Xiaomi Makes Autopilot Breakthrough With 10 Million Real Videos Training

Chinese tech giant Xiaomi has unveiled a major update to its driver assistance system, which is based on a fundamentally new approach to training artificial intelligence. Unlike traditional methods using computer simulations, Xiaomi's algorithm analyzed 10 million short videos of real-life driving situations recorded by cameras, lidars, and radars on test vehicles.

Each 30-60 second fragment demonstrated specific driving scenarios: from traffic jams and sudden braking to complex intersections and avoiding obstacles. This approach allowed the system to not only follow programmed rules, but also learn from examples of human behavior. The volume of the training base has more than tripled compared to the original 3 million videos.

The results are impressive: control smoothness has improved by 57%, obstacle avoidance efficiency has increased by 67%, and intersection navigation accuracy has increased by 23%. The update is already available for the SU7 Pro, Max, and Ultra electric vehicles, and will also be standard on the upcoming YU7 sedan.

This achievement was made possible by Xiaomi's massive investment in autonomous technology - 5.79 billion yuan ($807 million), the work of 1,800 engineers, and a fleet of 400 test cars collecting data. The proprietary MiMo neural network continues to train the system, which promises further improvements to the autopilot in future updates.