AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 13, 2026

Laplace Intelligence, an Ann Arbor-based startup spun out of the University of Michigan, is building automated red-teaming for robotics and physical AI — stress-testing a system by deliberately trying to make it fail, before it operates in the real world. The company is one of 20 startups competing in Startup Battlefield 200 at TechCrunch Disrupt in San Francisco, which runs October 13-15; Laplace is scheduled to present on the Disrupt Stage on the afternoon of Tuesday, October 13.

CEO and co-founder Howie Sun founded the company a year ago with co-founder and chief scientist Henry Liu, his former PhD adviser. Sun spent eight years in academia researching how to challenge autonomous driving AI in simulation. In 2023, the pair published a paper on the subject that appeared on the cover of Nature, which Sun says prompted autonomous vehicle companies to reach out about implementing their ideas. The company takes its name from "Laplace's demon," an 1814 thought experiment by Pierre-Simon Laplace imagining that knowing the exact state of everything in the universe would allow total precision in predicting the future. "We are essentially using simulation to predict the real world's future," Sun said. His stated goal is to help companies find and fix problems before deployment and, over time, build a shared foundation of evidence customers, insurers and regulators can trust.

At the core of the product is a high-fidelity simulation reconstructed from the real world, paired with an AI agent that continuously probes for weaknesses and failure points. The simulation environment is a post-trained version of Nvidia's Cosmos, an AI world simulation platform, grounded in real-world data. It can reconstruct settings ranging from an Arizona highway to a Paris pedestrian crossing by blending a mesh of real-world information — for example, using a Google Street View image as a reference from which an entire street is generated, combined with real-world map data so the recreated environment is true to scale rather than random. Developers can change the setting in real time with language prompts, such as adding afternoon traffic; Sun says Laplace pulls real-time traffic and congestion information from navigation providers like TomTom.

Inside those simulations, other actors are simulated too, not just the environment. In one demo, Sun directed the test agent vehicle to swerve into oncoming traffic, and the car it would have hit — or passed through, as Sun notes is common with some generative world models — moved out of the way to avoid a collision. He described this as a genuinely interactive environment in which everyone reacts to everything, closer to a world model with interactive participants than a static scene.

But the company's stated differentiator is the AI agents it is developing to automatically design corner cases — the tricky, rare scenarios that trip up an AI system — moving humans further out of the loop. According to Sun, those agents can design tests from news, police collision reports and other information sources to actively challenge autonomous vehicles and other physical AI. This matters most, in his view, for generalized robots, where data is scarce and no one yet knows what the real-world edge cases will be.

Today, autonomous vehicle developers largely handle edge cases by turning recorded footage into simulations or relying on humans to invent new scenarios. Several AV companies have begun using AI to generate edge cases for training; TechCrunch points to self-driving truck company Waabi as an example. Laplace's bet is that AI testing AI becomes a new paradigm that turns physical AI validation into a dynamic, interactive process — what Sun calls the company's biggest moat.

Commercially, Laplace is already generating revenue and turning a modest profit despite its founders' academic rather than Silicon Valley pedigrees, according to the report. Most revenue comes from autonomous driving use cases, with customers including Japanese trucking manufacturer Isuzu and Texas-based autonomous trucking firm Bot Auto. Nvidia is also a customer, per Sun. The company, Nvidia and T-Mobile are working together on a free smart city pilot for San Jose, California, aimed at high-level traffic operations. Laplace is also in early-stage work with partners building humanoid robots for warehouses.

Most customer contracts today are project-based, and the company is working to convert that into recurring revenue. Its near-term goal — which it hopes competing at Disrupt will support — is to raise funds and scale to compete with rivals. The most prominent is Applied Intuition, a late-stage startup that has raised $1.2 billion to date at a $15 billion post-money valuation, and which counts Isuzu as a customer as well. Sun says he is not overly worried, arguing the core methodology of using AI to test AI is a new paradigm that changes physical AI testing entirely.

For developers and product teams, the interesting structural point is where simulation fidelity ends and adversarial generation begins. Reconstruction from map data and imagery plus real-time traffic feeds gives the environment scale and plausibility; the AI agents supply the adversarial intent, turning validation from a curated scenario library into a search for failure modes. That reframes validation work: rather than hand-authoring scenarios, teams would define objectives, constraints and coverage criteria, then review what the agents surface. A likely trade-off is trust — automatically generated corner cases are only useful if they are physically and semantically valid, otherwise teams spend time triaging artifacts rather than real defects. Sun's answer is that interactivity is the filter: simulated actors that react to the tested system's behavior are harder to game than largely static generative scenes.

Why it matters: as robots and autonomous vehicles move from labs and pilot fleets toward wider deployment, validation becomes the gating factor, and the evidence trail behind it is what insurers and regulators will scrutinize. Teams building physical AI may soon need tooling that searches for failures automatically instead of waiting for humans to imagine them — and should expect project-based engagements to shift toward recurring validation contracts. Whether that market consolidates around incumbents like Applied Intuition or splits between training-focused and testing-focused vendors is an open question, but the competitive ground is clearly moving toward adversarial, simulation-based assurance rather than footage replay.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

University of Michigan spinout Laplace Intelligence uses AI agents and interactive simulation to stress-test robots and self-driving vehicles before deployment. The profitable startup is competing in Startup Battlefield at TechCrunch Disrupt, October 13-15 in San Francisco.