作者 | 周智宇 2026年的春节前夕，中国互联网的火药味比往年更浓烈了。 2月10日，华尔街见闻从百度内部人士处获悉，百度内部已于近期悄然启动代号为"O计划"的项目。不 同于以往的部门协作，这次由搜索和云团队联合办公，意味着项目打破部门藩篱，MEG与ACG联合作 战。 另有知情人士透露，这次合作的项目和百度App相关。百度瞄准用户的日常场景，将以百度App为枢 纽，借助文心助手，调动百度内部生态及合作伙伴的服务，解决用户的实际需求。 这一系列动作释放一个强烈信号。百度意识到，AI入口的竞争不再仅仅是前端UI的竞争，而是算力成 本、推理速度与数据调度能力的综合战争。 说白了，UI只是一个对话框，谁都能做；但真正拉开代际差距的，是背后的调度逻辑与算力账本。当 用户提出一个复杂需求，系统需要瞬间调动全生态的服务。如果调度能力差，AI就只是个会聊天的花 架子，没法真正解决问题；而如果算力效率低，单次回答的成本就会贵得离谱，商业上根本跑不通。 百度启动"O计划"并非没有征兆。百度创始人李彦宏在2026年初的一次内部分享中明确定调，训练的目 的是推理。 这意味着，推理效率将决定搜索的代际竞争。如果百度不能通过搜索与云的 ...

Core Viewpoint - The article discusses the WHALES dataset, which aims to enhance cooperative perception and scheduling in autonomous driving, addressing the limitations of existing single-vehicle systems in non-line-of-sight scenarios [2][3][4]. Group 1: WHALES Dataset Overview - WHALES (Wireless enHanced Autonomous vehicles with Large number of Engaged agentS) is the first large-scale dataset designed for evaluating communication perception agent scheduling and scalable cooperative perception in vehicular networks [4]. - The dataset integrates detailed communication metadata and simulates real-world communication bottlenecks, providing a rigorous standard for evaluating scheduling strategies [4]. - WHALES includes 70,000 images, 17,000 frames of LiDAR data, and over 2.01 million 3D annotations, making it a comprehensive resource for research in cooperative driving [14][29]. Group 2: Key Features and Contributions - The dataset supports V2V (Vehicle-to-Vehicle) and V2I (Vehicle-to-Infrastructure) perception, optimizing the CARLA simulator for speed and computational cost, achieving an average of 8.4 cooperative agents per driving scenario [14][29]. - WHALES introduces a novel Coverage-Aware Historical Scheduler (CAHS) algorithm, which prioritizes agents based on historical coverage, outperforming existing methods in perception performance [4][19]. - The dataset allows for the evaluation of various scheduling algorithms, including Full Communication, Closest Agent, and the proposed CAHS, enhancing the understanding of cooperative perception tasks [19][27]. Group 3: Experimental Results - Experiments conducted on the WHALES dataset demonstrated that cooperative models significantly outperform standalone models in 3D object detection, with F-Cooper improving mAP by 19.5% and 38.4% at 50m and 100m detection ranges, respectively [25]. - The CAHS algorithm showed superior performance in both single-agent and multi-agent scheduling scenarios, indicating its effectiveness in enhancing cooperative driving safety [27][28]. - The dataset's design allows for a linear increase in time cost with the addition of agents, making it feasible for large-scale simulations [14][29].