Laplace Intelligence Red-Teams Robots Before They Ship

Laplace Intelligence, an Ann Arbor, Michigan startup building automated red-teaming for robots and self-driving cars, presented on the Startup Battlefield 200 stage at TechCrunch Disrupt 2026 on October 13, 2026, making its case that physical AI systems need the same kind of deliberate, adversarial stress-testing that cybersecurity teams have run against software for decades, except aimed at machines that can cause physical harm when they fail. The company's premise is specific and uncomfortable: the industry's current answer to whether a robot or autonomous vehicle is safe enough to deploy is largely an internal judgment call, and Laplace is betting that the only credible way to answer that question is to have an AI system try, deliberately and systematically, to make the robot fail before a human operator, insurer or regulator ever encounters the failure in the field.

Laplace was founded a year ago out of the University of Michigan by chief executive Howie Sun and chief scientist Henry Liu, Sun's former doctoral adviser, building on research the two published in 2023 that made the cover of Nature on challenging autonomous driving systems in simulation. That academic result is what originally drew interest from autonomous vehicle companies, interest that eventually convinced Sun to leave research and build the idea into a product. The company takes its name from Laplace's demon, the 1814 thought experiment in which the French scholar Pierre-Simon Laplace imagined that perfect knowledge of every particle in the universe would make the future entirely predictable, a framing Sun uses deliberately: Laplace is, in his words, using simulation to predict the real world's future before a robot has to live through it.

Reconstructing Reality to Break a Robot on Purpose

At the center of Laplace's product is a high-fidelity simulation environment built on a post-trained version of Nvidia's Cosmos world-simulation platform, grounded in real-world data rather than synthetic or purely hand-built scenes. The company reconstructs real locations, Sun cites an Arizona highway and a Paris pedestrian crossing as examples, by blending real-world map data with reference imagery that can include something as simple as a Google Street View capture, then uses that reconstruction as the stage on which an AI agent continuously probes a robot or vehicle's behavior for weaknesses and failure points. Developers can prompt the simulator in natural language to change conditions in real time, adding afternoon traffic to a scene, for instance, which lets a test engineer explore edge cases that would be expensive, slow or simply unsafe to stage in the physical world.

The goal, as Sun described it, is to catch and fix problems for client companies ahead of deployment, then gradually assemble a common base of evidence that customers, insurers and regulators can all rely on together. That last detail is the part worth taking seriously beyond the pitch: Sun is explicitly positioning Laplace's output, not just as an internal engineering tool, but as the kind of third-party-credible safety evidence that an insurance underwriter or a regulator could eventually rely on when deciding whether to cover or approve a physical AI deployment. Building that kind of trusted evidentiary layer, if Laplace can actually deliver it, would address a real gap in the industry: there is currently no standardized, independently verifiable way for a robotics company to demonstrate safety performance to an outside party the way a crash-test rating or a cybersecurity penetration test report does in other industries.

Why This Matters More Now Than It Would Have Two Years Ago

Sun's explicit reference point for why this work matters now is the recent incident involving an OpenAI-affiliated agent interacting unpredictably with Hugging Face infrastructure, an episode that demonstrated how an AI agent can behave in ways its own developers did not anticipate even in a purely digital environment. His point, made directly to reporters at Disrupt, is that the same unpredictability becomes categorically more dangerous once AI agents are controlling machines that move through physical space rather than software that merely processes data. Robots and autonomous vehicles are moving from controlled pilot programs into general commercial deployment at a pace that has outrun the industry's ability to systematically test for the long tail of situations a real environment eventually produces, which is precisely the gap Laplace is positioned to exploit if its simulation-based testing proves rigorous enough to catch what internal teams miss.

For a procurement or risk team evaluating a robotics or autonomous vehicle vendor, Laplace's emergence is a signal worth tracking even before the company has a mature commercial product: it suggests that systematic, adversarial safety testing is becoming a distinct service category in its own right, separate from a vendor's own internal quality assurance process, in roughly the same way independent penetration testing became a standard procurement requirement in enterprise software once internal security assurances stopped being sufficient on their own. A buyer evaluating a physical AI vendor today has no real way to compare safety rigor across competing suppliers beyond each company's own marketing claims; an independent red-teaming layer, if it becomes an industry norm, would give buyers an actual basis for that comparison.

How Red-Teaming Differs From the Safety Testing Buyers Already Know

Most physical AI buyers already budget for some form of safety validation: certification bodies, functional safety standards drawn from industrial automation, and vendor-run test tracks for autonomous vehicles are all established parts of a deployment checklist. What distinguishes Laplace's approach from that existing toolkit is the adversarial posture built into the testing itself. A certification process typically checks whether a system meets a defined specification under expected conditions; a red-teaming agent is explicitly trying to find the conditions nobody specified, by deliberately manufacturing the kind of edge case, a pedestrian stepping out from behind a parked delivery truck at an unusual angle, a sudden change in surface friction, a partially obscured obstacle, that a specification-driven test plan would not think to include. That distinction mirrors almost exactly how cybersecurity practice split decades ago between compliance-driven auditing and genuine penetration testing, and the gap between those two disciplines in software security is a reasonable preview of the gap Laplace is betting will open up in physical AI safety testing as well.

The practical implication for a procurement team is that red-teaming is a complement to existing certification and safety processes rather than a replacement for them. A robot or vehicle that has passed its relevant industrial or automotive safety certification has demonstrated compliance with a known specification; a system that has also survived sustained adversarial simulation testing has demonstrated something additional, that it degrades gracefully, or at least predictably, outside that specification's boundaries. Buyers assembling a vendor shortlist for a physical AI deployment in a genuinely unstructured environment, a public sidewalk, a crowded retail floor, a shared warehouse aisle, should treat the presence or absence of this kind of adversarial testing layer as a meaningful differentiator between vendors who have only cleared the baseline certification bar and vendors who have gone looking for their own failure modes before a customer finds them instead.

What a Buyer Should Ask Before Trusting a Red-Team Result

The obvious question Laplace's pitch raises, and one that any buyer relying on its results eventually needs answered directly, is how closely a reconstructed simulation actually tracks the physical world it is standing in for. A simulator grounded in real map data and street-level imagery is a meaningfully more rigorous foundation than a purely synthetic test environment, but a simulation is still, by definition, a model of reality rather than reality itself, and the entire value of Laplace's red-teaming depends on failures discovered in simulation actually predicting failures that would occur in the field. Sun's own framing, borrowed from Laplace's demon, is candid about the limits of that premise: perfect predictive power was always a thought experiment, not an achievable engineering standard, and a buyer should treat Laplace's results as a meaningfully better filter than no adversarial testing at all rather than as a guarantee that a system passing its gauntlet will never fail in deployment.

That caveat does not diminish the commercial logic behind Laplace's approach so much as it defines the right level of trust a buyer should extend to it today. A startup one year removed from a university lab, competing for attention among twenty companies on a single pitch stage, has not yet accumulated the track record that would let an insurer or regulator treat its results as definitive evidence on their own. What Laplace has built, a real-world-grounded simulation layer paired with an AI agent explicitly designed to find a robot's breaking points, is nonetheless a genuinely different approach to a problem the physical AI industry has mostly left to each company's own internal judgment, and it is the kind of infrastructure layer worth watching regardless of whether Laplace itself becomes the company that standardizes it.

This analysis synthesizes statements made on the record at TechCrunch Disrupt 2026 and published company materials as of the publication date and should not be read as investment, financial, or professional advice; it is provided for general information purposes only.