3 6 Ke·2025-09-12 02:45Core Viewpoint - The emergence of "World Models" has complicated the high-end intelligent driving landscape, leading to debates over the authenticity and effectiveness of various models like VLA, WEWA, and others [3][5]. Group 1: Company Perspectives - Xiaopeng Motors claims to be the only company in China that has truly developed VLA, criticizing competitors for creating modified versions [3][7]. - Huawei's CEO of Intelligent Automotive Solutions stated that the company will not pursue the VLA path, emphasizing a focus on World Action (WA) instead of language processing [3][5]. - Li Auto is developing a foundational model to support its MindVLA algorithm, which is positioned as a marketing strategy rather than a true VLA implementation [7][8]. Group 2: Technical Insights - VLA (Vision-Language-Action) is seen as an evolution of the end-to-end plus VLM (Vision-Language Model) approach, addressing some limitations of the previous models [5][7]. - Xiaopeng is developing a large-scale driving model with 720 billion parameters, utilizing cloud distillation to deploy smaller models to vehicles [8][15]. - The concept of World Models, initially proposed by Tesla, aims to create a virtual environment for autonomous driving learning and validation [9][11]. Group 3: Industry Trends - The industry is witnessing a shift from perception-driven to cognition-driven approaches, with various companies exploring different architectures for intelligent driving [12][13]. - The debate over the effectiveness of VLA versus World Models reflects a broader struggle within the industry to define the best methodologies for achieving autonomous driving capabilities [17]. - The integration of cloud and vehicle-based models is seen as essential for optimizing perception and decision-making in autonomous systems [17].