GM identifies long-tail scenarios as central to autonomous driving safety, including rare events such as a mattress on the road, bursting fire hydrants, and power outages disabling traffic lights, as well as everyday common-sense situations like queuing in parking lots or navigating construction zones with worker gestures (https://spectrum.ieee.org/gm-scalable-driving-ai).

The company deploys Vision Language Action (VLA) models based on standard vision language models with specialized decoding heads for vehicle trajectories, 3D object detection, and interpreting overrides such as a police officer's hand gesture over a red light; a Dual Frequency VLA runs high-level semantic decisions at lower frequency and immediate spatial control at high frequency (https://spectrum.ieee.org/gm-scalable-driving-ai).

GM runs millions of high-fidelity closed-loop simulations daily, equivalent to tens of thousands of human driving days in hours, using diffusion models for Seed-to-Seed Translation to create domain changes such as turning clear-day data into rainy or foggy conditions while preserving geometry, plus the GM World simulator to generate new traffic scenarios via natural language and bounding boxes (https://spectrum.ieee.org/gm-scalable-driving-ai).