Group 1 - The current development of autonomous driving systems is significantly lagging behind expectations, facing numerous challenges, particularly in achieving safety and advancing from L3 to L4 and L5 levels [1][2] - Traditional "data feeding" methods are insufficient for complex scenarios, necessitating a new paradigm of "self-learning + prior knowledge" to enhance safety and generalization in high-level autonomous driving [1][5] - The focus is shifting towards a human-centered technology approach, emphasizing the construction of cognitive capabilities that surpass human abilities [1][9] Group 2 - Intelligent vehicle safety is a critical national demand, especially in China, where complex road traffic scenarios and frequent accidents pose significant challenges [2][3] - Low-level intelligent vehicles have achieved high market penetration, but there are still many safety challenges to overcome as the industry moves towards higher levels of automation [2][3] - A complete "perception-cognition-decision" technology system is essential for rapid perception, accurate judgment, and efficient response to complex dynamic scenarios [2][3] Group 3 - Current intelligent vehicles struggle with accurate perception, cognition, and safety decision-making in unpredictable and complex situations [3][4] - The rule-driven approach is limited to known structured scenarios, while the data-driven approach suffers from a lack of interpretability and generalization capabilities, making it inadequate for L4+ level autonomous driving [3][4] - Both rule-driven and data-driven methods face critical challenges in adapting to complex environments and ensuring safety [4][5] Group 4 - To address the limitations of existing methods, a cognitive-driven approach is proposed, which combines the interpretability of rule-driven systems with the learning capabilities of data-driven systems [5][6] - This cognitive-driven approach aims to enhance the system's ability to generalize, evolve, and make reliable decisions by understanding the interactions and dynamics of the human-vehicle-road system [5][6] Group 5 - The cognitive-driven architecture encompasses three main layers: perception, cognition, and decision-making, integrating both rule-based and data-driven elements [6][7] - The first layer focuses on environmental perception, the second on risk cognition and prediction, and the third on adaptive safety decision-making [6][7] - This comprehensive approach aims to create a cognitive autonomous driving system capable of handling complex and unknown scenarios effectively [6][7] Group 6 - The future of intelligent vehicles is expected to evolve from rule-driven and data-driven approaches to a cognitive-driven model, enhancing generalization and safety in unknown and long-tail scenarios [7][8] - A "three verticals and three horizontals" technical architecture is proposed to support the systematic evolution of intelligent vehicles, focusing on key vehicle technologies, advanced information technologies, and foundational support technologies [8][9] - The emphasis is on ensuring that "smart cars" are also "safe cars," necessitating a transition to a brain-like cognitive architecture for intelligent vehicle safety [9]

智能汽车的安全与可靠性，始终是智能化进阶的核心命题。清华大学车辆与运载学院院长王建强在论坛 上表示，当前智能汽车正从低等级辅助驾驶向高等级自动驾驶跨越，但复杂长尾场景下的事故频发，暴 露出"感知-认知-决策"全链路的技术短板。 "特斯拉事故反映感知系统误检，Waymo事故暴露风险认知不足，Uber事故源于决策误判，"他直言，规 则驱动与数据驱动的传统路线已难以满足L4+级别自动驾驶需求。前者依赖预设规则，难以应对开放场 景；后者陷入"黑箱决策"，泛化能力受限于训练数据。 为此，王建强提出"认知驱动"的第三条技术路线：融合规则驱动的可解释性与数据驱动的学习能力，借 鉴人类基于常识与经验的决策模式，构建"感知-认知-决策"闭环体系。例如，在感知层融合物理状态估 计与语义理解，提升复杂环境重构精度；在认知层通过"行车安全场"建模与深度学习结合，实现风险动 态预测；在决策层引入大语言模型推理，支持类人自适应决策。这一路线已在部分场景验证其优势，能 显著提升智能汽车对极端场景的应变能力。 中国汽车工业协会日前发布的数据显示，今年上半年，新能源汽车产销分别完成696.8万辆和693.7万 辆，同比分别增长41.4%和40 ...

Core Viewpoint - Intelligent automotive technology is a key solution for traffic safety, which remains a perpetual theme in the development of smart vehicles [1] Group 1: Current State of Intelligent Vehicles - Low-level intelligent vehicles have achieved a high market penetration rate, but accidents still occur as the industry transitions to higher levels of autonomous driving [1] - There are significant challenges in safety technology that need to be addressed in the context of complex long-tail scenarios [1] Group 2: Technological Approaches - The early development of intelligent vehicles relied on rule-driven approaches, while current mainstream autonomous driving methods include data-driven techniques [4] - Rule-driven systems are observable and interpretable but are inflexible in complex environments, whereas data-driven systems utilize deep learning but suffer from a "black box" nature that obscures decision-making processes [4] - A proposed third route, "cognitive-driven," aims to combine the interpretability of rule-driven systems with the learning capabilities of data-driven systems, enhancing adaptability and transparency [4][5] Group 3: Cognitive-Driven Architecture - The cognitive-driven approach is based on a deep understanding of the interactions between humans, vehicles, and roads, leading to accurate modeling and digital representation of system characteristics [5] - The architecture consists of three layers: perception, cognition, and decision-making, integrating physical state estimation, semantic understanding, and human-like adaptive decision generation [5][6] Group 4: Future Trends and Goals - The evolution of autonomous driving is shifting from rule-driven and data-driven methods to cognitive-driven systems, focusing on human-like cognition, learning, and evolution [5] - A new paradigm of "self-learning + prior knowledge" is necessary to enhance environmental understanding and reasoning capabilities, improving safety and generalization in long-tail scenarios [5] - The ultimate goal is to develop a high-level intelligent driving system that possesses self-learning, self-reflection, and adaptive capabilities, ensuring safety and verifiability [6]