Breaking Robots on Purpose, Before They Break Something Else

Picture a warehouse robot at 2 a.m. A pallet has shifted four inches out of place, the overhead light is flickering, and a human walks into the aisle from a blind corner at an angle nobody scripted. In the real world, that moment costs money, or worse. In simulation, it costs a few seconds of compute and generates a log file.

That second version of the scenario is the business Laplace Intelligence is in. The Ann Arbor company, a University of Michigan spinout, builds automated red-teaming for robotics and physical AI. Translated out of security-speak: it uses AI agents and interactive simulation to deliberately try to make a robot or self-driving vehicle fail, over and over, before that system touches pavement or a loading dock. The startup is profitable and is competing in TechCrunch Disrupt’s Startup Battlefield.

Why this category exists at all

I review a lot of tools, and most of them promise to help you build faster. Very few promise to help you find out what you built wrong. That asymmetry is the whole story here.

Software teams figured this out years ago. You don’t ship a web service because it passed the happy-path test; you ship it after someone has thrown malformed input, race conditions, and traffic spikes at it. Red-teaming became standard practice because adversarial testing finds the bugs that polite testing never will.

Physical AI has been slower to adopt that mindset, and the reason is mostly logistics. You can replay a million malformed HTTP requests on a laptop. You cannot easily arrange a million near-miss intersections with a real vehicle, a real pedestrian, and real weather. So the testing that happens tends to be the testing that’s convenient, which is not the same thing as the testing that matters.

High-fidelity simulation changes the economics of that. If the adversarial cases can be generated automatically and run in software, the expensive part stops being the test and starts being the decision about what to test.

What the agent layer is actually for

The phrase “AI agents” gets thrown around loosely enough that I want to be specific about what it plausibly buys you in this context. Simulation on its own is old news. Driving simulators, physics engines, and digital twins have been around for a long time, and plenty of teams already run scenarios in them.

The bottleneck has never been the simulator. It’s been scenario authorship. A human engineer writes test cases based on failures they can imagine, which means the test suite inherits the engineer’s blind spots. Agents that search for failures on their own are attacking exactly that limitation. They explore the space of weird conditions without needing someone to think of each one first.

That’s the part I’d want to kick the tires on if I were evaluating this. Not the rendering quality of the simulation, but how good the agents are at finding failures a competent human team would have missed. That’s the only metric that justifies the layer existing.

The business question I keep circling back to

Laplace is profitable, which is unusual enough in this corner of AI that it deserves a mention. It’s also working to convert project-based contracts into recurring revenue, and that transition tells you something about where the product currently sits.

Project-based work in testing usually means bespoke engagements. A customer has a specific robot, a specific environment, and a specific deadline, and the vendor builds something tailored. It pays well and it proves the technology works. It also doesn’t scale cleanly, because every engagement starts partway from scratch.

Recurring revenue means a platform customers run themselves, continuously, as part of their development cycle. That’s a harder product to build and a much better one to own. Teams that are serious about safety don’t red-team once before launch; they red-team every build. Whether Laplace gets there is the thing I’d watch, because it’s the difference between a very good consultancy and a piece of infrastructure.

The partnerships signal

The company is working with partners including T-Mobile and Nvidia on smart city initiatives and warehouse robotics. Those are two meaningfully different deployment environments, which suggests the approach isn’t welded to a single use case.

It also puts Laplace adjacent to the compute and connectivity providers who are building out the physical AI stack from the other direction. Being the testing layer next to those players is a reasonable place to stand.

My honest read

Adversarial testing for physical systems is one of the few AI tool categories where the value proposition doesn’t require any imagination. A robot that fails in simulation is a story nobody tells. A robot that fails in a warehouse aisle is an incident report, a lawsuit, or a headline.

Where I stay skeptical is the gap between generating a lot of failure cases and generating the right ones. Volume is easy to demo. Relevance is what customers actually pay for, and it’s much harder to prove from the outside.

Still, the category is correct, the team came out of a serious research program, and the company makes money. In this market, that combination is rarer than it should be.

🕒 Published: October 14, 2026