Laplace Intelligence red-teams robots and self-driving cars before they hit the real world

The recent volatility seen in AI agent deployments has served as a sobering wake-up call for the tech industry. As these systems transition from controlled digital environments into the unpredictable physical world, the stakes for failure have shifted from software bugs to potential real-world catastrophes.

Howie Sun, CEO and co-founder of Laplace Intelligence, puts it bluntly: “Imagine if something like that happened in the physical world. AI agents are moving from computers into the real world, and we don’t really want physical AI to fail in the real world.”

Based in Ann Arbor, Laplace Intelligence is tackling this challenge head-on by pioneering automated "red-teaming"—a process of stress-testing systems by intentionally engineering scenarios designed to force a failure. The startup is currently one of the 20 companies competing in the Startup Battlefield 200 at TechCrunch Disrupt in San Francisco, where they are showcasing how AI can be the ultimate safeguard for its own kind.

From Academic Research to Real-World Safety

Laplace Intelligence emerged from the University of Michigan just one year ago, born from the collaboration between Sun and his former PhD adviser, Henry Liu. Their partnership is rooted in deep academic rigor; in 2023, the duo published a seminal paper on challenging autonomous driving AI in simulation, which earned a prestigious cover feature in the scientific journal Nature.

The transition from academia to entrepreneurship was driven by demand. Following the publication of their research, autonomous vehicle (AV) companies began reaching out, seeking to implement their methodologies.

“I realized that I wanted to build this and take it from purely academic research into a product that’s useful in the real world,” Sun explained.

The company’s name is a nod to "Laplace’s Demon," an 1814 thought experiment by Pierre-Simon Laplace, which posited that if one knew the precise state of every atom in the universe, the future could be predicted with absolute certainty. While the startup doesn't claim omniscience, they are using high-fidelity simulation to bring a new level of predictive precision to physical AI.

The Mechanics of a Digital Twin

At the heart of the Laplace platform is a sophisticated, high-fidelity simulation engine. By leveraging a post-trained version of Nvidia’s Cosmos AI world simulation platform, Laplace can reconstruct real-world environments—from busy Parisian intersections to sprawling Arizona highways—with remarkable accuracy.

The platform doesn't just generate static environments; it creates interactive, scalable worlds. By blending real-world map data with visual references like Google Street View, the system builds environments that are true to scale. Developers can then manipulate these environments in real time using natural language prompts, such as adjusting traffic density or weather conditions.

Crucially, the simulation is populated by interactive agents. If a test vehicle is commanded to swerve into traffic, the surrounding simulated cars react dynamically to avoid a collision, rather than simply clipping through the test vehicle as seen in less advanced models.

Automating the "Edge Case" Hunt

While the simulation environment is impressive, the true value proposition lies in Laplace’s proprietary AI agents. These agents are designed to autonomously hunt for "edge cases"—the rare, tricky, or dangerous scenarios that typically trip up AI systems.

By removing the human from the loop, Laplace’s agents can scan vast amounts of data, including news reports and police collision records, to synthesize complex scenarios that challenge autonomous vehicles and robotics.

Key Advantages of the Laplace Approach:

• Dynamic Stress-Testing: Moving beyond static, pre-recorded footage to active, interactive testing.
• Reduced Human Dependency: AI-driven scenario generation identifies risks that human engineers might overlook.
• Broad Applicability: While currently focused on AVs, the technology is highly relevant for the emerging field of generalized robotics, where data on real-world edge cases is notoriously scarce.

A Profitable Path in a Competitive Market

Despite being a young startup without the typical Silicon Valley pedigree, Laplace Intelligence has already achieved a rare milestone: profitability. Their revenue is currently driven by autonomous driving contracts, with a client list that includes Japanese trucking giant Isuzu, Texas-based Bot Auto, and Nvidia.

Furthermore, the company is collaborating with T-Mobile and Nvidia on a smart city pilot in San Jose, California, aimed at optimizing high-level traffic operations. They are also expanding into the warehouse robotics sector, working with partners to ensure humanoid robots can operate safely in complex, human-centric environments.

The Road Ahead

Laplace is entering a market dominated by well-funded incumbents, most notably Applied Intuition, which has secured $1.2 billion in funding at a $15 billion valuation. However, Sun remains undeterred. He views Laplace’s methodology—using AI to test AI—as a fundamental paradigm shift.

“That will be the biggest moat,” Sun said, emphasizing that the transition to a dynamic, interactive testing process is the only way to build the foundation of evidence required by insurers, regulators, and the public.

As Laplace Intelligence presents its vision at TechCrunch Disrupt, the company is looking to scale its operations and transition from project-based contracts to a recurring revenue model. By turning the unpredictability of AI against itself, Laplace is betting that the safest way to build the future is to simulate every possible way it could go wrong.