Laplace Intelligence Automates Red-Teaming for Physical AI
Ann Arbor-based startup Laplace Intelligence is developing automated red-teaming for robots, autonomous vehicles and smart-city systems. The company uses AI agents and high-fidelity simulation to probe physical AI for weaknesses before it is deployed in real environments.
Laplace was founded a year ago by CEO Howie Sun and chief scientist Henry Liu, who came out of the University of Michigan. Sun spent eight years in academia researching ways to challenge autonomous-driving AI in simulation, and the pair published related work in Nature in 2023.
Interactive simulations for failure testing
The startup’s platform combines a real-world reconstruction with an AI agent designed to search continually for failure points. Its aim is to test robots, self-driving vehicles and smart-city applications in simulation before they operate on roads, in warehouses or elsewhere in the physical world.
Laplace uses a post-trained version of Nvidia Cosmos, an AI world-simulation platform. The company says the environment is grounded in real-world information: a Google Street View image can serve as a reference, while map data helps recreate streets at the correct scale. The system can model settings ranging from an Arizona highway to a Paris pedestrian crossing.
Developers can alter conditions with language prompts, such as adding afternoon traffic. Sun said Laplace also obtains real-time traffic and congestion information from navigation providers including TomTom. The platform simulates not only the setting but the behaviour of other participants, so vehicles and other actors can react during a scenario rather than simply appearing as static elements.
AI agents create uncommon edge cases
The central product focus is on agents that automatically design corner cases: rare or difficult situations that can expose a system’s limitations. Sun said these agents can draw on material such as news reports and police collision reports to actively challenge autonomous vehicles and other physical AI agents.
Autonomous-vehicle developers commonly use recorded footage transformed into simulations, or have people devise new testing scenarios. Laplace is attempting to move more of that scenario design into automated agents. Sun argues that this is particularly relevant to general-purpose robotics, where there is less real-world operating data and the relevant edge cases are less established.
Early customers and a smart-city pilot
Laplace says it is generating revenue and has reached a modest profit. Its autonomous-driving customers include Japanese trucking manufacturer Isuzu and Texas autonomous trucking company Bot Auto. Nvidia is also a customer, Sun said.
Laplace, Nvidia and T-Mobile are working on a free smart-city pilot with the City of San Jose for high-level traffic operations. The startup is also in early-stage work with partners developing humanoid warehouse robots. Most contracts are currently project-based, while Laplace works toward recurring revenue.
The company is among 20 startups selected for Startup Battlefield at TechCrunch Disrupt and was scheduled to present on October 13. For businesses deploying physical AI, the practical implication is to treat adversarial simulation and evidence of behaviour under rare conditions as part of pre-deployment validation.