Autonomous Driving Strategies Diverge: End-to-End Mass Production Arrives, VLA Remains Long-Term Play

Current intelligent driving technology paths are clearly diverging. Solutions represented by the "single-stage end-to-end" approach—boasting engineering advantages and human-like driving experiences (such as smooth car-following and predictive behavior anticipation)—have already achieved mass production. Bosch’s collaboration with Horizon Robotics has been deployed on 150,000 RMB-class vehicles like the Chery Exeed Sterra ES and ET5, accelerating the adoption of advanced intelligent driving features. In contrast, while Vision-Language-Action (VLA) large models demonstrate the ability to understand visual inputs and language instructions—pointing toward a highly intelligent future—they face significant hurdles in the near term, including challenges in multimodal alignment, scarcity of training data, and demanding chip compute requirements, making them unlikely to replace end-to-end solutions soon. Industry consensus holds that the immediate focus should be on reliable mass production, while laying groundwork for future VLA integration. Bosch plans to introduce generative AI and VLA technologies by 2026. Ultimately, the winners in China’s intelligent driving race will be those companies that successfully balance user experience, safety, and technological integration.