Valgo

Companies deploying robotaxis and autonomous trucks need to know how their systems' crash risk compares to human drivers, but the underlying methodological choices in computing such baselines are often hidden behind a single reported number. Safety-critical autonomous systems are also hard to validate because failures are rare events, and traditional Monte Carlo simulation requires enormous sample sizes to estimate failure probabilities reliably. Regulators, insurers, and operators need defensible, transparent risk estimates rather than opaque point estimates, and existing tools do not expose the full range of methodological assumptions that go into a baseline crash rate.

Valgo is a public benefit corporation that builds tools and research to quantify risk for physical AI systems, organized around three areas: validation, deployment, and insurance. Its public tool, available at humanbaselines.com, computes human crash-rate baselines for robotaxis and autonomous trucks and reports a full set of defensible baselines via specification-curve analysis instead of a single number, making the underlying assumptions explicit. The company's research output includes methods for failure sampling using diffusion models and importance sampling, Bayesian safety validation for estimating failure probabilities of black-box systems with far fewer simulations than Monte Carlo, and risk-driven design of perception systems. Valgo's team has also co-authored a forthcoming MIT Press textbook on safety validation algorithms for autonomous and cyber-physical systems.