If there’s one lesson the OpenAI-Hugging Face incident taught us, it’s that AI agents can be unpredictable.

“Imagine if something like that happened in the physical world,” Howie Sun, CEO and co-founder of Laplace Intelligence, told TechCrunch. “AI agents are moving from computers into the real world, and we don’t really want physical AI to fail in the real world.”Ann Arbor-based Laplace is building automated red-teaming — essentially, stress-testing a system by deliberately trying to make it fail — for robotics and physical AI. The company is betting that the only way to make sure physical AI is safe is to have AI test it.

“Our goal is to help companies discover and fix problems before deployment—and, over time, build a shared foundation of evidence that customers, insurers, and regulators can trust,” Sun said.

Laplace is one of 20 startups competing in Startup Battlefield, TechCrunch Disrupt’s three-day pitch competition in downtown San Francisco. It will present on the Startup Battlefield stage Monday afternoon, October 13.

Sun, alongside his co-founder and chief scientist Henry Liu, founded Laplace a year ago out of the University of Michigan. Sun has spent the last eight years in academia researching how to challenge autonomous driving AI in simulation. In 2023, he and Liu — his then-PhD advisor — published a paper on the subject that made it onto the cover of the top scientific journal Nature. That’s when the calls started coming in from autonomous vehicle companies interested in implementing their ideas.

“I realized that I wanted to build this and take it from purely academic research into a product that’s useful in the real world,” Sun said.

Using their academic research as a base, Sun and Liu built Laplace, so named for “Laplace’s demon,” an 1814 thought experiment by French scholar Pierre-Simon Laplace that imagined that if you knew the exact state of everything in the universe, you could predict the future with total precision. “We are essentially using simulation to predict the real world’s future,” Sun said.

At the startup’s core is a high-fidelity simulation reconstructed from the real world, paired with an AI agent that constantly probes for weaknesses and failure points. The goal is to ensure that any physical AI — be it a robot, autonomous vehicle, or smart city solution — is safe in simulation before it’s deployed in the real world.

The simulation environment is a post-trained version of Nvidia’s Cosmos, an AI world simulation platform, that’s grounded in the real world, allowing Laplace to reconstruct anything from an Arizona highway to a Paris pedestrian crossing. Simulations are generated by blending a mesh of different real-world information. A Google Street View image could serve as a reference from which an entire street is generated. But rather than generating something random, the simulator combines real-world map data to recreate an environment that’s true to scale.

Developers can then use language prompts to change the setting in real time, like adding in afternoon traffic. Sun says Laplace pulls real-time traffic and congestion information from navigation providers like TomTom.

Within Laplace’s simulations, not only the environment is simulated, but also the behavior of other actors. For example, during a demo, Sun directed the test agent vehicle to swerve into oncoming traffic. The car it would have hit — or driven through, as is common among some generative world models — moved out of its way to avoid a collision.

“It’s truly an interactive simulation environment,” Sun said. “Everyone is reacting to everything. …This is more like a world model with interactive participants.”While the simulated environment is an important part of Laplace’s product, the real unique selling point is the AI agents the startup is developing to automatically design corner cases (tricky, rare scenarios that could trip up an AI system) moving the human further out of the loop.

“They can design from the news, from police reports on collisions — all kinds of information to actively challenge the autonomous vehicles and other physical AI agents,” Sun said.

Today, autonomous vehicle developers mostly train an AI to handle edge cases using recorded footage that they then turn into a simulation, or by relying on humans to come up with new ideas and design them. While many AV companies have done what Laplace is pitching and started embracing AI to create edge cases for training — self-driving truck company Waabi comes to mind — the need for this kind of technology will really become apparent when we start placing generalized robots in the real world.

“There’s less data on generalized robotics,” Sun said. “No one knows what the edge cases will be in the real world.”Despite Sun and Liu coming from academia, with no shiny Silicon Valley tech company pedigrees, the two have managed to start generating revenue and even turn a modest profit. Most of the money is coming from autonomous driving use cases, with customers including Japanese trucking manufacturer Isuzu and Texas-based autonomous trucking firm Bot Auto.

Nvidia is also a customer, says Sun. The two companies, alongside T-Mobile, are working together on a free smart city pilot for the City of San Jose to provide high-level traffic operations. Laplace is also in the early stages of working with partners building humanoid robotics for warehouses.

Most of Laplace’s customer contracts are project-based today, and the company is working towards turning that into recurring revenue. The startup’s near-term goal — which it hopes will be aided by competing at Disrupt — is to raise funds and scale so it can compete with its rivals.

Chief among its rivals is Applied Intuition, a buzzy late-stage startup that’s raised $1.2 billion to date at a $15 billion post-money valuation. Isuzu is also one Applied’s customers.

Sun isn’t too worried. He believes Laplace’s core methodology of using AI to test AI is a “new paradigm that changes the testing of physical AI entirely,” turning it into a “dynamic interactive testing process.”“That will be the biggest moat,” he said.