Laplace Intelligence wants an AI opponent to find a robot’s consequential mistake before the machine reaches a road or workplace. The University of Michigan spinout is building interactive simulations in which testing agents vary conditions, observe a target’s response and search again. The central question is not whether these worlds look convincing, but whether the failures they expose matter outside them.
The startup presented its approach among 20 competitors in Startup Battlefield 200 at TechCrunch Disrupt, held October 13–15 in San Francisco. CEO Howie Sun says the company, founded about a year earlier with chief scientist Henry Liu, has revenue, a modest profit and customers including Isuzu, Bot Auto and Nvidia. Revenue amounts, contract duration and retention were not disclosed.
That is early commercial interest, not an independently validated safety certification. Laplace faces larger simulation providers and an important burden: showing that an adaptive adversary produces better evidence than a fixed scenario library without replacing physical validation.
Search for the unusual instead of waiting for it
Red-teaming deliberately challenges a system before a real incident does. Laplace applies that practice to autonomous vehicles, warehouse robots and intelligent traffic systems. A developer can reconstruct a location, place the target system inside it and alter conditions with language prompts.
Rather than repeatedly checking one predefined pedestrian crossing, an adversarial agent can vary timing, visibility, road geometry and participant behavior. Afternoon congestion, an unusual driver or a difficult crossing becomes an opportunity to discover combinations that break the policy. Humans still define objectives and constraints.
Sun says collision reports, news and other records can inspire corner cases. Translating them into executable situations is not automatic truth: reports may omit timing, visibility or intent. A useful test needs those choices documented rather than silently treating generated details as observed facts.
Ordinary operation is inefficient at collecting rare safety events. Michigan’s 2023 research noted that demonstrating human-level safety under naturalistic driving could require hundreds of millions of miles, most uneventful. Searching dangerous interactions aims to obtain more evidence per test, not prove that routine miles are irrelevant.
The 2023 Nature study used dense reinforcement learning—agents learning through action feedback—to focus an intelligent environment on safety-critical situations. It removed many uneventful cases and connected virtual hazards to a real vehicle on a physical track. The work appeared on Nature’s March 23, 2023 cover and reported acceleration by orders of magnitude within studied scenarios.
A separate 2026 Michigan near-miss project reported a 90% safety improvement during Mcity tests and a 99.9% reduction in required testing miles. Those figures belong to that research setup. They are not measured gains for Laplace’s commercial product and must not be transferred to all robots.
A world model must react, not merely render
TechCrunch reports that Laplace uses a post-trained version of Nvidia Cosmos. Post-training adapts an existing model to a narrower task. Cosmos world foundation models can generate or predict physical environments from text, images, video and other inputs for synthetic data and simulation.
Laplace adds maps and references: a Street View image can guide appearance, map data can preserve scale and road structure, and traffic information can alter activity. Grounding matters because attractive imagery can still put objects in impossible positions or create unrealistic rules.
Interactivity makes another difference. The evaluated system changes its environment through action, so other participants must respond consistently. In TechCrunch’s described demonstration, Sun directed a test vehicle toward oncoming traffic and the other car moved aside, rather than passively allowing impossible overlap.
Above that feedback loop, the testing agent chooses changes to learn about the target’s limits. It could modify density, speed, visibility or another participant’s conduct, then respond to hesitation, a rule violation or an unrecoverable state with another scenario.
This is the mechanism behind “AI testing AI.” Its promise is adaptive search rather than asking engineers to anticipate every combination. Its weakness is that both sides can be wrong: a tester may miss hazards, invent implausible ones or optimize an objective that does not transfer to physical operation.
Reproducibility is therefore essential. Engineers need versioned worlds, stable criteria and detailed logs to rerun a failure after a software change. A changed scenario must not be mistaken for a fixed policy. Failures in the generated world also need separation from failures in the robot.
Physical machines widen the evidence burden
An incorrect chatbot answer need not move machinery. Physical AI connects perception and decisions to vehicles, cargo and people. Weather, lighting, other road users, objects and human behavior cannot all be controlled in advance.
Liu has argued for an autonomous-driving competency test comparable to a human driving test, including unprotected left turns and yielding to pedestrians. The source describes heavy US reliance on manufacturer self-certification and more direct third-party requirements in some European and Chinese contexts; that is not a claim that every jurisdiction has one identical regime.
A test can establish minimum competency and better evidence, not prevent every crash. Customers, insurers, regulators and deployment partners need repeatable methods and interpretable results. Realistic video alone cannot supply a shared basis of trust.
General-purpose robots may have even less structured incident data than driving companies. A warehouse changes layouts, objects, people and tasks between facilities or shifts. Blocked aisles, falling loads and unexpected human movements create a different set of edge cases from highway traffic.
That limits transfer between domains. Vehicle evaluation may emphasize collision rates, rule compliance and interventions; warehouses may emphasize separation from workers, dropped objects, task completion and safe recovery. A successful driving study does not validate a humanoid handling packages.
A focused method enters a broad platform market
Applied Intuition is the strongest commercial reference: simulation, validation, data and embedded software across automotive, defense, trucking and equipment. Its 2025 Series F raised $600 million at a $15 billion valuation. Product history, support, security reviews and integrations matter to customers with long programs.
Laplace’s proposed entry point is not replacing every simulator. It is finding meaningful failures missed by existing workflows. Its project-based contracts provide customer learning and revenue but can depend on custom labor; the company wants more recurring software income and continuous testing across releases.
Sun names Nvidia as a customer. Laplace, Nvidia and T-Mobile also participate in a free San Jose smart-city pilot focused on high-level traffic operations. The company is exploring warehouse humanoid work. A free pilot should not be counted as a disclosed paid contract, and exploration should not be described as proven broad deployment.
Waabi offers a different comparison: its AI-centered Waabi World supports its own trucking system rather than serving as a neutral testing layer for everyone. Buyers may combine internal simulation, broad platforms, open tools and specialist red-teaming rather than choose only one.
Nvidia is both an opportunity and a dependency. It supplies Cosmos, is a reported customer and pilot partner, while offering tools others can adapt. Applied Intuition and internal teams can also add automated adversaries. Laplace needs differentiated failure-search intelligence and useful integration above the underlying model, not novelty alone.
Failure scenarios and results are sensitive intellectual property. Buyers will also need clear ownership and access arrangements before exposing weaknesses to a testing service. Profit in early projects indicates willingness to pay, not proof of scalable recurring economics.
The simulation gap determines the value of more tests
Simulation simplifies sensors, physics and human behavior. A realistic pedestrian may move implausibly; a traffic scene may violate local conventions. Maps and responsive actors reduce some errors but do not eliminate physical testing.
An objective focused on spectacular collisions can miss repeated hesitation, uncomfortable movement, blocked work or poor fallback behavior. Engineers cannot investigate unlimited false alarms, so scenarios need ranking by plausibility, severity, novelty and reproducibility. A finite search can also keep revisiting familiar hazards and miss an interaction outside its assumptions.
The strongest validation would show a previously unknown simulated failure reproduced physically, then prevented by a software change without introducing another problem. Independent evaluators, tracks, universities, insurers and neutral safety organizations could compare generated cases with recorded and human-designed tests.
The three commercial signals follow from that requirement: project work converting to repeatable recurring use, published links between virtual findings and physical outcomes, and a lasting advantage or integration role as larger platforms respond.
Laplace need not prove simulation replaces real testing. Helping engineers select better physical experiments is a more credible and useful claim. Its wager is that one consequential weakness can justify another test cycle; its obligation is to show that the weakness is reproducible, plausible and meaningful beyond the screen.