If there is any lesson to be learned from the incident between OpenAI and Hugging Face, it is that the behavior of AI agents can be difficult to predict.

“Imagine what would happen if something similar occurred in the real world,” said Laplace Intelligence, CEO and co-founder, to TechCrunch. “Virtual assistants are moving from computers to the real world, and we certainly don’t want the physical components to fail in that reality.”

Laplace, headquartered in Ann Arbor, is building automated red-team tests for robots and physical AI – essentially, stress testing by deliberately causing the systems to fail. The company believes that the only way to ensure the security of physical AI is to let AI test it.

Laplace is one of the 20 startups in the TechCrunch Disrupt three-day roadshow competition Startup Battlefield. The company will make its presentation on stage at Startup Battlefield on Monday, October 13th, in the afternoon.

Sun and co-founder and chief scientist Henry Liu founded Laplace a year ago from the University of Michigan. Over the past eight years, Sun has been researching how to challenge autonomous driving in simulation environments within the academic community. In 2023, he and Liu, who was then his doctoral supervisor, published a related paper that appeared on the cover of the top scientific journal Nature. Subsequently, autonomous driving companies began to contact them, hoping to bring these ideas to fruition.

“I realized that I wanted to make this a reality, to transform it from pure academic research into a product that is truly useful in the real world,” said Sun.

Sun and Liu founded Laplace based on their academic research. The company name is derived from the "Laplace's Demon" – a thought experiment proposed in 1814 by the French scholar Pierre - Simon Laplace, which posits that if one knew the precise state of everything in the universe, one could predict the future with complete accuracy.

"Essentially, what we're doing is using simulations to predict the future of the real world," said Sun.

At the core of this startup is a high-fidelity simulation environment reconstructed from the real world, coupled with a AI agent that continuously detects weaknesses and failure points. The goal is to ensure that any physical AI – whether it's a robot, an autonomous vehicle, or a smart city solution – can be verified for safety in simulation before being deployed into the real world.

This simulation environment is retrained based on Cosmos of Nvidia. Cosmos is a AI world simulation platform that is connected to the real world, allowing Laplace to reconstruct various scenarios ranging from Arizona highways to Paris crosswalks. The simulation is generated by integrating multiple types of real-world information. Google Street View images can be used as a reference to generate an entire street. However, the system does not generate randomly; instead, it reconstructs the environment in true proportion by combining real-world map data.

Developers can then use language prompts to change the scenario in real-time, such as adding afternoon traffic conditions. Sun indicates that Laplace will obtain real-time traffic and congestion information from navigation service providers like TomTom.

In the simulation of Laplace, not only is the environment simulated, but also the behavior of other participants is mimicked. For example, during a demonstration, Sun instructed the test proxy vehicle to turn towards oncoming traffic. The vehicle that would have been collided with – or in some generative world models, it would have simply passed through without any impact – moved out of the way to avoid the collision.

“This is truly an interactive simulation environment,” said Sun. “Everyone is reacting to everything… It’s more like a world model with interactive participants.”

Although the simulation environment is an important component of the Laplace product, the real selling point of this company lies in their AI agents that they are developing. These agents can automatically design edge cases—those tricky and rare scenarios that could cause errors in the AI system—thereby allowing humans to step back further from the process.

"They can design scenarios from news and police reports about collisions – all kinds of information can be used to proactively challenge autonomous vehicles and other physical AI agents," said Sun.

Today, most autonomous driving developers train for edge cases by recording videos, then convert these videos into simulations, or rely on humans to come up with new ideas and designs. Although many autonomous driving companies are already doing what is described in Laplace and have begun to embrace the use of AI to create the edge cases needed for training – autonomous truck company Waabi is one example – the demand for such technologies will truly become evident when we start integrating general-purpose robots into the real world.

"General-purpose robots have even less data available," said Sun. "No one knows what the edge cases in the real world will be."

Although Sun and Liu come from academia and do not have impressive resumes from Silicon Valley tech companies, they have already started generating income and have even achieved small profits. Their revenue mainly comes from autonomous driving applications, with clients including Japanese truck manufacturer Isuzu and the autonomous truck company Bot Auto based in Texas.

Sun also indicates that Nvidia is a customer as well. Together with T-Mobile, these two companies are undertaking a free smart city pilot project in San Jose to provide advanced transportation operation support. Laplace has also begun early-stage collaboration with partners developing humanoid robots for warehousing applications.

Currently, for Laplace, the majority of customer contracts are still based on project-based models. The company is working hard to convert these into recurring revenues. The short-term goal of this startup—which hopes to gain assistance through the Disrupt competition—is to raise funds and expand its scale in order to compete with its rivals.

The main competitor among them is Applied Intuition. This highly regarded late-stage startup has so far raised $1.2 billion in funding, and its post-investment valuation has reached $15 billion. Isuzu is also one of the customers of Applied.

Sun isn't too worried. He believes that Laplace using AI to test the core methods of AI represents a "completely new paradigm for changing physical AI testing," turning testing into a "dynamic and interactive process."

"That will be the greatest moat," he said.