The other day I was scrolling through LinkedIn when I came across a post from Harry Mellsop, Co-founder of simulation startup Antioch, who said that his company had just tripled its revenue in a single month.
“The thing that nobody tells you about building a company, is that when it rains, it pours,” he said.
The post was designed to attract new talent as they are “beyond inundated”, so need to hire across “every part of the business”, from core engineering and simulation to operations and go-to-market.
After a quick search of their website, I discovered that Antioch has only been around since May 2025, but has already raised $8.5 million – and another much bigger raise is coming soon (more to follow).
Antioch is not the only young physical AI company making rapid progress. Munich-based Microagi, which collects factory and household data to train humanoids, raised $55 million in July, only ten months after launching. In the UK, Humanoid went from an idea to a $1.35 billion valuation in just two years.
What is going on here?
“We now have intelligence on tap which is opening the floodgates to some very exciting applications,” Melsop told IZON in an interview.
“Five years ago, the hardest problem in robotics was the intelligence problem,” he says. How do you get data about the world from around you? How do you make sense of that and turn that understanding into useful actions? For a long time the best approach was a cobbled together series of different models that each did small things, and so thinking about scaling that was really scary and intractable.”
Now, he argues, the situation is changing rapidly. The emergence of increasingly capable AI models, combined with the computing and simulation infrastructure being built by companies such as Nvidia, means that robotics companies can increasingly access intelligence as a commodity.
Building the infrastructure for physical AI
Antioch is a New York-based agentic simulation platform that allows robotics companies to design, train and test their physical systems in cloud simulation.
By moving development out of the field and into software, Antioch aims to reduce reliance on what it describes as expensive, slow and incomplete real-world testing.
Antioch’s customers, including Fortune 500 and major technology companies, span drones, autonomous ground vehicles, humanoids and intelligent perception systems such as advanced security cameras. The common thread is the intersection between hardware, software and an intelligence layer.
The company integrates and amplifies a variety of bleeding-edge simulation tools, including Nvidia’s Isaac Sim, a GPU-accelerated robotics simulator built on the Omniverse platform, and Isaac Lab, a robot learning framework for training reinforcement learning models.
“Do you think it’s fair to say that Antioch wouldn’t exist without Nvidia’s technology?” I asked Harry.
“Yeah, for us, I think that’s true. I don’t think we could have done what we’re doing today without Nvidia. The Isaac and Omniverse ecosystem is brilliant and means we can build on top as the connective tissue.”
Antioch’s three enablers for customers
Harry explains that his company supports customers in three broad ways.
“First is development. So as companies look to actively engineer these robotics systems, the status quo today is they have the hardware, they’re writing the software and they’re taking it into the field to test it. That’s your development loop so we aim to replace a lot of that.
“Step two is evaluation. When you’re looking to launch something, you want to run a comprehensive benchmark of the performance, of the safety, efficacy, these types of things. We allow teams to do this at scale in simulation.
“The third is with synthetic data generation. As you’re training these models, you know real-world data is always best if you can get access to that, but in many cases you can’t, and even market leaders can’t. And so that’s where simulation and world models come in to help fill the gaps and improve performance.”
Making simulation trustworthy and scalable
The crux of the challenge for platforms like Antioch is producing simulations that reliably match reality, what is known as the sim-to- real gap. In other words, virtual environments need to be realistic enough that robots trained or tested inside them can operate reliably in the physical world. If a model is put in charge of a real machine, customers need to know that the results will stack up.
Then there is the question of scale: once you have a trustworthy system, how do you go from running one simulation to hundreds of thousands in parallel to explore all the edge cases a robot might encounter in the real world?
There are plenty of off the shelf simulators available today, but Harry says Antioch differentiates itself in part by building domain-specific libraries on top of Nvidia’s software, making them easier and cheaper for customers to use, while extending their capabilities.
These libraries include solutions for “notoriously challenging” simulation problems, like modeling certain sensors or deformable materials. While a physical AI company might spend months solving one of these problems, Harry says, Antioch provides them off the shelf.
“We take the best out of our ecosystem and then invest deeply into them to improve them further for our customers – it’s a fairly hands-on engagement with our customers.”
The step-by-step
I asked Harry to imagine I was one of his automotive clients that needed a robot to do some heavy lifting by hauling tyres.
“There’s two parts to the problem here,” he says. “How do you create the digital twin of the robot, and how do you create the digital twin of the environment?”
For the robot, customers give Antioch everything that goes into its definition such as CAD files, 3D geometries, hardware components, sensors, motors, software and machine learning models.
“That gets you like 80% of the way there,” says Mellsop.
Antioch automates the process of putting all of that into a simulation, before the next stage: calibration.
This is where the company will often go on site to capture data about how the real robots are operating, allowing it to make the digital version behave more like the physical one. The environment is built in a similar way. Antioch can use photogrammetry, lidar and other techniques to scan a customer’s site, with the resulting data fed into tools for Gaussian splatting, mesh generation and other world generation techniques. The end result is a high-fidelity digital world containing a high-fidelity digital robot.
This digital twin is then put through the same paces as the real robot.
“In Antioch, you can literally see the real robot and the digital twin rolling out side by side,” Harry says. “The technology inspects for differences between the two. Ideally, those are negligible. If they’re not, then we continue iterating with the customer until they are. This process of putting simulation alongside the real system, what we call ‘pair setting’, guarantees that, in calibration scenarios, the two systems behave identically. It also provides a strong statistical basis for expecting that this agreement will generalise to settings outside the calibration set.”
What Antioch has cracked so far
Closing the sim-to-real gap is top of the agenda for Antioch, and Harry breaks their progress down in three ways, robotic perception, robotic dexterity, and modelling the physical phenomena of the world the robot interacts with.
“On the perception side, so certain types of cameras, radars, lidars – early in our journey these problems were very, very difficult for us,” Harry says. “But we’ve since gotten very good at that to the point where I’d tentatively claim those perception-to-real gaps are gone for our customers, and we’ve got the data to back that up.
“The dexterous sim-to-real gap, when you’re dealing with motors, actuators and interacting with grippers, remains much more difficult. We can either do it cheaply with a bit of a sim-to-real gap, or we can do it in a more expensive way with a very minimal gap.
“The final gap is modelling the physical phenomena of the real world; when a robot interacts with something like stamping and compression, cutting and these types of tasks. I would say we’re somewhere in the middle today. We’re getting results that are easily good enough to shape business decisions for our customers.”
What’s coming next
For the foreseeable future, Antioch’s focus is to bring on more clients, improve simulation fidelity and the range of systems that can be accurately modeled. The next step is to start opening up access.
“Ultimately, the model that we would like to run would be like ChatGPT, where anyone can make an account and get started using Antioch. We’re on a graduated pathway to get there and have some very exciting announcements coming soon.”
