He Xiaopeng Pushes L2-to-L4 AV Policy Leap

🔑 Enhanced Key Takeaways

• •XPeng's second-generation VLA model features 72 billion parameters, trained on massive real driving video data, and generates action commands directly from visual signals end-to-end without intermediate language translation[3].
• •XPeng plans to launch three mass-produced L4 Robotaxi models in 2026 using pure-vision perception without LiDAR or HD maps, equipped with four Turing AI chips delivering 3,000 TOPS computing power[2][5].
• •Volkswagen will be the first launch customer for XPeng’s second-generation VLA autonomous driving solution[4][8].
• •XPeng merged its Autonomous Driving Center and Intelligent Cockpit Center into a new General AI Center in February 2026, led by Liu Xianming, to integrate driving decisions and human-machine interaction[4][7].

🛠️ Technical Deep Dive

• •Second-generation VLA: 72 billion parameters; end-to-end generation of action commands from visual signals, skipping language translation; trained on massive real driving videos for physical world understanding[3].
• •Pure-vision system: No LiDAR or HD maps; supports global road environments and varying driving habits; redundant architecture in perception, steering, braking, communications, energy, and compute[2].
• •Hardware: Four in-house Turing AI chips per vehicle, providing up to 3,000 TOPS computing capacity[2][5].
• •Architecture shift: Breaks from traditional perception-prediction-decision-planning modules; commonality with Tesla FSD V14 in multi-modal input and emergence abilities like stopping at hand wave[6].
• •Platform: Built on SEPA 3.0 with steer-by-wire and rear-wheel steering for models like GX SUV[5].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline