The OpenAI-Hugging Face incident underscored a crucial point: AI agents can behave unpredictably.
“Consider the repercussions if such an event occurred in the physical realm,” stated Howie Sun, CEO and co-founder of Laplace Intelligence, to TechCrunch. “As AI agents transition from digital environments to the real world, preventing failures in physical AI becomes paramount.”
Laplace, based in Ann Arbor, specializes in automated red-teaming, a method of rigorously stress-testing systems by intentionally attempting to induce failures, specifically for robotics and physical AI. The company operates on the principle that AI itself is the most effective tool for ensuring the safety of physical AI.
“We aim to assist companies in identifying and rectifying issues prior to deployment, ultimately establishing a robust evidence base that fosters trust among customers, insurers, and regulators,” Sun explained.
Laplace is among the 20 startups vying for recognition in Startup Battlefield 200, a pitch competition hosted at TechCrunch Disrupt in San Francisco from October 13-15. The company is scheduled to present on the Disrupt Stage on the afternoon of Tuesday, October 13.
A year ago, Sun and his co-founder, chief scientist Henry Liu, established Laplace, originating from the University of Michigan. Sun dedicated eight years to academic research, focusing on methods to challenge autonomous driving AI within simulated environments. In 2023, he and Liu, who was his PhD adviser at the time, co-authored a groundbreaking paper on this topic, which was featured on the cover of the prestigious scientific journal Nature. This publication prompted inquiries from numerous autonomous vehicle companies keen on adopting their concepts.
“I recognized the potential to transform this from purely academic research into a practical, real-world product,” Sun stated.
Drawing upon their academic foundation, Sun and Liu developed Laplace, a name inspired by “Laplace’s demon.” This 1814 thought experiment by French scholar Pierre-Simon Laplace posited that complete knowledge of the universe’s state would allow for perfect future prediction.
“Our core approach involves leveraging simulation to foresee real-world outcomes,” Sun explained.
Central to the startup’s offering is a high-fidelity simulation, meticulously reconstructed from real-world data, combined with an AI agent that continuously seeks out vulnerabilities and potential failure points. This aims to guarantee the safety of any physical AI—whether a robot, autonomous vehicle, or smart city solution—within a simulated environment before its real-world deployment.
The simulation platform is an enhanced version of Nvidia’s Cosmos, an AI world simulation platform rooted in real-world data. This enables Laplace to accurately replicate diverse environments, from an Arizona highway to a Parisian pedestrian crossing. Simulations are crafted by integrating various real-world data sources; for instance, a Google Street View image can serve as a basis for generating an entire street. Crucially, the simulator leverages actual map data to reconstruct environments precisely to scale, avoiding random generation.
Developers have the capability to modify the simulation environment in real time using language prompts, such as incorporating afternoon traffic. According to Sun, Laplace integrates live traffic and congestion data from navigation services like TomTom.
In Laplace’s simulations, both the environment and the behavior of other entities are simulated. During a demonstration, for instance, Sun instructed a test agent vehicle to swerve into opposing traffic. The vehicle it was about to collide with—unlike in some generative world models where it might pass through—actively maneuvered to prevent impact.
“This is genuinely an interactive simulation environment,” Sun remarked. “Every element responds to every other element. …It functions more as a world model featuring interactive participants.”
Although the simulated environment is a significant component of Laplace’s offering, its primary differentiator lies in the AI agents the startup is developing. These agents are designed to autonomously generate corner cases—complex, infrequent scenarios that could challenge an AI system—thereby reducing human intervention.
“These agents can draw information from news articles, police reports on collisions, and various other sources to actively challenge autonomous vehicles and other physical AI systems,” Sun explained.
Currently, autonomous vehicle developers primarily train AI to manage edge cases either by converting recorded footage into simulations or by having humans devise and implement new scenarios. Although several AV companies, such as the self-driving truck firm Waabi, have begun adopting AI for generating training edge cases, as Laplace advocates, the full necessity of this technology will become evident with the deployment of generalized robots in real-world settings.
“Information on generalized robotics is scarcer,” Sun noted. “The specific edge cases that will emerge in real-world applications remain unknown.”
Despite Sun and Liu’s academic backgrounds and lack of a typical Silicon Valley tech pedigree, they have successfully generated revenue and achieved a modest profit. A significant portion of this income stems from autonomous driving applications, with clients such as Japanese trucking manufacturer Isuzu and Texas-based autonomous trucking firm Bot Auto.
Sun also mentioned Nvidia as a client. Laplace, Nvidia, and T-Mobile are collaborating on a complimentary smart city pilot project for San Jose, focusing on advanced traffic management. Additionally, Laplace is in preliminary discussions with partners developing humanoid robotics for warehouse applications.
Currently, the majority of Laplace’s customer agreements are project-specific, and the company is striving to transition these into recurring revenue streams. A key short-term objective for the startup, hopefully bolstered by its participation at Disrupt, is to secure funding and expand its operations to effectively contend with competitors.
A prominent competitor is Applied Intuition, a high-profile late-stage startup that has accumulated $1.2 billion in funding, achieving a $15 billion post-money valuation. Notably, Isuzu is also a client of Applied.
Sun remains unconcerned, asserting that Laplace’s fundamental approach of employing AI to test AI represents a “new paradigm that completely transforms the testing of physical AI,” evolving it into a “dynamic interactive testing process.”
“That will serve as our strongest competitive advantage,” he stated.