Core Insights - Xiaopeng Motors has launched its second-generation VLA, which innovatively eliminates the "language translation" step, enabling direct generation of action commands from visual signals [2] - This model is Xiaopeng's first mass-produced physical world model, functioning as both an action generation model and a physical world understanding and reasoning model [2] - The second-generation VLA can utilize a vast amount of real driving videos for training, with nearly 100 million clips used, eliminating the need for any manual data labeling [2] Model Capabilities - The second-generation VLA can simulate future decision scenarios based on its understanding of the physical world, generating more realistic long-tail scenarios for adversarial training, significantly enhancing its ability to address long-tail issues [2] - The model is designed to drive various applications, including cars, Robotaxis, robots, and flying cars, showcasing its cross-domain capabilities [2] Future Plans - At the 2025 Xiaopeng Technology Day, the company announced plans to invite pioneer users to co-create experiences with the second-generation VLA by the end of December 2025 [2]
颠覆传统架构 小鹏汽车VLA2.0将于2026年一季度发布