Leapmotor Selects One-Stage End-to-End World Model Architecture for 2026 Autonomous Driving Strategy
Chinese EV maker Leapmotor announced its shift to an in-house developed "Autonomous Driving 4.0" architecture, adopting a one-stage end-to-end world model following an internal competition among three technical routes. Third-party testing demonstrated that this 4.0 architecture achieved a human intervention takeover rate of just one-third compared to top-tier industry competitors. This technical shift dramatically cuts system underlying costs to 40% of its first-generation architecture, enabling Leapmotor to bring advanced world model autonomous driving to budget EVs priced around 100,000 RMB (~$14,000 USD). It highlights a broader automotive trend of replacing multi-stage perception-planning pipelines with unified end-to-end neural networks to achieve top-tier autonomous capabilities. Unlike older two-stage architectures that lost perception data during stage transitions and relied heavily on hand-coded rules, the 4.0 one-stage model directly feeds sensor images into the network to output decisions with minimal rule-based fallback. Leapmotor plans to offer free OTA updates for the new system starting next year on LiDAR-equipped models, supporting its roadmap to advance to L3 autonomy by 2027–2028.
## BACKGROUND
End-to-end autonomous driving replaces modular pipelines—which separate perception, tracking, and motion planning—with a unified neural network that maps raw sensor input directly to control outputs. World models enhance this architecture by enabling the AI system to simulate and predict future environmental states, improving dynamic interaction handling and response speeds.