Laplace Intelligence puts robots and self-driving cars through penetration (red-team) tests before they enter the real world. The Ann Arbor-based startup, spun out of the University of Michigan, develops automated “red-teaming” tests for physical AI systems, deliberately causing them to fail and exposing their weak points before they reach the real world. The goal is to find weaknesses early. The company is one of 20 startups competing in the Startup Battlefield 200 competition at TechCrunch Disrupt, held in San Francisco on October 13-15, and will present on the Disrupt stage on the afternoon of Tuesday, October 13.
CEO and co-founder Howie Sun said the OpenAI-Hugging Face incident showed that AI agents can be unpredictable, and that such a failure would be undesirable in the physical world. He spent eight years in academia. During this time, Sun, who stress-tested autonomous driving AI in simulation, published in 2023 the paper he wrote with Henry Liu, his doctoral advisor at the time, and the work made the cover of Nature magazine. They then began attracting interest from autonomous vehicle companies. The pair founded Laplace a year ago. The company’s name comes from Pierre-Simon Laplace’s 1814 “Laplace’s demon” thought experiment.
At the core of Laplace is a high-fidelity simulation rebuilt from the real world, together with an AI agent that constantly searches for weaknesses. The system reconstructs reality. The simulation environment runs as a post-trained version of Nvidia’s Cosmos world simulation platform. Any environment can be recreated, from an Arizona highway to a pedestrian crossing in Paris. The system uses Google Street View imagery as a reference. Environments true to scale are produced using real map data. Developers can make changes in real time with language commands, such as afternoon traffic. The company pulls real-time traffic data from navigation providers like TomTom. Other vehicles also react. In one demo, when the test vehicle veered into the oncoming lane, the other vehicle avoided the collision.
Its main difference is that it creates agents that automatically design difficult and rare scenarios (corner cases). These agents draw on news and police accident reports. They compile scenarios from different sources. Today, auton
Why it matters
One of the hardest questions for teams developing autonomous vehicles and robots is knowing how the system will respond to rare accident scenarios without trying them on real roads; Laplace’s approach seeks to fill this gap by automatically generating adverse scenarios in simulation, compiled from police reports and news, meaning safety testing is no longer limited by the imagination of the person designing the test. The founders’ work appearing on the cover of Nature demonstrates the technology’s academic foundation, while the interest coming from autonomous vehicle companies shows that it corresponds to a commercial need. The reference to the OpenAI-Hugging Face incident reflects the concern that if the unpredictability in software agents is carried over to physical systems, the risk will grow exponentially. The question that remains open is to what extent the weaknesses uncovered in simulation can represent safety under real road conditions.
Background
Intelligence is not a new name in the FikirPilot archive: we have published 6 news items featuring this name in the last 90 days; the most recent is dated 6 October 2026.
Term: agent
An AI agent is software that, instead of producing a single response, calls tools and carries out multi-step work to achieve a goal.