Before autonomous vehicles and robots enter the physical world, Laplace Intelligence wants to challenge them in a simulated one. The Ann Arbor-based startup combines reconstructed environments with AI agents that probe for weaknesses, aiming to surface problems early in development.
Testing physical AI in a simulated world
Laplace’s approach is automated red-teaming: deliberately trying to make a system fail so developers can find and address vulnerabilities. The company applies this method to robotics and other forms of physical AI, including autonomous vehicles.
Its platform pairs a high-fidelity simulation with an AI agent that searches for failure points. The simulation is based on real-world information and can represent settings ranging from an Arizona highway to a Paris pedestrian crossing. A Google Street View image can help guide the recreation, while map data helps keep the generated environment true to scale.
Developers can use language prompts to alter a scene as it runs, such as adding afternoon traffic. Laplace also draws on real-time traffic and congestion information from navigation providers like TomTom, according to CEO and co-founder Howie Sun.
Scenarios that respond to one another
The simulated world includes the behavior of other participants, not just the background environment. In a demonstration, Sun directed a test vehicle to swerve into oncoming traffic. The oncoming car moved out of its way, responding to the potential collision.
That interaction matters because the startup aims to model situations where multiple actors react to each other. Laplace’s simulations are intended to behave more like an interactive world than a static test scene, allowing developers to examine how an AI system responds as circumstances change.
The company’s distinctive focus, Sun says, is using AI agents to automatically design corner cases: rare or difficult scenarios that could expose a system’s weaknesses. Those agents can draw on information such as news reports and police collision reports to create challenges for autonomous vehicles and other physical AI systems.
From autonomous driving to general-purpose robots
Autonomous vehicle developers have typically used recorded footage to build edge-case simulations, or relied on people to imagine and design new scenarios. Some companies have begun using AI to generate edge cases for training. Laplace sees a larger need ahead as more generalized robots enter real-world settings, where there is less data and the difficult scenarios are not yet well understood.
Sun and co-founder and chief scientist Henry Liu started Laplace a year ago out of the University of Michigan. Sun had spent eight years researching ways to challenge autonomous driving AI in simulation. In 2023, he and Liu published a paper on the subject that appeared on the cover of Nature. Interest from autonomous vehicle companies helped prompt Sun to turn the research into a product.
The name Laplace refers to “Laplace’s demon,” an 1814 thought experiment by French scholar Pierre-Simon Laplace. It imagines that knowing the exact state of everything in the universe would make it possible to predict the future with total precision. The startup applies that idea through simulation, seeking to anticipate how physical AI could behave in the real world.
Customers, pilots, and the next stage
Laplace says it has begun generating revenue and has even turned a modest profit. Most of its income comes from autonomous driving work. Its customers include Texas-based autonomous trucking firm Bot Auto and a major Japanese automotive manufacturer.
The startup is also working with Nvidia and T-Mobile on a free smart city pilot in San Jose, California, focused on high-level traffic operations. Separately, it is in early discussions with partners building humanoid robots for warehouses.
Most customer agreements are project-based today, and Laplace is working to develop recurring revenue. Its near-term plan is to raise funds and scale up. The company is one of 20 startups competing in Startup Battlefield 200 at TechCrunch Disrupt in San Francisco, where it is scheduled to present on October 13.
Laplace faces competition from Applied Intuition, which the source describes as a late-stage startup that has raised $1.2 billion at a $15 billion post-money valuation. Sun argues that using AI to test AI could change physical AI testing into a dynamic, interactive process. For Laplace, the test is whether simulated challenges can help developers find and fix failures before a system encounters them outside the simulation.