Laplace Intelligence, an Ann Arbor-based startup, is applying automated red-teaming to physical AI systems — robots, self-driving cars, and smart city solutions — by deliberately trying to make them fail in high-fidelity simulation before they reach the real world. Co-founder and CEO Howie Sun frames the problem in terms of AI agents becoming less predictable as they move from computers into the physical world, citing an OpenAI-Hugging Face incident as a reminder of what can go wrong. The company is one of 20 startups in the Startup Battlefield competition at TechCrunch Disrupt in San Francisco, with its stage presentation scheduled for Monday afternoon, October 13. Sun says the goal is to help companies discover and fix problems before deployment and build a shared foundation of evidence that customers, insurers, and regulators can trust.

Sun and chief scientist Henry Liu founded Laplace about a year ago out of the University of Michigan. Sun spent eight years in academia researching how to challenge autonomous driving AI in simulation; in 2023, he and Liu — then his PhD advisor — published a paper on the subject that made the cover of Nature, and AV companies began contacting them. The startup is named after Laplace’s demon, an 1814 thought experiment by French scholar Pierre-Simon Laplace imagining that perfect knowledge of the universe’s state would allow perfect prediction of the future. Sun says Laplace is essentially using simulation to predict the real world’s future.

The core product combines a post-trained version of Nvidia’s Cosmos world simulation platform, grounded in the real world, with AI agents that continuously probe for weaknesses and failure points. Laplace can reconstruct anything from an Arizona highway to a Paris pedestrian crossing, using sources such as Google Street View and real-world map data to stay true to scale. Developers can use language prompts to change the setting in real time — for example adding afternoon traffic using real-time congestion information from navigation providers like TomTom. The simulation also models the behavior of other actors: during a demo, Sun directed the test vehicle to swerve into oncoming traffic, and the car it would have hit moved out of the way. It is, Sun says, truly an interactive simulation environment — more of a world model with interactive participants than a passive generative world.

Laplace’s main differentiator is the AI agents that automatically design corner cases — rare, tricky scenarios that could trip up an AI system — moving humans further out of the loop. Sun says they can design tests from news reports, police reports on collisions, and other information to actively challenge autonomous vehicles and other physical AI. Today, AV developers mostly use recorded footage converted into simulation or rely on humans to design new scenarios; Waabi is another company pursuing AI-generated edge cases. Sun argues the need becomes sharper with generalized robots, because there is less data and no one knows what the real-world edge cases will be.

Despite its academic origins, the startup says it has generated revenue and even a modest profit, mostly from autonomous-driving use cases. Customers include Japanese trucking manufacturer Isuzu, Texas-based autonomous trucking firm Bot Auto, and Nvidia. Laplace is also part of a free smart city pilot for the City of San Jose with Nvidia and T-Mobile, and is in early work with partners building humanoid warehouse robots. Most customer contracts are project-based today, and the company is working toward recurring revenue. Its near-term goal, which it hopes to advance through Disrupt, is to raise funds and scale against rivals such as Applied Intuition, a late-stage startup with $1.2 billion raised and a $15 billion post-money valuation; Isuzu is also one of Applied’s customers. Sun says Laplace’s core methodology of using AI to test AI will change physical AI testing into a dynamic, interactive process, and calls that its biggest moat.