If the OpenAI‑Hugging Face clash taught anything, it was that artificial‑intelligence agents can behave in unexpected ways. “Imagine if something like that happened in the physical world,” said Howie Sun, CEO and co‑founder of Laplace Intelligence, speaking to 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.”
Based in Ann Arbor, Laplace is creating an automated red‑team platform that deliberately tries to break robotic and other physical‑AI systems. The startup believes that the safest way to verify physical AI is to let AI itself probe for weaknesses. “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 explained.
Laplace is one of twenty companies competing in Startup Battlefield 200, the pitch competition at TechCrunch Disrupt in San Francisco from October 13‑15. The firm will take the Disrupt Stage on Tuesday afternoon, October 13.
Sun and chief scientist Henry Liu launched Laplace a year ago after meeting at 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 adviser, published a paper on the topic that appeared on the cover of Nature, sparking interest from autonomous‑vehicle firms. “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 company’s name references “Laplace’s demon,” an 1814 thought experiment by Pierre‑Simon Laplace that suggested perfect knowledge of the universe would allow exact prediction of the future. “We are essentially using simulation to predict the real world’s future,” Sun noted.
At the heart of Laplace’s offering is a high‑fidelity simulation reconstructed from real‑world data, paired with an AI agent that continuously searches for failure points. The aim is to verify that any physical AI—whether a robot, autonomous vehicle, or smart‑city system—is safe in a virtual environment before it is released.
The simulation runs on a post‑trained version of Nvidia’s Cosmos platform, which is anchored in real‑world measurements and can recreate settings ranging from an Arizona highway to a Paris pedestrian crossing. Real‑world map data and sources such as Google Street View are blended to generate accurate, to‑scale environments. Developers can modify the scene on the fly with language prompts, for example by adding afternoon traffic. Sun said Laplace pulls live traffic and congestion data from navigation providers like TomTom.
Within these virtual worlds, not only the surroundings but also the behavior of other actors is modeled. In a demonstration, Sun instructed a test vehicle to veer into oncoming traffic; the simulated car that would have been struck moved aside to avoid a collision, illustrating an “interactive simulation environment” where “everyone is reacting to everything… This is more like a world model with interactive participants,” he said.
While the realistic environment is a key component, Laplace’s distinctive advantage lies in the AI agents that automatically generate corner cases—rare, tricky scenarios that can trip up an AI system—thereby reducing the need for human designers. “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 explained.
Today, most autonomous‑vehicle developers train their systems on edge cases derived from recorded footage turned into simulations, or rely on humans to invent new scenarios. Some companies, such as self‑driving truck maker Waabi, have begun using AI to create edge cases, but the demand for such tools will grow as generalized robots enter everyday settings. “There’s less data on generalized robotics,” Sun observed. “No one knows what the edge cases will be in the real world.”
Despite lacking Silicon Valley pedigrees, Sun and Liu have begun generating revenue and even posted a modest profit. Their primary income comes from autonomous‑driving projects, with clients that include Texas‑based autonomous trucking firm Bot Auto and a major Japanese automotive manufacturer. Laplace is also collaborating with Nvidia and T‑Mobile on a free smart‑city pilot for San Jose, California, to improve high‑level traffic operations, and is in early talks with partners building humanoid warehouse robots.
Most of Laplace’s contracts are currently project‑based, but the company aims to shift toward recurring revenue. Its immediate objective—bolstered by participation at Disrupt—is to raise additional capital and scale up to compete with rivals such as Applied Intuition.