Picture a self-driving truck on an Arizona highway at rush hour. A car ahead brakes hard, a cyclist cuts across two lanes, and the sun drops right into the cameras. Engineers might never think to script that exact mix. Real roads will throw it at the truck sooner or later anyway.
That gap is what Laplace Intelligence wants to close. TechCrunch AI reports that the Ann Arbor startup is building automated red-teaming for robots and physical AI. Red-teaming means stress-testing a system by deliberately trying to make it fail. Laplace’s twist is that an AI does the attacking, so failures turn up in simulation before they happen on a real street.
“AI agents are moving from computers into the real world, and we don’t really want physical AI to fail in the real world,” CEO and co-founder Howie Sun told TechCrunch. Laplace is one of 20 startups in Startup Battlefield 200 at TechCrunch Disrupt in San Francisco this week.
🔬 From a Nature Cover to a Product
Sun and his co-founder and chief scientist Henry Liu, who was once his PhD adviser, started Laplace a year ago out of the University of Michigan. Sun spent eight years in academia researching how to challenge autonomous driving AI in simulation. In 2023, the pair published a paper on the topic that made the cover of Nature.
After that, autonomous vehicle companies started calling. Sun decided to turn the research into a product people could actually use.
The name comes from “Laplace’s demon,” an 1814 thought experiment: if you knew the exact state of everything in the universe, you could predict the future perfectly. “We are essentially using simulation to predict the real world’s future,” Sun said.
⚙️ How the Platform Works
The product has two main parts: a realistic simulated world and an AI agent that keeps hunting for weak spots in it. Here’s what’s under the hood:
- Built on Nvidia Cosmos: The simulation runs on a post-trained version of Nvidia’s world simulation platform. It can rebuild anything from an Arizona highway to a Paris pedestrian crossing.
- Grounded in real data: A single Google Street View image can seed a whole street. The system then combines it with map data so the scene stays true to scale.
- Language prompts: Developers can change the scene in real time by typing something like “add afternoon traffic.”
- Live traffic feeds: Laplace pulls real-time congestion data from navigation providers like TomTom.
- Reactive actors: Other cars and pedestrians respond to what’s happening. In one demo, Sun steered the test vehicle into oncoming traffic, and the other car swerved out of the way instead of passing straight through it, which is a common glitch in generative world models.
- AI-designed corner cases: This is the main selling point. Agents create tricky, rare scenarios based on news stories, police collision reports, and other sources, so humans don’t have to dream them all up.
🚗 Why It Matters
Today, AV developers mostly build edge cases from recorded footage or rely on engineers to imagine them. Laplace isn’t alone in using AI for this job. Self-driving truck company Waabi already does something similar.
The bigger opportunity is general-purpose robots. Self-driving has years of road data behind it. Humanoid robots working in warehouses have almost none. “No one knows what the edge cases will be in the real world,” Sun said. When nobody knows where the failures are, an AI built to find them becomes very useful.
I think the more interesting long-term bet is Sun’s goal of building “a shared foundation of evidence that customers, insurers, and regulators can trust.” If simulation-based safety evidence becomes something insurers and regulators accept, the company that produces it gains a lot of influence in how physical AI gets deployed.
💼 Customers and Business Model
Without big Silicon Valley names behind it, Laplace is already making money and even turning a small profit, according to TechCrunch AI. Its customers and partners include:
- Isuzu, the Japanese truck maker
- Bot Auto, a Texas autonomous trucking company
- Nvidia, which is also a customer
- A free smart city pilot in San Jose with Nvidia and T-Mobile, focused on high-level traffic operations
- Early-stage work with humanoid robotics partners building for warehouses
The article doesn’t give pricing. Most contracts are project-based for now, and Laplace wants to move toward recurring revenue.
⚠️ The Caveats
Laplace is small and competing in a crowded market. Its biggest rival is Applied Intuition, a late-stage startup that has raised $1.2 billion at a $15 billion valuation. Most of Laplace’s revenue still comes from autonomous driving, so its robotics pitch is mostly unproven. It’s also trying to raise money to scale, and Disrupt is part of that push.
The idea is solid: you test physical AI the same way you test software, by attacking it on purpose before it ships. Whether a two-person academic spinout can beat a $15 billion incumbent is the real test. You can find more details in the original TechCrunch AI report.