小鹏关于自己VLA路线的一些QA

Core Viewpoint - The article emphasizes the advantages of using cloud-based large models for distillation into smaller models for vehicle deployment, highlighting the benefits of scale, unified modalities, and enhanced reinforcement learning capabilities [1][2][4]. Group 1: Cloud Model Distillation - The company is training a 72 billion parameter VLA model in the cloud, which will be distilled into smaller models for vehicle deployment, allowing the smaller models to inherit the capabilities of the larger model [1]. - Distillation from a larger cloud model to a smaller vehicle model enhances performance due to the scale effect, allowing for better learning from more data [1][2]. - The approach addresses the "modal unification" challenge in autonomous driving, where different drivers may respond differently to the same scenario, potentially leading to confusion and model collapse [2]. Group 2: Vehicle Edge Deployment - The VLA model must be deployed on the vehicle rather than in the cloud to mitigate safety risks associated with network latency, especially in critical driving situations [3]. - Delays of even 300 milliseconds can pose significant risks in vehicle control, necessitating local deployment for real-time responsiveness [3]. Group 3: Importance of Vehicle Chips - The local chip is crucial for the effective deployment of AI models, with the company developing a chip that offers three times the computational power of mainstream vehicle chips [4]. - The integration of chip design with software and model architecture is essential for maximizing performance and efficiency [4]. Group 4: Sensor and Computational Strategy - The company adopts a light radar and heavy computation strategy, which enhances the responsiveness and safety of autonomous driving systems [5]. - By eliminating laser radar, the company saves 20% of perception computing power, significantly improving visual processing speed and safety [5]. - The AI camera system, with advanced capabilities, provides superior visual input, enhancing the overall system's performance [5].