Researchers deploy AI-trained autonomous driving on real vehicles without any real-world testing

Researchers built a system that takes driving policies trained entirely in simulation and runs them on actual vehicles without retraining on real data. The system works by translating simulator outputs into real-world commands and converting camera images into a format the simulated AI can understand, achieving 75–90% success rates on basic driving tasks.

The bottleneck for autonomous vehicles has always been the gap between simulation and reality — policies that work perfectly in CARLA (a video game built for testing self-driving software) usually fail catastrophically on real roads because the simulator doesn't capture the messiness of actual driving. If this sim-to-real transfer stays reliable as it scales, it collapses the cost and timeline for training autonomous systems, since companies can iterate in simulation for months or years before touching a real vehicle. The catch: these are basic tasks on a single vehicle type. Whether this works on a fleet, in bad weather, or at highway speeds remains unknown.