Laplace Intelligence Uses AI Agents to Stress-Test Robotics and Self-Driving Cars

The OpenAI-Hugging Face incident highlighted that AI agents can be unpredictable. Imagine if such unpredictability occurred in the physical world, Sun said. AI agents are transitioning from computers to the real world, and we want to avoid physical AI failures in the real world.

Based in Ann Arbor, Laplace is building automated red-teaming, or stress-testing, for robotics and physical AI. They believe the only way to ensure physical AI safety is to have AI test it. Our goal is to help companies discover and fix problems before deployment, building a shared foundation of evidence for customers, insurers, and regulators, Sun stated.

Sun, along with co-founder and chief scientist Henry Liu, founded Laplace a year ago at the University of Michigan. Sun spent eight years researching how to challenge autonomous driving AI in simulation. In 2023, they published a paper that made the cover of *Nature*, leading to interest from autonomous vehicle companies. Sun realized he wanted to transition this research from academia to a real-world product.

Laplace is named after Laplace's demon, an 1814 thought experiment about predicting the future. We use simulation to predict the future of the real world, Sun explained. The core product is a high-fidelity simulation reconstructed from the real world, paired with an AI agent probing for weaknesses. The goal is to ensure any physical AI is safe in simulation before deployment.

The simulation uses a post-trained version of Nvidia's Cosmos, grounded in reality to recreate environments like Arizona highways. It blends real-world data, such as map data and Google Street View images, to create true-to-scale environments. Developers use language prompts to alter settings in real-time, such as adding traffic congestion, by pulling real-time data from providers like TomTom.

The simulations model not just the environment but the behavior of other actors. In a demo, when the test agent swerved into traffic, the other car moved to avoid a collision. It is an interactive simulation where everyone reacts to everything, functioning like a world model with participants, Sun noted.

The unique selling point is the AI agents that automatically design corner cases. These agents can proactively challenge autonomous vehicles using data from news reports and police collision records. While some companies like Waabi are adopting similar methods, the need for this technology will increase with generalized robots, which have less training data.

Despite their academic backgrounds, Sun and Liu have generated revenue, including from Isuzu and Bot Auto. Nvidia, alongside T-Mobile, is working with Laplace on a San Jose smart city pilot project. The startup is moving from project-based contracts to recurring revenue. Competitor Applied Intuition has raised significantly more funding.

Sun isn't concerned, believing the AI-testing approach is a new paradigm that changes the testing of physical AI entirely. That will be the biggest moat, he said.