Laplace Intelligence Tests Robots by Trying to Break Them
Laplace Intelligence is building automated red-teaming for robotics and physical AI, a test harness that tries to make systems fail before deployment. 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 core product pairs two components. The first is a high-fidelity simulation reconstructed from the real world, in effect a detailed virtual copy. The second is an AI agent that operates inside that simulation to probe for weaknesses and failure points. Failure is the objective. The agent searches for inputs, interactions and edge conditions, meaning unusual or extreme situations, that break the policy under test, rather than confirming normal behavior.
The simulation environment is a post-trained version of Nvidia's Cosmos, grounded in the real world, meaning further trained to reflect real places and conditions. Developers can use language prompts to alter the simulation setting in real time, changing the scenario around the robot or vehicle without rebuilding the scene. The system also pulls real-time 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 large language model evaluation. Red-teaming began as human-led adversarial testing, then became increasingly automated for software and models. Physical AI is now following the same path, with the added constraint that failures carry kinetic consequences, meaning they can cause physical harm. Simulation becomes the only place to run thousands of adversarial iterations safely and cheaply.
In my view, the interesting technical bet is the combination of grounding and steerability. A reconstructed, data-fed world addresses the staleness problem that limits hand-built scenarios, which quickly go out of date. Prompt-based control addresses the coverage problem that limits log replay, which can only retest what was already recorded. For robotics teams, the practical question is always how to generate the rare, high-severity case that was not in the training or validation set. An agent tasked with causing failure, operating in a world that can be edited in natural language and refreshed with live traffic state, is a direct attempt to automate that search.
Looking at what this means for development workflows, the shift is from accumulating nominal miles or hours to directing compute toward failure discovery. If adversarial scenario generation can be made reliable and physically plausible, it changes where engineering time goes. Less time scripting test cases. More time fixing the weaknesses the agent surfaces. That pattern has held through prior transitions in software testing, and there is reason to think embodied systems will benefit in the same way, with safer behavior in warehouses, on roads and around people as the eventual result.
