Volkswagen 成為 XPeng VLA 2.0 首發客戶

🔑 增強重點摘要

• •XPeng VLA 2.0 employs a 'Vision–Implicit Token–Action' architecture that bypasses language translation for direct visual-to-action generation, enabling faster responses[1][2][8].
• •Trained on nearly 100 million real driving video clips without annotation, equivalent to 65,000 years of human driving experience, enhancing long-tail scenario handling[2].
• •Powers applications beyond vehicles, including next-gen IRON humanoid robot with smoother walking via three Turing chips (2,250 TOPS) and VLT + VLA + VLM integration[1][2].
• •Integrates FastDriveVLA framework, reducing visual tokens by 75% (from 3,249 to 812 per frame) and computational load by 7.5x on nuScenes benchmark while maintaining accuracy[3][7].

🛠️ 技術深入

• •Architecture: 'Vision-Implicit Token-Action' path eliminates language bottleneck, using latent and trajectory tokens with world simulation for retraining from video and ego info[1][2][5].
• •Training data: ~100 million unannotated real driving clips; generates realistic long-tail scenarios for adversarial training[2].
• •Compute: Runs on Turing chips (2,200+ TOPS per chip; up to 3,000 TOPS with four in GX SUV); 30-billion-parameter model processed locally[3][9].
• •Optimizations: FastDriveVLA (XPeng-PKU collab) uses adversarial foreground-background reconstruction for token pruning, achieving 7.5x compute reduction on nuScenes[3][7].
• •Capabilities: 'Narrow Road NGP' boosts takeover mileage 13x in complex environments; emergent skills like hand gesture recognition and traffic light response[2].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線