Ann Arbor-based Laplace Intelligence is automating stress tests for physical AI: an AI agent hunts for scenarios in which a robot or self-driving car fails. The startup showed it on 13 October at Startup Battlefield 200 during TechCrunch Disrupt.
Key takeaways
- The simulation environment is a post-trained NVIDIA Cosmos, grounded in real-world data
- AI agents design corner cases, drawing on news and police collision reports
- Other road users react to the test vehicle instead of letting it drive through them
- Founded a year ago out of the University of Michigan, the company is modestly profitable
- Customers include Isuzu, Bot Auto and NVIDIA, plus a smart city pilot in San Jose with T-Mobile
A simulation reconstructed from the real world
At the core sits a high-fidelity simulation reconstructed from the real world — a post-trained?post-training: Further training of a finished model on your own data, rather than training from scratch. Here: tuning the simulator so it sticks to real geometry. NVIDIA Cosmos that can rebuild anything from an Arizona highway to a Paris pedestrian crossing.
The scene is not random: a Google Street View image serves as reference, and map data fills in the geometry, so the scale matches reality. Traffic data comes from navigation providers such as TomTom, and the setting changes with a prompt in real time.
Not just the setting, but how others behave
The behaviour of other road users is simulated too. During a demo Sun steered the test vehicle into oncoming traffic — the car it would have hit moved out of the way. In many generative world models the vehicle would simply have driven through it.
It's truly an interactive simulation environment. Everyone is reacting to everything. This is more like a world model with interactive participants.
Howie Sun, CEO and co-founder of Laplace Intelligence.
The real product is the agent, not the scene
The differentiator is not the environment but agents that design corner cases themselves. They draw on news and police collision reports, pushing the human further out of the loop.
Today most self-driving developers turn recorded footage into simulations or rely on engineers. The method rests on Sun and Liu's University of Michigan paper on dense reinforcement learning, published in Nature in 2023.
| Dimension | Autonomy developers today | Laplace Intelligence |
|---|---|---|
| Source of corner cases | road footage, engineers' ideas | an AI agent reading news and police reports |
| Other road users | often static, the car drives through them | react to the vehicle under test |
| Environment | reconstructed from footage | post-trained Cosmos plus map data |
The diagram arranges the elements described by the company in TechCrunch's report. The ordering is editorial — the piece does not say how many scenarios pass through the loop, or how case difficulty is measured.
The main rival is Applied Intuition, valued at $15 billion and serving the same customer, Isuzu.
Why it matters
Automated red-teaming is then the only method that scales faster than collecting accidents. The company's ambition goes beyond a testing product: a shared body of safety evidence for physical AI that customers, insurers and regulators can trust. Without one, every robot deployment in public space comes down to trusting the manufacturer.
What's next
- The company, NVIDIA and T-Mobile are running a free smart city pilot for San Jose, California
- Talks with partners building warehouse humanoids are early-stage
- Contracts are project-based today — the goal is recurring revenue
