LI AUTO(US:LI) 理想TOP2·2025-11-20 10:42Core Viewpoint - The article discusses the advancements in autonomous driving technology, particularly focusing on the VLA (Vision-Language-Action) architecture developed by Li Auto and the insights shared by Xiaopeng's autonomous driving head, Liu Xianming, during a podcast. Liu emphasizes the removal of the intermediate language component (L) to enhance scalability and efficiency in data usage [1][4][5]. Summary by Sections VLA Architecture and Training Process - The VLA architecture involves a pre-training phase using a 32 billion parameter (32B) vision-language model that incorporates 3D vision and high-definition 2D vision, improving clarity by 3-5 times compared to open-source models. It also includes driving-related language data and key VL joint data [10][11]. - The model is distilled into a 3.2 billion parameter (3.2B) MoE model to ensure fast inference on vehicle hardware, followed by a post-training phase that integrates action to form the VLA, increasing the parameter count to nearly 4 billion [13][12]. - The reinforcement learning phase consists of two parts: human feedback reinforcement learning (RLHF) and pure reinforcement learning using world model-generated data, focusing on comfort, collision avoidance, and adherence to traffic regulations [15][16]. Data Utilization and Efficiency - Liu argues that using language as a supervisory signal can introduce human biases, reducing data efficiency and scalability. The most challenging data to collect are corner cases, which are crucial for training [4][6]. - The architecture aims to achieve a high level of generalization, with plans to implement L4 robotaxi services in Guangzhou based on the current framework [4][5]. Future Directions and Challenges - Liu acknowledges the uncertainties in scaling the technology and ensuring safety, questioning how to maintain safety standards and align the model with human behavior [5][18]. - The conversation highlights that the VLA, VLM, and world model are fundamentally end-to-end architectures, with various companies working on similar concepts in the realm of Physical AI [5][18]. Human-Agent Interaction - The driver agent is designed to process short commands directly, while complex instructions are sent to the cloud for processing before execution. This approach allows the system to understand and interact with the physical world like a human driver [17][18]. - The article concludes that the traffic domain is a suitable environment for VLA implementation due to its defined rules and the ability to model human driving behavior effectively [19][20].