Ann Arbor-based Laplace Intelligence is building automated red-teaming for robotics and physical AI, aiming to ensure these systems are safe before they hit the real world. The company’s CEO, Howie Sun, emphasizes the importance of stress-testing AI to prevent failures, inspired by the unpredictability of OpenAI-Hugging Face incidents.

The core of Laplace’s product is a high-fidelity simulation paired with an AI agent that constantly probes for weaknesses. This simulates various real-world environments and behaviors, from traffic to human actions, ensuring that any physical AI is safe in simulation before deployment.

Laplace uses Nvidia’s Cosmos, a real-world-grounded AI world simulation platform, to reconstruct environments like Arizona highways or Parisian pedestrian crossings. The simulation environment is highly interactive, with AI agents that can be directed to handle edge cases, such as swerving into oncoming traffic, ensuring the system reacts appropriately.

While Laplace’s technology is in demand, particularly for autonomous vehicle developers, the company is also working on projects for smart cities and humanoid robotics. The need for this kind of technology is expected to increase as generalized robots are deployed in the real world, where data is less abundant.

Despite its academic roots, Laplace has managed to generate revenue and turn a profit, with notable customers including Bot Auto and a Japanese automotive manufacturer. The startup is now working towards scaling and competing with rivals, with a focus on raising funds to achieve this goal.