Xiaomi EV, the electric vehicle division of Xiaomi, unveiled an upgraded version of its driver-assist system on Thursday, adding world model capabilities as part of the company’s continued push into advanced assisted driving technologies.
The enhanced system, called Xiaomi HAD (Hyper Autonomous Driving) Enhanced Edition, debuted at the opening of the 2025 Guangzhou auto show. Xiaomi said the updated version incorporates reinforcement learning and world model functionality built on its dataset of 10 million driver-behavior clips. The company noted that its ViSE algorithm won the world-model category at the 2025 ICCV professional challenge, while a paper related to its Genesis generative model was accepted by the NeurIPS AI conference.
Xiaomi said the upgrade improves overall user experience, including enhanced traffic-condition analysis. China’s major EV makers have increasingly explored end-to-end training approaches in recent years, and many have begun adopting world model frameworks to better simulate human driving behavior.
The latest Xiaomi HAD release includes improvements to AEB (Autonomous Emergency Braking) and adds a new AES (Automatic Emergency Steering) function. The company reiterated that its system is an assistive tool rather than an autonomous driving solution, stating that “drivers must remain fully attentive at all times.”
More than 90% of Xiaomi EV users currently engage the company’s assisted driving functions, which Xiaomi says have helped prevent 457,000 potential collisions. The firm plans to increase research spending on assisted driving next year, with AI investment exceeding RMB 7 billion ($1 billion). Xiaomi’s smart driving team includes more than 1,800 members, among them 108 PhDs.
The 23rd Guangzhou International Automobile Exhibition opened Thursday, with public access beginning November 22. Xiaomi EV’s booth, located in Hall 17.2 of Zone D, is exhibited alongside brands including Nio’s Onvo, Xpeng, Avatr, Luxeed and Changan Mazda.
Xiaomi began rolling out the HAD feature earlier this year to eligible vehicles at no additional cost. A major upgrade in July applied the 10-million-clip training dataset, which the company said significantly boosted assisted-driving performance. The clips represent short video segments of human driving behavior used to train AI models.