Laplace Intelligence Develops AI-Driven Red Teaming to Stress-Test Robots and Autonomous Vehicles

Image: Howie Sun, CEO, Laplace Intelligence.Image Credits:Laplace Intelligence · Source

Laplace Intelligence, spun out from the University of Michigan, uses AI agents within high-fidelity real-world simulations to proactively identify failure points in robotics and self-driving cars before deployment, aiming to enhance safety and reliability.

As AI systems increasingly move from digital environments into physical applications like robotics and autonomous vehicles, ensuring their reliability and safety becomes paramount. Laplace Intelligence, a startup founded in 2025 by University of Michigan researchers Howie Sun and Henry Liu, is addressing this challenge by employing AI-driven red teaming within interactive, high-fidelity simulations. Their approach aims to expose potential failures before such systems enter the real world.

Laplace’s technology combines detailed simulations grounded in real-world data with autonomous AI agents tasked with probing and stress-testing physical AI systems. The simulations are constructed using Nvidia’s Cosmos platform, allowing accurate reconstruction of various environments—from highways in Arizona to pedestrian crossings in Paris—using geographic and visual data such as Google Street View imagery. Users can modify scenarios dynamically, incorporating real-time conditions like traffic levels sourced from providers like TomTom.

What sets Laplace apart is its focus on AI agents that actively generate rare and challenging scenarios, known as corner cases, to test autonomous systems. These cases can derive from diverse inputs, including news events and police collision reports, pushing physical AI to handle unforeseen or difficult situations. This represents a shift from traditional testing methods relying on human-designed scenarios or recorded footage.

The company has found initial traction primarily in autonomous driving. Its clientele includes Japanese truck manufacturer Isuzu, autonomous trucking firm Bot Auto in Texas, and Nvidia, alongside whom Laplace is collaborating with T-Mobile on a smart city pilot project in San Jose, California. Early engagements also extend to humanoid robotics used in warehouses. Although still early stage, Laplace has begun generating revenue and operating profitably.

Laplace’s co-founder Sun emphasizes their methodology as a new paradigm in testing physical AI, transforming it into a dynamic, interactive process that can better predict and prevent failures. Their work builds on academic research previously published in the journal Nature, highlighting the rigorous scientific foundation behind the startup’s product.

With competitors like Applied Intuition commanding substantial funding and market presence, Laplace aims to scale up through upcoming funding rounds and greater market penetration. Participation in competitions such as TechCrunch Disrupt’s Startup Battlefield helps raise the startup’s profile as it endeavors to become a trusted provider of rigorous AI safety testing for robotics and autonomous vehicles.

In summary, Laplace Intelligence leverages AI to test AI by creating realistic, adaptive simulations that expose vulnerabilities prior to real-world deployment. This focus on preemptive failure discovery reflects the growing imperative to ensure physical AI systems operate safely and reliably as they become more widespread.