Physical AI systems—from delivery bots to autonomous vehicles—face unique challenges once they leave the lab. Laplace Intelligence, a startup spun out of the University of Michigan, is tackling those risks head-on. Their platform automates red-teaming: stress-testing robotic and vehicular AI systems via simulated rare or dangerous scenarios before they ever hit the streets. Founded just over a year ago by AI researchers Howie Sun and Henry Liu, Laplace believes this type of safety testing should be the norm—not the exception.
Simulating the Real World, Interacting with Agents
At its foundation is a high-fidelity simulator based on a modified version of Nvidia’s Cosmos, pulling in real-world map, image, and traffic data to build realistic environments—whether that’s a Paris crosswalk or an Arizona highway. Using tools like Google Street View and live congestion feeds from navigation services, the platform recreates true-to-scale surroundings. What sets it apart is its interactive behavior: agents in the simulation react to one another. For example, if a test vehicle veers into oncoming traffic, other simulated actors steer to avoid collision—bringing a whole new realism to edge-case modeling.
AI That Finds the Unseen Errors
Rather than relying primarily on human testers or existing footage, Laplace deploys AI agents that scour reports—police records, news stories, collision data—and then design so-called corner cases. These are rare or unexpected scenarios that tend to cause failure in autonomous systems, but human designers might not foresee them. By automating this discovery process, Laplace aims to push physical AI systems into safer territory before they’re ever exposed to risk.
Early Traction, Real Customers
Though the founders come from academia, their operation is already generating revenue and small profits. Their early customers include Isuzu and Bot Auto, both in the autonomous trucking sphere, as well as Nvidia and partners in a smart city pilot with San Jose. Laplace is also engaging with companies building warehouse robotics. Today, contracts tend to be one-off projects; the startup is working toward subscription or recurring models, which would help scale their business sustainably.
Standing Out vs. an Established Field
The red-teaming safety space is warming up. Applied Intuition has raised over $1.2 billion and holds a valuation around $15 billion, and counts some of the same customers as Laplace—like Isuzu. But Laplace claims its strength lies in its methodology: using AI to test AI, in interactive, dynamicallychanging simulation environments. The startup believes this paradigm gives it a defensible edge—or moat—over incumbents that use more static or human-led approaches.
Laplace has also appeared at TechCrunch Disrupt’s Startup Battlefield competition in San Francisco on October 13, 2026, where it will further pitch its vision to investors and industry watchers.
Why this matters: As physical AI systems proliferate in transportation, logistics, and public infrastructure, undetected failure modes could carry serious consequences. Laplace’s approach offers a scalable way to uncover potential hazards early. Key to watch: how well their simulations align with real-world outcomes, whether customers embrace AI-led red-teaming over traditional testing, and how they manage to convert project work into steady recurring business.
