A new research framework uses two local AI models on a single laptop to dream up and grade rare failure scenarios for autonomous systems, aiming to generate safety evidence without ever touching the cloud.
A new paper describes a system that runs a local language model through a loop of thinking, acting, observing, and updating to generate scenario descriptions, using a project's system requirements, system description, and environment description as background knowledge. A second, independent local model then critiques each scenario against domain-specific criteria before it gets refined. Refined scenarios run through simulation, and the resulting metric scores guide what happens next: similar scenarios get generated to dig deeper into known trouble spots, or novel ones get created to probe untested territory. Clustering those scores alongside a thematic read of the AI-written scenario text then surfaces functional gaps and the conditions that trigger them.
Edge cases are the whole point of safety testing for self-driving cars, drones, and other autonomous systems - the rare situations that don't show up in normal operation but can cause real failures. Most prior LM-based scenario generators lean on human reviewers or fixed rules, which caps how strange or diverse the generated scenarios can get. Running everything locally, with one model generating and another critiquing, removes both the human bottleneck and the need to ship sensitive system data to an outside API.
What the paper doesn't offer yet is a comparison against real crash data or larger cloud-based scenario generators, so whether a laptop-sized setup actually finds the failures that matter is still an open question.