Laplace Intelligence, an Ann Arbor-based startup founded by Howie Sun and Henry Liu, is commercializing academic research on stress-testing autonomous systems. The company creates interactive simulations of real-world environments—reconstructed from map data, street imagery, and live traffic feeds—then deploys AI agents to deliberately provoke failures in robots, self-driving cars, and other physical AI. Sun argues that as AI moves from computers into the physical world, companies need to discover and fix problems before deployment, not after.

The core technology builds on a 2023 Nature cover paper by Sun and Liu, his former PhD advisor, on challenging autonomous driving AI in simulation. Laplace's simulation environment is a post-trained version of Nvidia's Cosmos, grounded in real-world data so that a Google Street View image can generate an entire street. Unlike some generative world models, the simulation is interactive: other actors react to the test vehicle's actions, such as swerving into oncoming traffic. The startup's differentiator is automating the design of corner cases, drawing on sources like news reports and police collision records to generate tricky scenarios without heavy human input.

Laplace has already generated revenue and a modest profit, mostly from autonomous driving projects. Customers include Japanese trucking manufacturer Isuzu, Texas-based autonomous trucking firm Bot Auto, and Nvidia itself. The company is also in the early stages of working with humanoid robotics partners for warehouses. Its near-term goal is to raise funds and scale, competing against Applied Intuition, a late-stage startup valued at $15 billion that also counts Isuzu as a customer. Sun is presenting at TechCrunch Disrupt's Startup Battlefield on October 13 to help attract investment. The source notes that while some AV companies already use AI to create edge cases, the need for this kind of testing will grow as generalized robots enter the real world, where edge cases are still largely unknown. No conflicting sources were provided.}```markdown? Wait, the user asked