Laplace Intelligence uses AI to crash-test robots before reality does

A one-year-old startup spun out of the University of Michigan is already profitable selling something the autonomous vehicle industry mostly still does by hand: finding the rare, dangerous scenarios that make physical AI fail.

Laplace Intelligence, based in Ann Arbor, builds automated red-teaming for robotics and self-driving systems — essentially, using an AI agent to deliberately try to break another AI inside a high-fidelity simulation before it ever touches a real road or warehouse floor. The company counts trucking manufacturer Isuzu, autonomous trucking firm Bot Auto, Nvidia, and a smart city pilot with San Jose among its customers.

It will pitch on the Startup Battlefield stage at TechCrunch Disrupt in downtown San Francisco on Monday afternoon, October 13, as one of 20 competing startups.

Why does anyone need AI to test AI?

CEO and co-founder Howie Sun points to a recent lesson from the software world. "Imagine if something like that happened in the physical world," Sun told TechCrunch, referring to the OpenAI-Hugging Face incident as proof that AI agents can behave unpredictably. "AI agents are moving from computers into the real world, and we don't really want physical AI to fail in the real world."

The stakes go beyond engineering. Sun framed the company's mission in terms of accountability: "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."

That last word matters. As robots and autonomous vehicles move into public space, insurers and regulators will need evidence of safety that goes beyond a company's own testing claims.

What's under the hood?

Laplace's platform pairs two pieces. The first is a simulation environment rebuilt from real-world data — a post-trained version of Nvidia's Cosmos world-simulation platform, grounded with map and sensor information. A Google Street View image can serve as a reference from which the simulator generates an entire street, but real map data keeps the recreation true to scale. It can reconstruct anything from an Arizona highway to a Paris pedestrian crossing, and developers can change conditions with language prompts — say, adding afternoon traffic, using real-time congestion data pulled from navigation providers like TomTom.

The second piece, and according to Sun the real differentiator, is the AI agent that designs corner cases — tricky, rare scenarios that could trip up an AI system — with minimal human involvement.

"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.

The simulation is also interactive. During a demo, Sun directed a test vehicle to swerve into oncoming traffic, and the oncoming car — rather than being driven through, as happens with some generative world models — moved aside to avoid the 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."

The name itself signals the ambition: "Laplace" comes from "Laplace's demon," the 1814 thought experiment by French scholar Pierre-Simon Laplace proposing that knowing the exact state of everything in the universe would let you predict the future perfectly. "We are essentially using simulation to predict the real world's future," Sun said.

Who's behind it, and who's paying?

Sun spent eight years in academia researching how to challenge autonomous driving AI in simulation. In 2023, he and Henry Liu — his then-PhD advisor, now co-founder and chief scientist — published a paper on the subject that landed on the cover of Nature. That's when autonomous vehicle companies started calling, interested in implementing the 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.

The customers followed. Isuzu and Texas-based Bot Auto anchor the autonomous driving revenue, which makes up most of Laplace's income — enough that the startup has turned a modest profit. Nvidia is also a customer, and together with T-Mobile the two are running a free smart city pilot for the City of San Jose providing high-level traffic operations. Laplace is in early-stage work with partners building warehouse humanoid robots as well.

Today, most AV developers handle edge cases by replaying recorded footage in simulation or asking humans to invent new scenarios. Some, like self-driving truck company Waabi, already use AI to generate edge cases for training. Sun argues the real demand arrives with generalized robotics: "There's less data on generalized robotics. No one knows what the edge cases will be in the real world."

Can it survive a $15 billion rival?

Laplace's chief competitor is Applied Intuition, which has raised $1.2 billion at a $15 billion post-money valuation — and counts Isuzu among its own customers too.

The company's contracts are mostly project-based today, and it's working to convert them into recurring revenue. Its near-term plan is to raise funds and scale, a push it hopes competing at Disrupt will support.

Sun isn't rattled by the size gap. He calls Laplace's method of using AI to test AI a "new paradigm that changes the testing of physical AI entirely," turning safety validation into a "dynamic interactive testing process."

"That will be the biggest moat," he said.

By Sophie Lindqvist