If the incident between OpenAI and Hugging Face taught us anything, it’s that the behavior of AI agents can be unpredictable.
"Imagine what would happen if something similar occurred in the real world," said Howie Sun, CEO and co-founder of Laplace Intelligence, to TechCrunch. "AI agents are moving from computers into the real world, and we don't want physical AI to malfunction in the real world."
Laplace, headquartered in Ann Arbor, is building automated red teaming tests for robots and physical AI—essentially stress-testing systems by deliberately causing them to fail. The company bets that the only way to ensure the safety of physical AI is to have AI test itself.
Laplace is one of the 20 startups competing in the Startup Battlefield, a three-day pitch competition at TechCrunch Disrupt. The company will present on the Startup Battlefield stage on Monday, October 13.
Sun and co-founder and chief scientist Henry Liu founded Laplace a year ago after meeting at the University of Michigan. Over the past eight years, Sun has conducted academic research on challenging autonomous driving AI in simulation environments. In 2023, he and Liu, who was then his PhD advisor, published a related paper that appeared on the cover of the premier scientific journal Nature. Following this, autonomous driving companies began reaching out to them to implement these ideas.
“I realized I wanted to turn this into something real—to take it from pure academic research into a product that’s genuinely useful in the real world,” Sun said.
Sun and Liu founded Laplace based on their academic research. The company’s name derives from “Laplace’s Demon,” a thought experiment proposed in 1814 by French scholar Pierre-Simon Laplace, which imagines that if one knew the precise state of everything in the universe, one could predict the future with complete accuracy.
“We are essentially using simulations to predict the future of the real world,” said Sun.
At the core of this startup is a high-fidelity simulation environment reconstructed from the real world, paired with an AI agent that continuously probes for weaknesses and failure points. The goal is to ensure that any physical AI—whether robots, autonomous vehicles, or smart city solutions—is verified for safety in simulation before being deployed in the real world.
This simulation environment is post-trained using NVIDIA's Cosmos, an AI world simulation platform connected to the real world, enabling Laplace to reconstruct scenarios ranging from Arizona highways to Parisian crosswalks. The simulation is generated by integrating multiple real-world data sources. Google Street View images serve as references to recreate entire streets, but the system does not generate content randomly—it reconstructs environments at accurate real-world scales by combining real-world map data.
Developers can then use language prompts to dynamically alter the scene, such as adding afternoon traffic. Sun said that Laplace will obtain real-time traffic and congestion data from navigation providers like TomTom.
In Laplace’s simulation, not only is the environment modeled, but the behaviors of other participants are also simulated. For example, during one demonstration, Sun instructed the test autonomous vehicle to turn into oncoming traffic. The vehicle that would have been collided with—or, in some generative world models, simply passed through—moved out of the way to avoid the collision.
“This is truly an interactive simulation environment,” Sun said. “Everyone is reacting to everything… It’s more like a world model with interactive participants.”
Although the simulation environment is a key component of Laplace’s products, the company’s real selling point is the AI agents it is developing to automatically design edge cases—challenging, rare scenarios that could cause AI systems to fail—thereby enabling further human withdrawal from the process.
“They can design scenarios from news reports and police reports about collisions—various types of information can be used to actively challenge autonomous vehicles and other physical AI agents,” said Sun.
Today, autonomous driving developers mostly train AI to handle edge cases by recording videos, converting those videos into simulations, or relying on humans to generate new ideas and design them. Although many autonomous driving companies are already doing what Laplace describes and beginning to embrace using AI to create the edge cases needed for training—Waabi, an autonomous trucking company, is one example—this technology’s demand will become truly apparent when we start deploying general-purpose robots into the real world.
"General-purpose robots have less available data," said Sun. "No one knows what the edge cases in the real world will be."
Although Sun and Liu come from academia and lack glamorous Silicon Valley tech company backgrounds, they have already begun generating revenue and even achieved modest profitability. Their income primarily comes from autonomous driving-related applications, with clients including Japanese truck manufacturer Isuzu and Texas-based autonomous trucking company Bot Auto.
Sun also noted that Nvidia is a customer. Together with T-Mobile, the two companies are collaborating on a free smart city pilot project in San Jose to provide high-level traffic operations support. Laplace is also in early-stage collaboration with partners developing humanoid robots for warehouse scenarios.
Currently, most of Laplace's client contracts are project-based, and the company is working to transition them to recurring revenue. The startup’s near-term goal—which it hopes to achieve with the help of the Disrupt competition—is to secure funding and scale up to compete with rivals.
Its main competitor is Applied Intuition, a highly watched late-stage startup that has raised $1.2 billion to date, with a post-money valuation of $15 billion. Isuzu is also one of Applied’s customers.
Sun isn’t overly concerned. He believes that Laplace’s core approach of using AI to test AI represents a “completely new paradigm for physical AI testing,” transforming testing into a “dynamic, interactive process.”
“That would be the widest moat,” he said.