AI Red-Teaming for Robots: How Laplace Tests Physical AI
Laplace Intelligence uses AI agents and realistic simulations to test robots and self-driving cars. See how its approach could change physical AI safety.
Before Robots Meet the Real World, Laplace Wants AI to Try to Break Them
A software bug can ruin an app. A failure in a self-driving vehicle or warehouse robot can hurt people and damage property. That's why testing physical AI isn't just another engineering step. Developers have to find the situations their systems can't handle - before those systems meet the real world.
Laplace Intelligence, based in Ann Arbor, is building a way to do that. The University of Michigan spinout combines real-world-informed simulation with AI agents that actively look for failure scenarios. Instead of relying only on recorded examples or asking engineers to imagine every edge case, Laplace aims to produce new tests and then watch what happens when a system is pushed.
This is a sensible direction, but it's not a safety guarantee. Simulations can miss the messy details of reality. And even if a system passes a test suite, it can still fail in a situation nobody anticipated. The real question is how to plug simulation into a broader safety workflow. Below is what Laplace is building, why its approach matters for robotics and autonomous vehicles, and what companies should evaluate before treating AI-driven testing as proof of readiness.
Why physical AI needs a different kind of testing
AI systems are moving beyond screens. Autonomous vehicles share roads with people. Robots operate in warehouses. Smart-city systems interpret and respond to traffic. In these settings, a mistake isn't just a wrong output - it can become an unsafe action in an environment that doesn't stay still.
Testing is hard partly because the most important situations are often rare. A vehicle may handle familiar roads well and then struggle with an unusual obstruction, surprising traffic behavior, or a tricky interaction among road users. A general-purpose robot runs into a similar problem: its environments and tasks are too varied to cover with a short list of scripted tests.
The edge-case problem
An edge case is an uncommon or difficult situation that can expose a system weakness. Often it's not a single odd event, but a combination: a specific road layout, poor visibility, unusual movement by another road user, and a decision the system hasn't really seen during training.
Developers have a few ways to find these cases. They can analyze recorded footage, convert past incidents into simulations, or ask human experts to come up with scenarios. Those approaches still matter, but they have obvious limits. Recorded data only covers what happened and what got captured. Human-designed tests take time and expertise, and even good experts can't reliably think of every interaction that matters.
Laplace's premise is that AI agents can help widen the search. The idea is straightforward: an agent can probe a system, change conditions, and look for responses that reveal weaknesses. The goal isn't only to generate more tests. It's to make the testing process more active - less about replaying a fixed set of situations and more about hunting for conditions where the system breaks.
Why this matters beyond self-driving cars
Autonomous vehicles are a good starting point because they already rely heavily on simulation and have clear safety requirements. But the longer-term challenge shows up in broader robotics too. As Laplace CEO and co-founder Howie Sun told TechCrunch, there's less data for generalized robots and fewer established lists of their real-world edge cases.
That creates a painful loop. You can't test a robot against every situation ahead of time, but deployment is when you discover what you missed. Laplace argues simulation can help interrupt that loop by letting teams explore plausible failures before deployment. It won't eliminate uncertainty - but it can surface some of it earlier, with less risk.
The deeper point is that physical AI testing needs to account for interaction, not just isolated model outputs. A vehicle or robot operates in a world where other actors also act. So the realism of the simulated world - and how its participants respond - ends up mattering more than it might in other kinds of AI evaluation.
How Laplace combines simulation and AI agents
Laplace's product has two connected pieces: a simulated environment meant to resemble real places, and AI agents intended to stress the system being tested. The company describes the result as an interactive simulation, where the environment and its participants respond to each other rather than acting like static scenery.
Sun and co-founder and chief scientist Henry Liu started the company about a year before the TechCrunch report, building on work they did at the University of Michigan. Sun spent eight years studying how to challenge autonomous-driving AI in simulation. In 2023, Sun and Liu published related work in Nature. Interest from autonomous-vehicle companies helped push the project from academic research toward a commercial product.
Reconstructing places, not just generating scenery
According to Sun, Laplace uses a post-trained version of Nvidia's Cosmos world-simulation platform. Laplace says it can reconstruct environments ranging from an Arizona highway to a Paris pedestrian crossing by combining multiple sources of real-world information.
A Google Street View image, for instance, can provide visual reference. The simulation also uses real-world map data to keep the environment true to scale, rather than just "looking right." Engineers can then steer conditions using language prompts - like adding afternoon traffic. Sun also said Laplace draws on real-time traffic and congestion data from providers such as TomTom.
And yes, this distinction matters. A scene that looks realistic isn't automatically useful for testing. Road geometry, distances, traffic flow, and how other road users behave all affect outcomes. Visual fidelity helps, but the simulation also has to represent the conditions relevant to whatever decision the system is making.
Testing interactions, not just images
In a TechCrunch-described demo, Sun had a test vehicle swerve into oncoming traffic. The approaching car moved to avoid a collision. That example points to Laplace's focus on simulated participants that react to events, not just objects placed in the scene.
A world where every other actor follows a prewritten path might be enough for some tests. It's less useful when you want to see how a system behaves as the situation changes in response to its own actions. Interactive simulation makes those "action changes what happens next" scenarios easier to probe.
The second piece is the test agent. Laplace says its agents can use information like news reports and police collision reports to design challenging scenarios for autonomous vehicles and other physical AI systems. In principle, that's a way to turn descriptions of real incidents into candidate tests, then explore variations inside the simulation.
There's still engineering judgment in the loop. An agent can generate something novel but irrelevant, or it can produce a failure that doesn't really match the original incident. The useful metric isn't how many scenarios get generated. It's whether teams can understand them, reproduce them, prioritize them, and learn from the ones that expose meaningful weaknesses.
Where the business opportunity - and the limits - are
Laplace is selling into a market where the cost of missing a failure can be huge, while the cost of exhaustive real-world testing is also substantial. Simulation offers a way to run repeatable experiments without putting a vehicle, robot, or bystander into every test. If AI scenario generation works well, it can also make the process more flexible - especially when engineers don't know in advance which edge cases to look for.
Laplace started with autonomous driving, and the company says most of its current revenue comes from that. Its customers include Japanese truck manufacturer Isuzu and Texas-based autonomous trucking company Bot Auto. Sun also described Nvidia as a customer. Laplace, Nvidia, and T-Mobile are working on a free smart-city pilot in San Jose, California, focused on high-level traffic operations. The company is also in early discussions with partners building humanoid robots for warehouses.
From project work to repeatable revenue
Sun told TechCrunch that the startup is generating revenue and has reached modest profit. Most contracts, though, are project-based, and Laplace is trying to build recurring revenue. That matters when you evaluate the business, not just the technology.
Project work can help a young company learn what customers need and show that its product solves a real problem. But projects can come with bespoke integrations, custom environments, and ongoing human effort. If each deployment requires a fresh build, growth can get limited by staffing and delivery capacity. To be repeatable, the product needs common workflows: reusable scenario libraries, clear evaluation methods, integrations into customer development processes, and a real reason customers keep using it after the initial engagement.
Laplace also competes with well-funded players. TechCrunch identified Applied Intuition as a major rival, citing $1.2 billion raised and a reported post-money valuation of $15 billion. Isuzu is also an Applied Intuition customer. According to Sun, Laplace differentiates with a methodology where AI tests AI through dynamic, interactive scenarios.
That's a solid product thesis, but a "method" isn't automatically a long-lasting moat. Customers ultimately care whether the tool finds failures they would otherwise miss, whether those findings can be reproduced, and whether the workflow improves safety and development without turning review into a time sink. Evidence, workflow fit, and trust will matter as much as the novelty of the approach.
Simulation is evidence, not proof
The biggest technical limitation is the gap between simulation and reality. A simulation might capture road geometry and traffic patterns, but still miss details that matter to how a system behaves. Even if the test agent searches effectively within the scenario space it can represent, it can still overlook failure modes outside that model. So passing a simulation test doesn't mean the system is safe in every real-world condition.
Teams should treat simulation as one layer in a wider validation process. That can include software testing, hardware-in-the-loop evaluation, closed-course trials, supervised pilots, operational monitoring, and incident review. Which layers matter depends on the product and the risk. A warehouse robot in a controlled space and an autonomous truck on public roads don't face the same exposure.
None of this is meant to dismiss simulation. It's about being precise about what results actually show. A useful test report should spell out what was modeled, what scenarios ran, what failures appeared, and what stayed untested. That's more informative than a broad claim that an AI system was "validated."
What companies should look for before deployment
If you're evaluating AI-driven simulation, don't start with the question, "How many scenarios can this platform generate?" Volume is easy to measure and hard to interpret. A better evaluation is whether the generated tests are relevant, repeatable, and tied to decisions engineering and safety teams actually have to make.
A practical review can start with four questions:
- Can you trace a scenario to its source? If a test comes from an incident report, a data source, or a generated variation, the team should understand where it came from and what its limitations are.
- Can you reproduce the result? A failure that can't be replayed consistently is hard to debug - and hard to use as evidence.
- Does the simulation represent the factors that matter? Visual realism isn't enough. Teams should check the geometry, timing, participant behavior, and other assumptions that affect the system's decisions.
- What happens after a failure is found? The workflow should connect the result to diagnosis, fixes, regression testing, and a real decision about deployment risk.
Build a feedback loop around failures
The strongest use of a red-team style system isn't a one-time pass or a pass/fail label. It's a loop: identify a scenario, reproduce the behavior, figure out why the system failed, make a change, and rerun the test alongside earlier cases. If the same failure pops up after a software update, the suite should make that obvious.
This also changes how teams prioritize. Not every unusual scenario deserves equal attention. A useful triage process considers severity, likelihood, how confident you are in the simulation, and whether the failure can be mitigated. Without that prioritization, automated scenario generation can create a growing backlog of interesting examples that still doesn't help anyone decide what to do next.
Keep human review where judgment is needed
Automation can reduce the work involved in proposing and running tests. It doesn't remove the need for human review. Engineers and safety specialists still need to judge whether a scenario is plausible, whether the outcome is a real system weakness, and what level of evidence is enough to close an issue.
That balance matters even more when agents use unstructured sources like news reports or police records. These sources can point teams to leads, but they may not include the details required to recreate an event accurately. A responsible workflow distinguishes a useful lead from a validated test case and keeps track of the assumptions made along the way.
Laplace's approach is most convincing when you view it this way: not as a replacement for test engineers or real-world validation, but as a tool to help teams search a bigger space of potential failures before deployment. The payoff depends on the quality of that search and how rigorously teams work through the results.
Conclusion
Laplace Intelligence is trying to address a real gap in physical AI development: hard-to-predict situations that can expose weaknesses in vehicles and robots. Its combination of interactive simulation and AI-driven scenario generation could help engineering teams find more problems before deployment, especially as robotics moves beyond tightly defined tasks.
The claim worth keeping in perspective is this: testing isn't the same as proof. Simulation results are only as useful as the assumptions, scenarios, and follow-up work behind them. If you're evaluating tools like this, start with a narrow pilot - pick a specific system and risk area, define what a meaningful failure looks like, and then see whether the platform produces reproducible issues that genuinely improve your validation process.
For more context, read the TechCrunch report on Laplace Intelligence.