NIO announced today that its World Model is scheduled for a significant new update. The latest version will first be equipped on the upcoming executive flagship SUV, the NIO ES9, which is set to begin deliveries soon. Users of existing vehicles with the Banyan, Cedar, and Cedar S intelligent systems will also receive an over-the-air update for the new version in June.
In January 2026, NIO plans to fully roll out an intelligent driving assistance version based on the World Model + closed-loop reinforcement learning technical architecture to the hundreds of thousands of vehicles equipped with the Banyan, Cedar, and Cedar S systems. In the three months following the January update, NIO users' mileage and duration of using the city navigation pilot feature increased by 92% and 116%, respectively, compared to the previous period.
The new version introduces three pioneering technologies. First, it is China's first intelligent driving assistance system capable of directly controlling the steering wheel and pedals. Traditional systems first generate a sampled trajectory via a model, which is then translated into vehicle operations for the steering wheel and acceleration/brake pedals. NIO's new World Model version is the first in China to eliminate the trajectory step, allowing the system to directly operate the steering wheel and pedals. This results in a shorter path, lower latency, and significantly enhances the smoothness and precision of vehicle control. This end-to-end modeling approach also forms the underlying control foundation for next-generation intelligent driving assistance systems.
Second, it features the industry's first three-layer training framework combining a complete world model, supervised fine-tuning, and closed-loop reinforcement learning. Building upon the World Model + closed-loop reinforcement learning architecture, the new version upgrades to this complete three-layer framework. The newly added supervised fine-tuning training utilizes high-quality behavioral data for precise behavior sculpting, enabling the model to learn human-like performance and refresh the human-like driving experience. Each layer of the training framework has a distinct role, allowing the model to simultaneously achieve high baseline performance, human-like behavior, and regulatory compliance.
Third, it is the first automaker-developed intelligent driving assistance system capable of recognizing overhead road signs. By reconstructing the information representation method of sensors, the new World Model version enables, for the first time in an automaker-developed system, real-time recognition and understanding of overhead signs for tidal flow lanes and variable lanes. This adds the capability to handle such scenarios, further improving lane selection accuracy. Without relying on high-definition maps or enhanced navigation maps, the system continuously expands the boundaries of user application scenarios.
Comments