The vision for physical AI is that engineers will eventually program physical agents with the same ease as digital ones. We haven't reached that point yet. Robotics continues to face limitations due to a scarcity of data from real-world environments. To train their systems, companies often construct mock warehouses for testing, and an entire industry is emerging to monitor factory lines and gig workers, gathering data to train deep learning models for robots. An alternative approach is simulation; creating detailed virtual replicas of real-world settings could offer the scalable data and testing environments roboticists require. Antioch, a startup developing simulation tools for robot developers, aims to address what the industry terms the sim-to-real gap—the difficulty of making virtual environments so realistic that robots trained within them can function reliably in the physical world.

EMBED_PLACEHOLDER_0

"How can we do the best possible job reducing that gap, to make simulation feel just like the real world from the perspective of your autonomous system," said Antioch CEO and cofounder Harry Mellsop. The company recently secured $5 million in seed funding at a $60 million valuation, led by venture firms A* and Category Ventures, with additional investment from MaC Venture Capital, Abstract, Box Group, and Icehouse Ventures.

Mellsop established the New York-based company with four cofounders in May of last year. Two of the other founders, Alex Langshur and Michael Calvey, previously collaborated with him to found Transpose, a security and intelligence startup that was sold to Chainalysis for an undisclosed sum. The remaining two founders, Collin Schlager and Colton Swingle, have backgrounds at Google DeepMind and Meta Reality Labs, respectively.

The demand for improved simulation is central to the efforts of many leading autonomy companies. For instance, in the self-driving car sector, Waymo utilizes Google DeepMind’s world model to test and assess its driving algorithms. Theoretically, this approach will reduce the data collection needed to deploy Waymo vehicles in new regions, a significant expense in scaling autonomous vehicle technology.

Developing and employing these models to test robots involves a distinct skill set compared to creating a self-driving car. Antioch intends to build a platform that addresses this challenge for newer companies lacking the capital to develop such capabilities independently. These smaller firms also cannot afford to construct physical testing facilities or operate sensor-equipped vehicles for millions of miles.

"The vast majority of the industry doesn’t use simulation whatsoever, and I think we’re now just really understanding clearly that we need to move faster," Mellsop remarked.

Antioch’s executives liken their product to Cursor, a popular AI-powered software development tool. Their platform enables robot builders to launch multiple digital versions of their hardware and link them to simulated sensors that replicate the data the robot’s software would encounter in reality. These virtual environments allow developers to test edge cases, conduct reinforcement learning, or generate new training data—provided the simulation is of sufficiently high fidelity.

The core challenge is ensuring the physics within the simulation aligns with reality, so that when the model controls an actual machine, no issues arise. The company begins with models from Nvidia, World Labs, and others, then develops domain-specific libraries to enhance usability. By collaborating with multiple customers, Antioch gains a breadth of context for refining its simulations that surpasses what any single physical AI company could achieve alone.

"We do a lot of work on dev tools, and we love that vertical, but the challenges are different. With software, you can have these bad coding tools, and the risk is generally pretty contained to the digital world. In the physical world, the stakes are much higher," Mellsop explained.

Currently, Antioch primarily focuses on sensor and perception systems, which represent the majority of needs in automated cars and trucks, agricultural and construction machinery, and aerial drones. Broader ambitions for physical AI to power generalized robots capable of replicating human tasks remain more distant. Although Antioch targets startups, some of its earliest partnerships have been with large multinational corporations already making substantial investments in robotics.

Adrian Macneil possesses deep expertise in this field. As an executive at the self-driving startup Cruise, he developed the company’s data infrastructure, and in 2021, he founded Foxglove, a company providing similar data pipelines to physical AI startups. Macneil is supporting Antioch as an angel investor.

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

Macneil hopes to see the emergence of tools akin to those that fueled the SaaS revolution—platforms like GitHub, Stripe, and Twilio—to support physical AI. "We genuinely all think that anyone building an autonomous system for the real world is going to do so in software primarily in two to three years," Mellsop said. "It’s the first time you can have autonomous agents iterate on a physical autonomy system, and actually close the feedback loop."

Experiments are already underway in this direction. David Mayo, a researcher at MIT’s Computer Science and Artificial Intelligence Laboratory, is using Antioch’s platform to evaluate large language models (LLMs). In one experiment, Mayo has AI models design robots, then tests them in Antioch’s simulator. The platform can even stage simulated contests between models, such as pushing a rival bot off a platform. Providing LLMs with a realistic sandbox could establish a new paradigm for benchmarking their capabilities.

However, before a world of AI engineers becomes a reality, more work is needed to bridge the gap between digital models and the physical world. If this can be achieved, developers will be able to create a data flywheel—a concept Macneil identifies as crucial to the success of industry leaders like Waymo, where engineers grow increasingly confident that each successive model will outperform the last. For other companies aspiring to replicate such success, they will need to either develop these tools themselves or acquire them.