This simulation startup wants to be the benchmark for physical AI

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 scarcity of data coming from physical spaces. To train their machines, companies need to build model warehouses to test their machines, while an entire industry is emerging around monitoring factory lines and autonomous workers to train deep learning models to operate robots.

Another option is simulation. Detailed virtual replicas of real-world environments can provide the data and workspaces roboticists need to do this work in a scalable way.

Antioch, a startup-building simulation tools company for robotics developers, wants to bridge what the industry calls the simulation-to-reality gap — the challenge of making virtual environments realistic enough so that robots trained within them can operate reliably in the physical world.

“How can we do the best job possible to bridge that gap, to make the simulation look like the real world from the perspective of your autonomous system?” said Antioch CEO and co-founder Harry Mellsop.

To do so, the company told TechCrunch today that it has raised an $8.5 million seed round valued at $60 million, led by venture firm A* and Category Ventures, with additional participation from MaC Venture Capital, Abstract, Box Group, and Icehouse Ventures.

Mellsop started the New York-based company with four co-founders in May last year. Two other co-founders, Alex Langeshore and Michael Calvey, helped him found Transpose, a security and intelligence startup, and Sold to Chainalysis For an undisclosed sum. The other two – Colin Schlager and Colton Swingle – previously worked at Google DeepMind and Meta Reality Labs, respectively.

The need for better simulation is at the core of what many major independent companies do. In the field of self-driving cars, for example, Waymo is using Google’s global model DeepMind to test and evaluate its driving model. In theory, this technology would make deploying Waymo vehicles in new areas require less data collection, a key cost in scaling self-driving vehicle technology.

Building and using these models to test robots is arguably a different skill set than creating a self-driving car, and Antioch wants to build a platform that solves this problem for newer companies that don’t have the capital to do it all themselves. These small companies also don’t have the capital to build physical proving grounds or drive cars equipped with sensors a few million miles.

“The vast majority of the industry is not using simulation at all, and I think we now clearly understand that we need to move faster,” Millsop said.

Antioch executives compare their product to Cursor, the popular AI-powered software development tool. Antioch allows robot manufacturers to spin up multiple digital instances of their devices and connect them to simulated sensors that mimic the same data that robot software receives in the real world. These environments allow developers to test edge cases, perform reinforcement learning, or create new training data.

If so, the simulation has high enough accuracy. The challenge here is to ensure that the physics in the simulation matches reality so that when the model is put in charge of a real machine, nothing goes wrong. The company starts with models created by Nvidia, World Labs, and others, and builds domain-specific libraries to make them easier to use. Working with multiple customers gives Antioch a depth of context for improving its simulations that no actual AI company can match on its own, executives say.

“What happened with software engineering and LLMs is just starting to happen with physical AI,” Cagla Kaymaz, partner at Category Ventures, told TechCrunch. “We do a lot of work on development tools, and we love the industry, but the challenges are different. With software, you get these bad programming tools, and the risks are generally contained in the digital world. In the physical world, the risks are much higher.”

Antioch’s focus now is primarily on sensing and perception systems, which account for the bulk of the need for automated cars and trucks, farm and construction machinery, or drones. Aspirations for physical AI to power generalized robots to replicate human tasks are far-fetched. While Antiok targets startups, some of its early engagements have been with huge multinational companies that are already investing heavily in robotics.

Adrian Macneil has a strong understanding of this space. As an executive at self-driving startup Cruise, he built the company’s data infrastructure, and in 2021 founded Foxglove, a company that provides the same kind of data pipelines to actual AI startups. McNeil supports Antioch as an angel investor.

“Simulation is really important when you’re trying to build a safety case or deal with very high-precision tasks,” he said at the Ride.AI conference in San Francisco on Wednesday. “It’s not possible to drive enough miles in the real world.”

McNeil would like to see the same kind of tools that led the SaaS revolution — platforms like Github, Stripe, and Twilio — emerge to support physical AI. “We need to bring more of the entire toolchain to the shelf,” he told TechCrunch.

“We all honestly believe that anyone who builds an autonomous system for the real world will be doing so primarily in software within two to three years,” Millsop said. “It’s the first time you can use autonomous agents to replicate on a physical autonomous system, effectively closing the feedback loop.”

There are already experiments in this direction. David Mayo, a researcher at MIT’s Computer Science and Artificial Intelligence Laboratory, uses the Antiok platform to evaluate his MBA. In one experiment, Mayo commissioned AI models to design robots, then used an Antioch simulator to test them. It can also pit models against each other in simulated competitions, such as pushing a rival robot off a platform. Giving LLMs a realistic framework can help provide a new model for measuring them.

However, before the world of AI engineers emerges, there is still more work ahead to bridge the gap between digital models and the real world. If that can be done, developers will be able to create the kind of data flywheel that McNeil believes is key to the success of category leaders like Waymo, as engineers are increasingly confident that next month’s model will be more capable than the last.

If other companies want to replicate this success, they will need to build those tools themselves, or buy them.