This Startup Crash-Tests Robots and Self-Driving Cars Before They Hit the Road

A new startup called Laplace Intelligence wants to make sure robots and self-driving cars are safe before they ever meet you on the street. The Ann Arbor company builds detailed computer simulations of real places — a highway in Arizona, a busy crosswalk in Paris — and then uses AI to deliberately try to crash or confuse the vehicles inside them. Think of it like a flight simulator for robots, but the simulator is constantly inventing tricky situations to see what breaks. The company was founded a year ago by Howie Sun and Henry Liu, whose earlier research on testing self-driving AI made the cover of the journal Nature in 2023.

Why should you care? Self-driving cars and delivery robots are already on some roads, and they will only become more common. If they fail, people can get hurt. Laplace's simulations let companies find and fix problems before a real accident happens. The simulations are interactive: if one car swerves, other cars and pedestrians react, just like in real life. The AI can even design new "corner cases" — rare, dangerous scenarios — by reading news reports and police collision data, so humans don't have to think of every possible bad thing.

There is a catch. Simulations are never perfect. A virtual world can't capture every weird thing that happens on a real street, like a sudden hailstorm or a person in a chicken costume. And right now, Laplace is still a small startup competing against bigger companies. But its bet is simple: the only way to make physical AI safe is to have AI test it first. If it works, you might never notice — except that the robots around you will be a lot less likely to make mistakes.