How This Startup Breaks Robots on Purpose to Keep Us Safe

Laplace Intelligence is building automatic tests that try to break robots and self-driving cars before they are used in real life. The Ann Arbor company is one of 20 startups competing in Startup Battlefield at TechCrunch Disrupt, and was scheduled to present on the Battlefield stage on Monday afternoon, October 13, 2026. TechCrunch
The company was founded out of the University of Michigan about a year before October 2026 by Howie Sun and Henry Liu. Sun serves as CEO and co-founder. Liu serves as co-founder and chief scientist.
Its product has two parts. The first is a lifelike virtual world copied from the real world. The second is an AI tester that works inside that world to find weak spots. Causing failure is the goal. The tester looks for inputs, encounters and rare situations that break the robot's software, instead of just checking normal behavior.
The virtual world is a version of Nvidia's Cosmos that was trained further on real-world data. Developers can type plain written commands to change the scene in real time, without rebuilding it. The system also uses live traffic and congestion information from navigation providers like TomTom. The company name refers to Laplace's demon, the 1814 thought experiment by Pierre-Simon Laplace.
The broader context here will be familiar to anyone who has worked in security or in testing large language models. Red-teaming started as people trying to break systems by hand, then became more automatic for software and models. Physical AI is now following the same path, except failures can cause real physical harm. Simulation is the only place to run thousands of these hostile tests safely and cheaply.
In my view, the key idea is combining a realistic world with easy control. A world copied from real data avoids the problem of hand-made tests going stale. Typed commands avoid the problem of only replaying past trips, which misses new dangers. Robotics teams always face the same hard question of how to find the rare but serious case that was not in training or testing data. A tester whose job is to cause failure, in a world that can be rewritten in words and updated with live traffic, is a direct way to automate that search.
Looking at what this means for how robots are built, the shift is from collecting normal miles or hours to using computers to hunt for failures. If this kind of hostile test creation can be made dependable and lifelike, engineers will spend less time writing tests by hand and more time fixing the flaws the tester finds. We have seen this shift before in software testing. The payoff should be safer machines in warehouses, on roads and around people.


