If there’s one lesson the OpenAI-Hugging Face incident has taught us, it’s that AI agents can be unpredictable.
“Imagine if something like that happened in the physical world,” Howie Sun, CEO and co-founder of Laplace Intelligence, told TechCrunch. “AI agents are moving from computers to the real world, and we don’t want physical AI to fail in the real world.”
Ann Arbor-based Laplace builds automated red teams (essentially stress tests that intentionally try to cause a system to fail) for robotics and physical AI. The company is betting that the only way to ensure that physical AI is secure is to have it tested.
“Our goal is to help companies find and fix problems before deployment, and over time build a shared foundation of evidence that customers, insurers and regulators can trust,” Sun said.
Laplace is one of 20 startups participating in the Startup Battlefield 200, a startup pitch competition being held this week from October 13th to 15th at TechCrunch Disrupt in San Francisco. It will be presented on the Disrupt Stage on Tuesday afternoon, October 13th.
Sun founded Laplace a year ago from the University of Michigan with co-founder and chief scientist Henry Liu. Sun has spent the past eight years in academia researching ways to challenge self-driving AI through simulation. In 2023, he and Liu, who was his PhD advisor at the time, published a paper on the topic that was featured on the cover of the top scientific journal Nature. That’s when he started getting calls from self-driving car companies interested in making his idea a reality.
“We wanted to build this and take it from purely academic research into a product that could be useful in the real world,” Sun says.
Sun and Liu built Laplace on the basis of academic research. Laplace was named after the 1814 thought experiment by French scholar Pierre-Simon Laplace, “Laplace’s Demon.” Laplace imagined that if we knew the exact state of everything in the universe, we could predict the future with perfect accuracy.
“We’re basically using simulation to predict the future of the real world,” Sun said.
At the core of the startup is a high-fidelity simulation reconstructed from the real world, paired with an AI agent that constantly probes for weaknesses and points of failure. The goal is to ensure that physical AI, such as robots, self-driving cars, and smart city solutions, is safe in simulations before being deployed in the real world.
The simulation environment is a post-training version of Nvidia’s Cosmos, a real-world-based AI world simulation platform that allows Laplace to reconstruct everything from Arizona highways to Paris crosswalks. Simulations are generated by blending different meshes of real-world information. Google Street View images serve as a reference for generating the entire street. However, rather than randomly generating something, the simulator combines real-world map data to recreate a true-to-scale environment.
Developers can use language prompts to change settings in real time, such as adding afternoon traffic. Sun said Laplace gets real-time traffic and congestion information from navigation providers such as TomTom.
Within a Laplace simulation, not only the environment but also the behavior of other actors is simulated. For example, during the demonstration, Sun instructed the test agent vehicle to swerve into oncoming traffic. As is common in some generative world models, the cars that would have collided with or passed through moved out of the way to avoid the collision.
“This is truly an interactive simulation environment,” Sun says. “Everyone is reacting to everything…This is more like a world model with interactive participants.”
While simulated environments are a key part of Laplace’s product, the real unique selling point is the AI agent the startup is developing to automatically design corner cases (tricky and rare scenarios that can trip up AI systems) and keep humans further out of the loop.
“They can design all kinds of information to actively challenge self-driving cars and other physical AI agents, such as news and police reports about collisions,” Sun said.
Currently, most self-driving car developers use recorded footage to train AI to handle edge cases and turn it into simulations, or rely on humans to come up with new ideas and design. While many AV companies are starting to do what Laplace suggests and employ AI to create edge cases for training—self-driving truck company Waabi comes to mind—the need for this kind of technology will really become apparent as we begin to deploy generalized robots in the real world.
“There is less data on generalized robotics,” Sun said. “No one knows what will happen in the real world.”
Even though Sun and Liu come from academia and don’t have the pedigree of Silicon Valley’s shiny tech companies, they’ve managed to start making money and even make a small profit. Most of the funding is coming from self-driving use cases, with customers including Japanese trucking manufacturer Isuzu and Texas-based self-driving truck company Bot Auto.
NVIDIA is also a customer, Sun said. The companies are collaborating with T-Mobile on a free smart city pilot in San Jose, California, to provide a high level of transportation operations. Laplace is also in the early stages of working with partners to build humanoid robots for warehouses.
Currently, most of Laplace’s customer contracts are project-based, and the company is working to turn them into recurring revenue. The startup’s short-term goals, which it hopes will be helped by competition at Disrupt, are to raise funding and scale to compete with competitors.
Its leading competitor is Applied Intuition, a high-profile late-stage startup that has raised $1.2 billion to date at a post-money valuation of $15 billion. Isuzu is also one of Applied’s customers.
San isn’t too worried. He believes Laplace’s core methodology of using AI to test AI is “a new paradigm that will completely change the testing of physical AI” and transforms AI into a “dynamic, interactive testing process.”
“That’s going to be the biggest moat,” he said.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
