Laplace Intelligence is developing AI agents that create rare, challenging situations to see whether robots and self-driving cars can cope with real-world conditions. The agents can use material such as news coverage and police collision reports to build scenarios, then test AI systems in interactive simulations before they are deployed.
Autonomous vehicle testing today leans on recorded footage and human-designed scenarios. Developers generally turn recorded footage into simulations or ask people to come up with and design edge cases. Laplace, based in Ann Arbor, is building automated red-teaming for robotics and physical AI. The process deliberately pushes a system toward failure so developers can identify and address problems before deployment.
“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,” Laplace CEO and co-founder Howie Sun told TechCrunch. The company is betting that the only way to make sure physical AI is safe is to have AI test it. Sun said the company wants to help businesses find and fix problems before deployment and, over time, build a shared body of evidence that customers, insurers, and regulators can trust.
Sun said AI agents may act unpredictably once they move out of computers and into the physical world. Citing the OpenAI-Hugging Face incident, he 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.”
Sun and Henry Liu, Laplace’s co-founder and chief scientist, started the company at the University of Michigan a year before the article was written. Before that, Sun spent eight years in academia studying how to challenge autonomous-driving AI in simulations. In 2023, he and Liu, who was then Sun’s PhD adviser, published a paper on the topic that appeared on the cover of the scientific journal Nature. Sun said autonomous-vehicle companies began contacting them after the paper came out to discuss applying the research.
“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 name Laplace refers to “Laplace’s demon,” an 1814 thought experiment from French scholar Pierre-Simon Laplace. In it, a person with exact knowledge of everything in the universe could predict the future with complete precision. “We are essentially using simulation to predict the real world’s future,” Sun said.
The company’s simulation environment is based on a post-trained version of Nvidia’s Cosmos, an AI world simulation platform grounded in real-world data. By combining different kinds of real-world information, it can recreate places from an Arizona highway to a Paris pedestrian crossing. For instance, a Google Street View image can guide the creation of a whole street. The simulator also uses real-world map data to keep the recreated setting to scale instead of generating a random scene. Laplace’s goal is to ensure that physical AI – including robots, autonomous vehicles, and smart-city solutions – is safe in simulation before it’s deployed in the real world.
Language prompts let developers alter a simulated scene as it runs, for example by adding afternoon traffic. Sun said Laplace also gets live traffic and congestion data from navigation providers such as TomTom.
Laplace simulates how other actors behave, not just the environment around them. In one demonstration, Sun told a 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.”
Laplace’s real unique selling point is its agents, which are being developed to generate corner cases automatically. Those are rare, tricky situations that can expose weaknesses in an AI system. The agents are intended to move humans further out of the loop.
Sun said this matters most for generalized robotics, where there’s less data on how systems will behave. “There’s less data on generalized robotics,” he said. “No one knows what the edge cases will be in the real world.” Laplace has begun working with partners developing humanoid robots for warehouses.
Although Sun and Liu came from academia rather than established Silicon Valley tech companies, Laplace has started bringing in revenue and has turned a modest profit. Most of its income comes from autonomous-driving projects. Customers include Japanese trucking manufacturer Isuzu and Texas-based autonomous trucking firm Bot Auto. Sun said Nvidia is a customer as well.
Laplace, Nvidia, and T-Mobile are collaborating on a free smart-city pilot in San Jose, California, to provide high-level traffic operations. The company’s customer agreements are mostly project-based for now, and it is working to shift toward recurring revenue.
Laplace is among 20 startups in Startup Battlefield 200, the pitch competition held at TechCrunch Disrupt in San Francisco from October 13-15. It is set to present on the Disrupt Stage on Tuesday afternoon, October 13. The startup’s near-term goal is to raise funds and scale so it can compete with rivals, and Sun hopes competing will help.
One competitor is Applied Intuition, which has raised $1.2 billion at a post-money valuation of $15 billion. Isuzu is also a customer of Applied Intuition. Sun said he is not especially concerned about the competition. He believes Laplace’s use of AI to test AI is a “new paradigm that changes the testing of physical AI entirely,” making testing a “dynamic interactive testing process.”
“That will be the biggest moat,” Sun said.
