AUTOSIGNAL
Human-curated. Expert-annotated. Every signal traced to source.
人工精选 · 专家点评 · 每条信号可溯源
Industry Moves行业动向
Xpeng Unveils VLA 2.0 Autonomous Driving AI, Trained on 100 Million Video Sequences小鹏发布VLA 2.0自动驾驶AI，用1亿段视频序列训练
S 1.5
T3
1 sources1 个来源
R1-money-or-metal
- Xpeng, a Chinese automaker listed in New York and Hong Kong, unveiled VLA 2.0, its next-generation autonomous driving AI model, at its Media Experience Day on March 2, trained on approximately 100 million video sequences — equivalent to 65,000 years of human driving.
- VLA 2.0 uses a Vision-Implicit Token-Action architecture that maps camera images directly to vehicle controls such as steering, braking and acceleration without a natural-language processing step, delivering inference speeds 12 times faster than the previous generation.
- CEO He Xiaopeng says the model drives "intuitively," handling complex scenarios such as narrow streets, unmapped roads and unpredictable pedestrian behavior — an advance the report says is significant enough to unsettle Wall Street investors in the automaker.
- 在纽约和香港两地上市的中国车企小鹏，于3月2日Media Experience Day上发布新一代自动驾驶AI模型VLA 2.0，训练数据约1亿段视频序列，相当于65000年人类驾驶经验。
- VLA 2.0采用“视觉-隐式Token-动作”（Vision-Implicit Token-Action）架构，无需自然语言处理环节即可将摄像头画面直接转化为方向盘、刹车、油门等车辆控制指令，推理速度较上一代提升12倍。
- CEO何小鹏表示该系统驾驶表现“直觉化”，能应对狭窄街道、未绘制地图道路及行人不可预测行为等复杂场景——据报道，这一进展之强足以令华尔街投资者为之震动。