TechCrunch reports on Laplace Intelligence, an Ann Arbor startup founded by Howie Sun and Henry Liu out of the University of Michigan that red-teams physical AI before deployment. The company reconstructs real-world environments in simulation, then uses AI agents to deliberately try to make autonomous vehicles and robots fail. “Our goal is to help companies discover and fix problems before deployment,” CEO Howie Sun told TechCrunch.
Laplace's simulation layer is a post-trained version of Nvidia's Cosmos, grounded in real-world data so it can recreate settings from an Arizona highway to a Paris pedestrian crossing. Developers can adjust scenarios with language prompts, and simulated actors respond to each other in real time. The startup's core differentiator is its automated corner-case generation: AI agents draw on news, police reports, and other inputs to design rare, tricky scenarios that could trip up a physical AI system.
Sun spent eight years in academia testing autonomous driving AI in simulation, and he and Liu published a Nature cover paper in 2023 that attracted interest from AV companies. Laplace now generates revenue and a modest profit, with customers including Texas-based autonomous trucking firm Bot Auto and a major Japanese carmaker. It is also working with Nvidia and T-Mobile on a smart-city pilot in San Jose. The company is competing at TechCrunch Disrupt's Startup Battlefield 200, where it will present on October 13. Its main rival is Applied Intuition, but Sun argues that using AI to test AI represents a new paradigm for physical AI safety.