The vibe check for physical AI
Ever seen those AI agents glitch out or hallucinate on your screen? Now, imagine that happening to a robot walking down a sidewalk or an autonomous truck merging onto a highway. It’s giving… disaster. Howie Sun, CEO and co-founder of Laplace Intelligence, is basically trying to prevent those real-world fails before they happen.
Laplace is building automated "red-teaming" for physical AI and robotics. In tech speak, red-teaming means stress-testing a system by trying to break it. The goal is to see if the AI can handle the chaos of the real world—like sudden traffic or unexpected pedestrian moves—in a safe, simulated environment before it’s unleashed on the public.
How it works
Laplace’s tech is built on a high-fidelity simulation, reconstructed from real-world data like Google Street View images and mapping data from TomTom. It’s not just a flat image; it’s a dynamic world model where everything reacts to everything else. If you tell a simulated car to swerve into traffic, the other cars actually move to avoid it.
Sun and his co-founder/chief scientist Henry Liu (who was Sun’s PhD advisor) founded the company a year ago out of the University of Michigan. They aren’t just throwing stuff at a wall; they literally wrote the book on challenging autonomous driving AI in a paper published in Nature back in 2023.
The “Ok, but why should you care?”
Right now, most autonomous vehicle companies train their AI using recorded footage or by having humans dream up risky scenarios. But for the new generation of generalized warehouse robots, we literally have no idea what the "edge cases" (the rare, tricky stuff) will be. Laplace uses AI to design these scenarios automatically, pulling from things like news reports and accident data, which helps developers get way ahead of potential bugs.
They’ve already snagged customers like Isuzu, Bot Auto, and even Nvidia. They’re also working with T-Mobile on a smart city pilot for San Jose. They’re competing in the Startup Battlefield at TechCrunch Disrupt this week, looking to scale up and take on the big dogs like Applied Intuition.
Why it matters
As robots start moving out of controlled labs and into our grocery stores and city streets, we need a way to trust them. Laplace’s methodology—using AI to test AI—could become the industry standard for insurance and regulation, making sure that when robots finally enter our space, they don't cause any unnecessary chaos. Lowkey, they’re the unsung heroes of the robot uprising.
