Laplace Intelligence is using AI to stress-test robots and self-driving cars before deployment

Updated October 13, 2026 9:23 pm

Laplace Intelligence is building automated red-teaming tools for robots, autonomous vehicles and other physical AI systems, aiming to catch dangerous failures in simulation before those systems are released into the real world. The Ann Arbor startup, which is presenting at TechCrunch Disrupt in San Francisco on October 13, says its approach could become a new safety standard for AI that moves beyond screens and into streets, warehouses and cities.

The company’s pitch is simple but ambitious: if AI is increasingly being asked to operate in physical environments, then AI should also be used to challenge it. That matters because errors in robotics or self-driving systems are no longer just software bugs. They can turn into collisions, injuries, blocked roads or costly shutdowns.

Co-founder and CEO Howie Sun says the industry has already seen enough evidence that AI can behave unpredictably, even before it is embodied in a machine. In his view, putting that unpredictability into a car, a robot or a smart-city system raises the stakes dramatically.

Sun argues that the industry cannot afford to discover those failures after deployment, when mistakes can affect people and property in the physical world.

What Laplace Intelligence actually does

Laplace is developing a simulation-based testing platform that automatically creates difficult scenarios designed to break autonomous systems. In practice, that means the startup builds digital environments that resemble real roads, crossings and city streets, then deploys AI agents to probe those environments for weak points.

The company describes this as automated red-teaming for physical AI. Instead of waiting for a human tester to think of an unusual edge case, Laplace’s system tries to generate those corner cases on its own.

That testing process is central to the startup’s thesis: the safer the simulation, the safer the deployment. If a robot or autonomous car can be pushed into failure in a virtual environment, engineers can fix the problem before the machine reaches public roads or shared spaces.

Why the company says AI should test AI

Laplace believes traditional testing methods are becoming too slow and too limited for the next generation of physical AI. Developers often train vehicles and robots using recorded data, then convert that footage into simulations or ask human testers to imagine rare scenarios. That approach works to a point, but it becomes increasingly difficult as the systems get more general-purpose and are expected to navigate unfamiliar settings.

The startup argues that AI agents can do this job more efficiently by continuously generating and refining challenge scenarios. In effect, one AI system is tasked with finding the blind spots of another.

Sun says that over time the company wants to help create a shared evidence base that can be trusted by customers, insurers and regulators. That would make testing not just a development tool, but part of a broader safety and compliance framework.

How the simulation platform works

Laplace’s product combines a high-fidelity virtual environment with interactive agents that can alter conditions in real time. The simulation is grounded in real-world geography and infrastructure rather than invented landscapes, allowing developers to reproduce specific places with notable precision.

The company says its environment layer is built on a post-trained version of Nvidia’s Cosmos world simulation platform. Using public and commercial mapping inputs, Laplace can reconstruct places such as a highway in Arizona or a pedestrian crossing in Paris, then modify the scene as needed for testing.

That reconstruction process can start with visual references like street-level imagery and combine them with map data to preserve real-world scale and layout. Developers can then change variables through language prompts, adding conditions such as heavier afternoon traffic or more complicated traffic flow.

Laplace also says it brings in live traffic and congestion data from navigation providers, including TomTom, so the simulation reflects conditions that are closer to the real moment a vehicle might face on the road.

Why interactivity matters

Laplace says its simulator does more than render a background scene. The other actors in the environment behave dynamically, which is essential if the goal is to test real decision-making under pressure. A test car that swerves into oncoming traffic, for example, does not simply pass through the other vehicle as if it were a static object.

In the company’s demo, the simulated vehicle being confronted reacted to the threat rather than disappearing through it. That behavior is important because it turns the environment into a living test bed where every movement can change the outcome.

Sun described the setup as a kind of world model with participants that respond to one another, rather than a one-way simulation where only the tested agent behaves realistically.

Why the startup is focusing on corner cases

Laplace’s biggest differentiator is its use of AI agents to create edge cases automatically. Corner cases are the rare, tricky scenarios that can expose weaknesses in autonomous systems but are hard for humans to anticipate consistently.

Those scenarios can be created from multiple sources, including news reports and police collision records. By feeding those real-world signals into the testing pipeline, Laplace aims to produce situations that are not merely synthetic, but grounded in the kinds of incidents engineers actually worry about.

This is especially relevant for robotics, where the range of possible real-world failures is much broader than in tightly constrained driving environments. Sun says there is simply less historical data available for general-purpose robots than for road vehicles, which makes predictive testing harder and more valuable.

Sun says the biggest question in generalized robotics is not just how to make a robot move, but how to anticipate the failures nobody has seen yet.

From autonomous driving to general-purpose robots

Autonomous vehicles remain the startup’s most commercially mature market, but Laplace is already looking beyond cars and trucks. Warehouses, humanoid robots and smart-city systems are all potential customers for simulation-driven testing.

That expansion reflects a wider shift in the AI market. As robots become more capable and more mobile, they will need testing systems that can handle messy, unstructured environments. A warehouse robot might need to work around pallets, people, doors, forklifts and shifting inventory, while a delivery robot may encounter weather, roadworks, pedestrians and damaged surfaces all in one shift.

Laplace’s bet is that these environments are too complex to rely on human imagination alone. AI, it argues, must become part of the safety layer.

How did Laplace get here?

Laplace was founded by Sun and chief scientist Henry Liu, who built the startup from academic research at the University of Michigan. Sun spent about eight years studying how to challenge autonomous-driving systems in simulation before turning that work into a company.

The turning point came after Sun and Liu published a paper in 2023 on the topic. The research was significant enough to appear on the cover of Nature, and the attention quickly spilled from academia into industry. Autonomous vehicle companies began reaching out to discuss practical applications.

That interest convinced Sun that the technology had moved beyond theory. Instead of remaining a research project, it could become a product that engineering teams could actually use.

The company’s name comes from Laplace’s demon, a famous thought experiment proposed by French mathematician Pierre-Simon Laplace in 1814. The idea imagines that if an intelligence knew the exact state of every particle in the universe, it could predict the future with perfect accuracy.

Sun says the metaphor fits the startup’s mission: use simulation to model what the real world might do next, then stress that model until it reveals where the autonomous system breaks.

Who is using Laplace today?

Laplace says its early revenue is coming mostly from autonomous-driving applications. Among its customers are Japanese trucking manufacturer Isuzu and Texas-based autonomous trucking startup Bot Auto.

The company also counts Nvidia as a customer, according to Sun. In addition, Laplace is participating in a free smart-city pilot in San Jose with Nvidia and T-Mobile to support traffic operations at a high level.

Beyond transportation, Laplace says it is beginning to work with partners building humanoid robots for warehouse use. Those engagements are still at an early stage, but they suggest the company is already moving into sectors where physical AI failures could be both costly and visible.

• Isuzu — uses Laplace for autonomous driving-related testing
• Bot Auto — another trucking customer focused on autonomy
• Nvidia — customer and platform partner
• T-Mobile — collaborating on smart-city traffic operations in San Jose
• Warehouse robotics partners — early-stage humanoid robot work

What the business model looks like

Most of Laplace’s deals are project-based right now, which is common for young infrastructure and enterprise startups. The company is trying to turn those engagements into recurring revenue, a move that would give it more predictable cash flow and a stronger basis for growth.

Sun says the startup is already generating revenue and is reportedly profitable on a modest scale, a notable position for an early-stage company coming out of academia. Still, he acknowledges that scaling the business will require outside capital.

That is part of why the company is using TechCrunch Disrupt as a launchpad. The pitch competition offers visibility to investors, customers and potential partners, all of whom could help Laplace compete more aggressively.

How does Laplace compare with its rivals?

Laplace is entering a market that already includes well-funded simulation and autonomy-testing companies, most notably Applied Intuition. The late-stage startup has raised about $1.2 billion and is valued at roughly $15 billion, making it a formidable competitor.

Applied Intuition also counts Isuzu among its customers, which underscores how crowded and strategically important the space has become. If Laplace wants to stand out, it will need not just technical credibility but also a compelling reason for customers to switch or diversify their testing stack.

Sun believes the answer lies in Laplace’s core methodology. He argues that using AI to find weaknesses in other AI systems creates a fundamentally different testing loop — one that is more dynamic, more scalable and harder to replicate with manual methods alone.

That claim may prove difficult to validate quickly, especially in a sector where trust tends to be earned over years rather than months. But it also captures the logic behind the startup’s strategy: if physical AI is becoming more autonomous, the testing methods used to verify it may need to become autonomous too.

Why the timing matters

The startup is arriving at a moment when robotics and self-driving technology are both under pressure to move from demos to deployment. Companies want faster iteration, broader coverage of edge cases and stronger safety evidence before launching products into public spaces.

At the same time, regulators, insurers and enterprise buyers are becoming less willing to accept vague assurances about safety. They want proof. Laplace’s pitch is that simulation, if pushed hard enough and grounded in real-world data, can provide that proof earlier in the product lifecycle.

That may be especially valuable for sectors where a mistake can trigger a public backlash or costly liability dispute. A tool that can identify failure before release could become part of the go/no-go decision for enterprise customers.

What comes next for Laplace?

In the short term, Laplace is focused on using its Disrupt appearance to raise money and grow its customer base. Longer term, the company wants to scale its platform beyond one-off projects and become a recurring part of how physical AI systems are tested.

If it succeeds, the startup could help establish a new expectation in robotics and autonomous driving: not just build the machine, but prove that it can survive the worst-case scenario before anyone turns it loose in the real world.

For now, Laplace is still early, and the market it is pursuing is intensely competitive. But the underlying problem it is trying to solve is real and likely to grow. As more AI steps out of the browser and into physical spaces, the need for rigorous pre-deployment testing will only become more urgent.

That urgency is what Laplace is betting on — and what it is hoping investors, customers and regulators will eventually share.

Timeline of Laplace Intelligence’s rise

For an industry racing to put more intelligence into machines that move through the world, Laplace is offering a counterintuitive but increasingly necessary idea: let AI break AI before reality does.