Image source: © 2025 Krish Capital Pty.Ltd

Highlights

  • One-stage end-to-end architecture combines perception and control for simultaneous “see-and-act” driving.
  • Core validation completed; mass production and deployments targeted for late 2025, subject to integrations.
  • Designed for varied sensors, scalable compute, and rapid model iteration using driving data.

WeRide (Nasdaq: WRD) announced WePilot AiDrive, a one-stage, end-to-end advanced driver assistance system developed with Tier-1 supplier Bosch. The release arrives roughly six months after both companies began mass production of a two-stage, end-to-end ADAS, and shifts the processing model from sequential perception–planning to a unified pipeline intended to “see and act” concurrently. The company frames this as a step toward shorter reaction times and more efficient routing while maintaining tolerance to faults.

According to WeRide, WePilot AiDrive has completed core function validation and is tracking for mass production and vehicle deployments later in 2025. The roadmap places the system within planned model-year cycles for automakers that are currently evaluating next-generation Level 2 feature sets. Actual timing will depend on OEM integration schedules, hardware availability, and regional certification.

Functionally, the system is engineered to handle a range of everyday and complex road scenarios: lane changes in dense traffic, detours around unplanned construction, unprotected turns at intersections, smoother acceleration when following vehicles, and nuanced interactions with pedestrians, oncoming cars, and roadside obstacles on narrower urban roads. These are common pain points for existing ADAS stacks that rely on brittle rule-based logic or require extensive calibration across markets.

The architecture emphasizes three practical advantages for automakers and suppliers:

Scalable computing footprint. WePilot AiDrive runs on high-performance platforms and can be distilled to mid- and lower-power systems. Proprietary middleware is designed to decouple algorithms from hardware and foundational software, supporting cross-platform portability. Modular Level 2 feature packages aim to simplify deployment and shorten integration cycles.

Sensor adaptability. The stack supports both pure-vision configurations and multi-sensor fusion, allowing OEMs to align cost, performance, and regional regulatory requirements without rewriting core logic.

Rapid iteration. Daily training pipelines leverage extensive driving data and automated label generation to refine the model. This approach targets faster improvements for edge cases and long-tail events, reflecting the principle that “the more you drive, the better it drives.”