicly released code for the comparison, Arc needed 0.6 seconds to generate each second of data. 

Cosmos Nano took 23 seconds. 

Cosmos Edge took nine.

That's a big claim from a small company. The trick, according to the company, is putting physics into the architecture. Instead of learning how the world works from mountains of video — or hardcoding physical laws — Saturn has distilled mathematical symmetries and physical principles into how Arc is built and trained, so the model doesn't have to relearn the same rules in every new situation. 

The result is a model that learns from fewer demonstrations and generates very, very fast.

"We can go at 250 frames per second, while the models are at one frame per second or 0.8," Seghetti told me on the Humanoid (almost) Daily podcast. "It is a massive gap in speed that allows all the systems to train faster, to execute faster, and run on edge at faster-than-real-time execution."

Speed matters because Saturn isn't trying to be the robot brain. It wants to be the gym. 

Seghetti's bet is that robots will need reinforcement learning environments generated at scale, the same way LLMs needed RL post-training before they were actually useful. Off-the-shelf robot policies, she says, top out somewhere around 70% or 80% success, and "RL, a reinforcement learning environment, is what is going to make it possible for a model to go from 80% success rate off the shelf to 99.99999%."

That's the gap between a cool demo and a deployment.

Saturn is also making a contrarian data bet. 

While Figure and others pour money into egocentric data, Saturn doesn't touch it: "we don't collect egocentric data," Seghetti says, preferring high-quality data from its own industrial deployments. 

(Others, of course, are betting differently: one startup recently raised $32 million on the idea that robots don't need a billion dollars of real-world data, and another UK company is betting $40 million it can train robots instantly via simulation.) 

In our conversation we also talked about ...

• why she doesn't think every task is a humanoid task
• why robotics today looks like the auto industry in the '90s
• Standard Bots letting factory workers vibe code a robot
• the researchers who got a hacked robot to stab a human hand in a test. 

Her prediction for the next 12 to 18 months: we'll have fewer robot data providers, and more new capabilities emerging from high-quality deployment data.

Here's the full transcript of our conversation.

Full transcript: Elisa Seghetti on Saturn Dynamics

John Koetsier: There's so much going on in robot training. Different techniques, different technologies, entirely different approaches. We've seen multiple companies raise money at huge valuations recently with entirely different ideas about how to do it. Real-world data, simulated data, teleoperated data, robots doing things, failing, and having some sort of evolutionary algorithm that rewards success. All these different things.

Today, I'm chatting with another company that is doing something different in robot data, and they're doing something actually really impressive that is beating some of what NVIDIA is doing as well.

We're talking to the CEO. Her name is Elisa Seghetti. The company is Saturn Dynamics. Welcome, Elisa.

Elisa Seghetti: Thank you for having me.

John Koetsier: Super pumped to have you. Great to have this conversation. Give us the 30 seconds first off on Saturn Dynamics. What are you guys doing?

Elisa Seghetti: Saturn Dynamics develops very efficient world models for robotics and industrial applications. We believe that one of the missing components in robot learning is the ability to generate, at scale, reinforcement learning environments that are going to make it possible for these robots to learn in a very efficient fashion.

In the same way in language we had LLMs helping Claude Code and Codex become possible, we need the same exact infrastructure for robotics.

John Koetsier: Okay, cool. So I mentioned at the top so many different methodologies for training robots. What do you see as the best, and what's most interesting? What are you doing that's different from a lot of other people?

Elisa Seghetti: Fundamentally, the way we see the market is that all these different approaches are not excluding one another. We're actually converging on some basic ideas, all of us.

The basic idea is that we need a lot of data that we don't have. And so you have people collecting it, people generating synthetic data, people going from teleoperation or egocentric data. But fundamentally, all of these modalities of data are useful. Then it's a slightly different priority for an individual company to prioritize one compared to the other.

The second thing that we all know is that we need a scalable way to train models that can somewhat follow the similar behavior that we have seen in language. The scalable way of training models will become better and better depending on the different architectures.

But one of the things we are sure about is that RL, a reinforcement learning environment, is what is going to make it possible for a model to go from 80% success rate off the shelf to 99.99999%.

That is what we actually need to deploy. Because any model, even GPT-3, GPT-4, was good but was not used for purposes before actually starting to apply RL post-training, et cetera.

We see that a lot of folks in the space see world models as another algorithm or another simulator. The way we see it instead, we see world models as something much more powerful, something that can help a lot of other off-the-shelf policies, so VLAs, even Astra that is trying to control robots, and other models to become better and to bridge that gap between 70% off the shelf to 99.9%, because this is fundamentally what we are actually missing to get to deployment.

John Koetsier: I love that you said Astra because there's another approach here that I totally didn't even mention right off the top, which is just frontier models zero-shotting or one-shotting it and just controlling robots and figuring it out and running some internal tests. There's been some amazing, cool things that have been released in that recently, and that looks interesting, right?

Of course, how often are you going to be able to run a frontier model on a warehouse robot, right? Assuming you can afford the tokens, right? I mean, it's just not reasonable to run that on 1,000 robots in a warehouse or something.

Elisa Seghetti: No, exactly. We totally agree with you, meaning that one of the biggest bottlenecks that we have in physical AI, let's assume that we have solved the data problem, that we have enough data, but then you get to a threshold where your model cannot be so big, like a trillion parameters, that you cannot deploy it on edge or you cannot deploy it at a reasonable cost.

Because at the end of the day, the real use case that we are seeing for humanoid robotics is the human workforce. That's the number one competitor. There are clear economics that cannot make sense if you need to run a 100,000-GPU cluster to have your robot powered, on top of the cost of the hardware.

So we absolutely need to find a way to get these models to run locally, to get them powerful while being small enough to make sure that we can deploy them locally.

And this is also one of the reasons that at Saturn, one of our key metrics is computational efficiency. Because we don't want to have an overbloated model that is very good at doing a lot of stuff, but not good enough to be deployed on edge when we're actually maybe distilling it or making it smaller.

We need to have something that, by definition, can be run in the conditions of deployment.

John Koetsier: Interesting. And you brought up post-training and other things like that. So where is your model sitting there? Does your model give a general sense of, hey, this is how the world works, and pretty good capability, and you still need some post-training?

Elisa Seghetti: So the way the world model works is you have an environment. It can generate frames. So you take an input initial frame and the intention of the robot to do a specific action, and then you can spin out future simulations of how the world is going to evolve.

That capability is interesting because instead of controlling the robot directly, it gives you information about how the future may evolve.

And so with this capability, what you can do is apply the world model across a lot of different moments in the development cycle.

A few people are starting to use world models for generating synthetic rollouts. And so you can generate capabilities or specific sequences of actions that would be hard for your robot to execute, or conditions and edge cases that would be hard to reproduce in real time.

People instead are using world models as a planner. So I have my policy that is calling my world model to see how the future may look, and then the planner will choose the best action, and I will be able to choose the best possible choice among a lot of them.

And the last way people are starting to use world models is the idea of world action models. So I use the prediction of the world model to generate an action that is going to get me to that state.

And this is the moment where, fundamentally, the world model takes over the role of a VLA and takes over the role of a controller and allows you to control the robot.

Saturn is mostly sitting in the first part of the use case, so we are mostly covering and supporting other policies in becoming better and accelerating the path. We are going to start working on world action models later on, but for now, the main focus is the first one.

John Koetsier: It's super interesting, right? Because you see all kinds of different things, and I did a story on Standard Bots two weeks ago or something like that, and that's of course not exactly humanoid. They're mostly smart arms or relatively smart arms, and they've deployed a training system where anybody on the factory floor with an iPad can train a robot what it needs to do.

That's pretty interesting. That's just making it super fast, super quick, super cheap. And so there's all kinds of evolutions at all the different levels that you were talking about.

Elisa Seghetti: No, absolutely. And we are seeing a new generation of robotics companies that are built for purpose to be AI-native.

Because if you think about companies like ABB, KUKA, Universal Robots, they have always been built to be programmed by system integrators. So you have an external company that goes in and helps you deploy your robot.

While there are companies like Standard Bots or the new generation of humanoids that have open SDKs, people can train and integrate in an extremely fast fashion, a bit like an MCP for robots.

And we want to partner with more and more of those because we see that this is going to empower internal teams inside the manufacturing floor to become users of the model, and not only the robotics companies or the system integrators or the end-to-end deployers to use the model.

In the same way, now it's completely normal for a non-technical person to call Codex and start doing some activities. In the same way, if you have very easy access to programming your robot, even the internal team could fix some deployment, could adjust some behaviors the moment your production line changes in three weeks.

John Koetsier: Which is huge, of course, right? Because you don't have to call an integration team in, you don't have to have three PhDs and five engineers come in and spend half a million dollars because you just moved something in your warehouse, and you can do it in a couple of days maybe, or a couple of hours or something like that.

Talk about your technology, ARC, and how that differs from Cosmos or Isaac Sim. You've got some pretty big claims there in terms of way faster and way more effective. Talk about that.

Elisa Seghetti: Yes. So compared to NVIDIA Isaac Sim, it's a very different product in that the world model is by definition a system that learns from data, while Isaac Sim requires 3D assets to be used in the simulation.

Compared to Cosmos, Cosmos is the perfect example for a comparison. It's a product that has been built mostly from a lot of video data coming from autonomous driving, drones, robotics.

ARC is built for purpose to be trained and tested in industrial robotics use cases. We call it, like, we build for machines that make. So a machine that needs to make other activities and do other tasks.

What we have seen is that these machines need to be fast, need to be able to adapt to a lot of different conditions, and so we need our model to match them in that requirement.

Doing this, we have built an architecture from the ground up that prioritizes speed, efficiency, and physical accuracy. And this architecture allows us to have a model that learns from way fewer demonstrations compared to others, and is much better at understanding the physical properties.

These two characteristics together make all the generation, all the activity of the model much faster and much easier.

And so we can go at 250 frames per second, while the models are at one frame per second or 0.8. It is a massive gap in speed that allows all the systems to train faster, to execute faster, and run on edge at faster-than-real-time execution.

John Koetsier: Super interesting. Where do humanoids fit into what you're doing? Obviously, that's the hot, buzzy area in robotics right now, and there's a lot of investment there. There's a lot of data being created there, and that's all wonderful and good.

Most of the real automation is still happening with somewhat more traditional automation functionality, although that's getting smarter as well. We're seeing AI being injected into that as well.

Where do you sit for all of it?

Elisa Seghetti: So fundamentally, the way we see humanoids is an amazing form factor for a lot of use cases, especially the wheeled ones.

When we discuss with our partners in manufacturing plants, in the industrial ecosystem, we see that there are a lot of tasks that require mobility across an area and interaction with the hands or with the gripper.

Today, the only solution alternative to the traditional humanoid form factor is an AMR, so a mobile base with a couple of robotic arms installed on top. That can be a solution that is cost-effective and also certified for actual deployment in a manufacturing plant.

So with a lot of the partners we are in conversation with, we are actually engaging towards this kind of form factor that looks a bit like a humanoid, but is a bit more certified and is a bit more mature in terms of the cost and the supply chain and the integration platform that you have.

We believe that humanoids are going to become particularly useful for all these kinds of tasks. We still see a gap in the price, so the business case is definitely really hard to justify for a lot of these platforms, and performance in terms of battery, payload at the wrist, and the hands.

The moment where hardware is actually not going to be a bottleneck anymore, supply chains will be more established, costs will be driven down, and we will see a lot of use cases.

That said, I don't believe in a world where every single task is a humanoid task.

I always say that we already have robots in our apartments. We have a washing machine, we have a dishwasher, and they are much better than me with my hands at washing the dishes and washing the clothes. It's a specialized form factor only for that use case.

So at some point we will realize which are the tasks that require a specialized machine and which are the tasks where the actual form factor of "let's move and let's operate with objects" is useful.

John Koetsier: I wonder if we can go way high back and kind of look at the entire training ecosystem as it's developing right now, and if you can give kind of a macro comment.

And it's probably challenging because you're in it and you're in a certain space and you see what you see. But you talked off the top and said there's value in all these different modalities, collection modalities, and technologies and everything like that.

A big question in my mind is when you see, and you don't have to talk about specific companies, but you see a company like Figure and they're investing, they say billions. They haven't raised billions, so whatever. But they're investing hugely in Index and egocentric data, and you see others.

And there's others that are making huge bets in different directions, and some are making bets on the frontier models, right? They're just making bets on the frontier models and some post-training.

Not all those bets are likely to pay off. Are people wasting money?

I mean, every nascent industry, every startup, before it's consolidated, before everything's done, people are wasting money. That's a guarantee, right?

But what do you see when you look at everything that's going on in training? Do we need everything that's happening? And do you think there are areas like yours, but others as well, that are the most interesting to pursue?

Elisa Seghetti: The way I see this market is, as you have said exactly, a very nascent industry, like a very young supply chain.

A lot of people always have the debate of vertical integration versus buy versus build, or horizontalization.

If you look at the market of robotics today, it looks a bit like the automotive market in the '90s, where everyone was trying to rebuild everything from the ground up internally, having all of that vertically integrated and having control of the full supply chain.

So what you have is that then your supply chain is extremely cost-inefficient because you cannot drive competition between the different layers of the supply chain because you have everything verticalized in-house.

What has happened then across decades is that you have tier one, you have tier two, you have people that are specialized in building wheels, specialized in building a specific component or chip that goes into the car.

Companies that become the definition of a software stack for automotive, and everyone in that layer starts having their winners and starts having their players in the space.

Because at that level, then you can drive competition across the horizontal, and you don't need to drive competition across the vertical layer.

So in the same way, when I see Figure, I see 1X, I understand why you're building everything in-house because you don't have a supply chain that is established enough for which you can go and buy the best world model from Saturn, the best policy from Pi, the best controllers and arms from someone else.

Each one of us is still in the early days, is still a seed, pre-seed company in terms of the volume of the capital, but in terms of the technology readiness that you have.

Then if we jump ten years from now, we are going to see this market much closer to having a clear, well-defined supply chain, infrastructure layers, where you are going to have some clear winners and some other followers.

Like in e-commerce, now you have Stripe that is dominating the payment system. You have Shopify that is the orchestrator around that, and you have all these different layers across an industry where 20 years ago everyone was trying to rebuild the final e-commerce in-house.

John Koetsier: Yes. I remember trying to build e-commerce back then, and it was hard. Something similar is probably true about robots right now.

That said, there are reasons to verticalize, right? Certainly in the US, we see competition to vertically integrate. Part of that is just due to import restrictions.

Part of that is due to, in terms of AI, knowing what's in your robot, knowing what's in your robot brain.

I also just reported, I think it was a week and a half ago, about what some researchers did. They injected some bad data, I guess, into some AI that robots were using and LLMs, and you saw that it had some dangerous results, right?

It could then be triggered to do something dangerous. In this case, they had it attack a human hand with a knife.

As we get humanoids in our homes and we're sleeping and they know where the kitchen is and they know where the knives are, that becomes interesting, right?

So there are some reasons to keep that vertical, but I suppose that there are also ways to ensure that partners that you're bringing in can be certified and known good.

Elisa Seghetti: Exactly. In the same way, now we start having the problem of how we can make sure that AI agents have cybersecurity layers, harnesses, and guardrails so that we don't have my AI agent helping me draft my email, taking control and writing to my bank account.

This is all that we are seeing already in the digital world.

Luckily for us, we can start seeing the effect in the digital world, taking measures, understanding how it works, putting harnesses, putting guardrails before we have this technology coming to the physical world, where it's then going to have, of course, a much bigger impact across multiple use cases, across multiple deployments.

Secondly, the thing that I'm also confident about is that deployments in B2B normally happen faster than deployment in the consumer market.

So before we start having volume of deployment in the physical world, we need to get through the classic procurement process of a large B2B company that will have all their ISO certification, SOC, cybersecurity tests, and everything we need to make a product compliant and ready for deployment before we actually start distributing it into apartments, into actual homes.

So I see the reason why we need to put a higher degree of control and a higher degree of integration compared to other markets.

But in the same way, if you're using Lovable, you're not controlling the brain that is building your product, that is behind your product. You're actually just using it.

In the same way, we are getting used to the fact that someone is going to be extremely good at building the intelligence, someone is going to be extremely good at assembling together the best hardware for that use case, and these companies are going to start partnering.

Like Copilot integrating OpenAI into their distribution channel.

I very much foresee a future where every robot is much more similar to a laptop. Each one of us is going to have multiple laptops or multiple mobile phones. We are going to have multiple robots helping us in different tasks for different use cases with different form factors.

The intelligence that I'm running on top of that can be because I want to choose that operating system.

Some of them are going to be closed, like the Apple ecosystem. Some are going to be more open, but it's going to depend on the preferences and the use case that you are going to choose one compared to the other.

John Koetsier: Super, super interesting. Love it. I've loved this conversation.

Maybe we'll end here. How do you see things evolving over the next six to 18 months for your company and for training data and training methodologies as a whole?

Elisa Seghetti: For us, I think that the big milestone in the next few months will be to start actively increasing the amount of data that we collect via our deployments.

We have a core belief that high-quality data is going to always win over volume and volume of low-quality data.

For this reason, for example, we don't collect egocentric data. We don't use ego data in our training corpus, and we get our data that we collect via our own deployments into real, high-value industrial tasks.

We see that we are at the very start of the exponential of the data that we can collect, the diversity of this data, and we want to see in the next 12 to 18 months what is the effect of this exponential in diversity and high quality of data, and how we can impact our model.

So with the scaling laws, what are we going to see out of that paradigm of more data, better capabilities, better generalization, and emerging behavior that we can see out of this distribution?

In general, what I'm seeing the more we move forward in the market is that the new modalities, the new architectures that come up, all of them are now getting closer to quality rather than brute force.

A few little demonstrations of high value. Yesterday, you talked with someone who did an incredible demonstration, and then the robot is able to execute the demonstration following that, rather than throwing a bunch of low-quality data and having your model that is overbloated.

So I will see the market of data providers shrinking a bit in the next 12 to 18 months, the emergence of new behaviors and new capabilities coming from the high-quality data coming from the deployments.

So the initial part of the exponential of the deployment.

John Koetsier: That's super interesting to hear because I've got people who are pinging me and saying, "Hey, should I get into robot data creation? Seems like a hot market," right?

I mean, they see things like Skild raising money. They see things like Physical Intelligence, which is raising now at potentially a $20 billion valuation.

If you do this well, there's good money in it for you, and you're doing something interesting as well.

Thank you so much. I really appreciate this time. It's been super interesting.

Elisa Seghetti: Thank you so much for all the questions, and it's been amazing meeting you.

Sources:

• Saturn Dynamics
• You Can Now Vibe Code A Robot: Standard Bots
• Nightmare Humanoid Robot Scenario: Hacked Robot Stabs 'Human Hand' In Test
• UK Startup Bets $40 Million That It Can Train Robots 'Instantly' Via Simulation
• Here's A $32 Million Bet That Robots Don't Need A Billion Dollars Of Real-World Data