WeRide Releases One-Step End-to-End Assisted Driving Solution, Expected to Be Mass-Produced in Vehicles by 2025

On August 21st, WeRide officially launched the one-stage end-to-end assisted driving solution, WePilot AiDrive, in collaboration with Bosch. This comes just half a year after the mass production of the "two-stage end-to-end" solution from their partnership started.

WePilot AiDrive has completed the core functionality verification and is expected to achieve mass production and vehicle integration by 2025, facilitating the global assisted driving industry to advance towards a more intelligent, efficient, and universally applicable large-scale application stage.

It is reported that compared to the traditional two-stage architecture of perceiving first and then making decisions, WeRide's WePilot AiDrive is more like an "experienced driver on the road" — it doesn't need to "see first and then think"; instead, it starts steering at the very moment it sees, resulting in shorter paths, faster reactions, and higher fault tolerance.

Image source: WeRide

According to the official video released, WeRide's WePilot AiDrive can achieve a balance of safety, comfort, and efficiency in various complex road conditions. It can easily handle the following scenarios: decisive and accurate lane changes in dense traffic, detours on construction sites with unmarked roads, unprotected intersection turns, avoidance in mixed traffic of pedestrians and vehicles, navigating sharp curves, instant and delay-free following and starting, as well as complex interactions with pedestrians, oncoming vehicles, and roadside obstacles on narrow roads in urban villages.

The architecture-level optimization brought by the one-stage end-to-end technology provides WeRide's WePilot AiDrive with three key advantages: scalable computing power adaptation, adaptability to different sensor configurations, and rapid iteration capability measured in days.

WeRide's WePilot AiDrive supports both multi-modal sensor fusion perception and pure vision. It can be adapted to high-computing-power platforms, and through model distillation, it can quickly adapt to medium and low-computing-power platforms. Middleware is used to decouple the algorithm layer from the hardware and basic software layers, making L2 level assisted driving function modular, thereby enabling rapid deployment and mass production.

At the same time, leveraging massive driving data, WeRide's WePilot AiDrive can automatically generate training labels, continuously expanding the model's coverage of edge cases, further enhancing the system's understanding and handling of long-tail scenarios. This enables a continuous, short-cycle, and low-cost improvement of product performance for consumers, truly achieving "always fresh with frequent use" and "the more you drive, the better it gets."

Currently, WeRide’s WePilot AiDrive has demonstrated highly human-like L2 assisted driving capabilities and exceptional safety performance across multiple scenarios. It is scheduled for mass production and vehicle integration in 2025, providing end users with a more efficient and safer L2 assisted driving experience.