The promise of physical AI is that engineers will be able to program physical agents in the same way they program digital agents.
We’re not there yet. Robotics is still hampered by the lack of data from physical space. To train machines, companies need to build mock-up warehouses to test them, while the entire industry has begun efforts to train deep learning models to operate robots, especially monitoring factory lines and gig workers.
Another option is simulation. Detailed virtual replicas of real-world environments could provide roboticists with the data and workspace they need to do this work in a scalable way.
Antioch, a startup that develops simulation tools for robot developers, wants to bridge what the industry calls the “simulation-to-reality gap”: the challenge of making virtual environments realistic enough to ensure that robots trained within them can operate reliably in the physical world.
“How can we reduce that gap and do the best job of making the simulation feel like the real world from an autonomous systems perspective?” said Harry Melsop, co-founder of Antioch.
To that end, the company told TechCrunch today that it has raised an $8.5 million seed round at a $60 million valuation, led by venture firms A* and Category Ventures, with additional participation from MaC Venture Capital, Abstract, Box Group, and Icehouse Ventures.
Melsop founded the New York-based company with four co-founders last May. Two of the other founders, Alex Langshur and Michael Calvey, joined him to co-found and sell security and intelligence startup Transpose to Chainaosis for an undisclosed amount. The remaining two, Collin Schlager and Colton Swingle, worked at Meta Reality Labs and Google DeepMind, respectively.
tech crunch event
San Francisco, California
|
October 13-15, 2026
The need for better simulation is at the heart of what many large municipal companies are working on. For example, in the self-driving car space, Waymo uses Google DeepMind’s world model to test and evaluate its driving models. In theory, this technology would reduce the data collection required when deploying Waymo vehicles to new regions, reducing key costs when scaling up self-driving vehicle technology.
Building and using these models to test robots likely requires a different set of skills than creating self-driving cars, and Antioch wants to build a platform that solves that problem for startups that don’t have the capital to do everything in-house. These small companies also don’t have the capital to build physical test arenas or drive millions of miles on cars equipped with sensors.
“The vast majority of the industry doesn’t use simulation at all, and I think we’re at the point now where we clearly understand that we need to move faster,” Melsop said.
Antioch executives compare their product to Cursor, a popular AI-powered software development tool. Antioch allows robot builders to launch multiple digital instances of hardware and connect them to simulated sensors that mimic the same data that the robot’s software receives in the real world. These environments allow developers to test edge cases, perform reinforcement learning, and generate new training data.
That is, if the simulation fidelity is high enough. The challenge here is to make sure that the physics in the simulation match reality so that nothing goes wrong when the model is assigned to a real machine. The company starts with models built by Nvidia, World Labs, and others and builds domain-specific libraries to make them easier to use. Working with multiple customers gives Antioch deep context to refine its simulations, which executives say is not possible with a single physics AI company alone.
“What happened with software engineering and LLMs is exactly what is starting to happen with physical AI,” Çağla Kaymaz, partner at Category Ventures, told TechCrunch. “We do a lot of work with development tools, and we love this space, but the challenges are different. With software, you can have these poor quality coding tools, but the risks are usually much lower in the digital world. In the physical world, the risks are much higher.”
Antioch currently focuses primarily on sensors and perception systems, which account for the majority of needs for self-driving cars and trucks, agricultural and construction equipment, or drones. The aspiration for physical AI to power general-purpose robots that replicate human tasks is even further afield. Although Antioch is pitching to startups, some of its early work was with large multinational companies that were already investing heavily in robotics.
Adrian McNeil has a solid understanding of this field. He was an executive at self-driving startup Cruise, where he built the company’s data infrastructure, and in 2021 founded Foxglove, a company that provides similar data pipelines to physical AI startups. Mr. McNeil supports Antioch as an angel investor.
“Simulation is very important when you’re trying to build a safety case or when you’re trying to deal with very high-precision tasks,” he said Wednesday at the Ride.AI conference in San Francisco. “It’s impossible to drive far enough in the real world.”
McNeil hopes to see the same kinds of tools that propelled the SaaS revolution (platforms like Github, Stripe, and Twilio) emerge to support physical AI. “We need to make the entire off-the-shelf toolchain available in greater quantities,” he told TechCrunch.
“We all really believe that people who build autonomous systems for the real world will primarily be building them in software within two to three years,” Melsop says. “This is the first time we can iterate autonomous agents on a physical autonomous system and actually close the feedback loop.”
Experiments in this direction have already been carried out. David Mayo, a researcher at MIT’s Computer Science and Artificial Intelligence Laboratory, is using Antioch’s platform to evaluate LLMs. In one experiment, Mayo had an AI model design a robot and test it using Antioch’s simulator. You can also pit your models against each other in mock contests, such as pushing rival bots off the platform. Providing a realistic sandbox for LLMs could potentially provide a new paradigm for benchmarking LLMs.
However, before the world of AI engineers arrives, much work still needs to be done to bridge the gap between digital models and the real world. That would allow developers to create the kind of data flywheel that McNeil believes is key to the success of category leaders like Waymo, and engineers are increasingly confident that next month’s model will be more capable than last month’s.
If other companies want to replicate that success, they’ll need to either build those tools themselves or buy them.
