如果您有任何科研辅导需求，欢迎联系我们！ 自驾方向： 大模型、VLA、端到端自动驾驶、3DGS、BEV感知、目标跟踪、毫米波雷达视觉融合、激光视觉融合、 多传感器标定、多传感器融合、车道线检测、在线地图、轨迹预测、世界模型、3D目标检测、Occupancy、高性能计 算、NeRF、语义分割、决策规划等。 具身方向： VLA、视觉语言导航、端到端、强化学习、Diffusion Policy、sim2real、具身交互、抓取点预测与位姿估 计、机器人决策规划、运动规划、3DGS、SLAM、触觉感知、双足/四足机器人、遥控操作、零样本学习等； 3D视觉 相关： 点云处理、3DGS、SLAM等； 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 今年的CVPR和ICCV工作陆续放出，从今年的2大顶会来看热点研究方向，主要集中在四个部分：通用cv、自动驾驶 相关、具身相关、3D视觉相关。下面是4个领域中细分的子方向。 计算机视觉与图像：diffusion、图像质量评估、半监督学习、零样本、开放世界检测等； 自动驾驶相关，集中在端到端、闭环仿真3DGS、多模态大模型、扩 ...

下周又要看关税后续了 90天到了。 目前越南的20% ，周五市场解读为利好。 不知道后续nv卡 等有啥招数。 唯一不变的是 三四季度 一批asic卡要面世了。 国内卡荒将得到一定缓解。 但是nv的政策，现在也不知道走向何处？？？ 1）Anthropic年化收入达到40亿美元，较年初增长近四倍； 2）Cursor通过引入Anthropic的高管增强市场竞争力； xAI控制台新增Grok4及Grok4Code引用， 全球第二Anthropic年化收入已达40亿美元，较年初增长近4倍、与 Cursor 竞争加剧 海外基建 落地： 阿里云将设立首个AI全球能力中心，并在马来西亚、菲律宾新增数据中心 阿里云将设立首个AI全球能力中心，计划联合超过1000家企业打造10余个行业AI示范项目，并与全球 120多所高校合作，每年培养10万名AI人才。 接下来还是 asic逻辑和稳定币还是不受关税影响。其它不知道。 1）Grok4作为xAI的旗舰模型，专注于自然语言处理、数学推理和综合推理能力的提升； 2）Grok4Code专为编程优化，计划与代码编辑器无缝整合，提高开发效率； 3）xAI通过API提供Grok4访问权限，未来 ...

本周专题：特斯拉 Q2 交付下滑 小鹏 G7 正式上市 汽车 2025 年 07 月 06 日 特斯拉 Q2 交付下滑 小鹏 G7 正式上市 行 业 7 月 2 日，特斯拉公布其今年第二季度的全球销量达 384122 辆， 较去年同期的 443956 辆下降 13%，已连续第二个季度出现下滑。这一 下滑主要归因于公司更实惠车型上市延迟、首席执行官埃隆·马斯克 的政治立场争议以及在美国市场销量的大幅下滑。 定 期 报 告 特斯拉目前面临着多重困境交织的局面。美国国会通过巨额减税 与支出法案，其中取消电动车税收抵免政策的部分，将对特斯拉构成 潜在冲击。同时面对通用汽车、大众、宝马、比亚迪、小米等推出的 新车型，特斯拉沿用多年的产品线正在承受越来越大的竞争压力。而 面对汽车业务的疲软，特斯拉积极挖掘自动驾驶软件、无人驾驶网约 车服务和人形机器人 Optimus 的发展潜力。 7 月 3 日，小鹏汽车正式发布全新 AI 智能家庭 SUV——小鹏 G7。 新车共推出 602 长续航 Max、702 超长续航 Max、702 超长续航 Ultra 三个版本，官方指导售价分别为 19.58 万元、20.58 万元和 22. ...

Core Viewpoint - Huawei's announcement of the open-source Pangu Pro MoE model with 720 billion parameters and the Pangu model with 70 billion parameters has sparked industry discussions regarding the similarities between the Pangu Pro MoE model and Alibaba's Tongyi Qwen-2.5 14B model in terms of parameter structure [1][2]. Group 1 - The Noah's Ark Lab stated that the Pangu Pro MoE model is developed and trained based on the Ascend hardware platform and is not an incremental training of other vendors' models [1]. - A GitHub study found a high correlation of 0.927 in attention parameter distribution between the Pangu Pro MoE model and the Tongyi Qwen-2.5 model, indicating significant structural similarities [1]. - Noah's Ark Lab clarified that while some code implementations of the Pangu Pro MoE model reference industry open-source practices, they strictly adhere to open-source license requirements and clearly mark copyright statements [1][2]. Group 2 - Industry analysts suggest that the Pangu Pro MoE model likely did not use the pre-trained weights of the Tongyi Qwen-2.5 model as initialization parameters, as there are essential differences in the absolute value distribution of biases [2]. - The structural consistency between the two models may stem from shared architectural design principles, which is common in large models as good structures are widely adopted [2]. - Noah's Ark Lab emphasized that the Pangu Pro MoE model features key innovations, being the first mixed expert model designed for the Ascend hardware platform, and introduced a grouped mixed expert model (MoGE) architecture to enhance training efficiency [2]. Group 3 - Noah's Ark Lab expressed gratitude to global developers and partners for their support of the Pangu model and highlighted the importance of constructive feedback from the open-source community [3]. - The lab aims to optimize model capabilities through collaboration with like-minded partners, accelerating technological breakthroughs and industry applications [3].

Core Viewpoint - Longqi Technology, a leading ODM manufacturer in smart hardware, has submitted its listing application to the Hong Kong Stock Exchange after just over a year on the Shanghai Stock Exchange, aiming to leverage its strong market position and diverse product offerings in the consumer electronics sector [1][2]. Group 1: Business Model and Market Position - Longqi Technology provides comprehensive ODM services, integrating hardware design, software development, testing, and supply chain management, with a diverse product range including smartphones, AI PCs, automotive electronics, tablets, smartwatches, and smart glasses [1]. - The company is the second-largest consumer electronics ODM globally and ranks first in smartphone ODM shipments as of 2024 [1]. - Major clients include leading brands such as Xiaomi, Samsung, Lenovo, Honor, OPPO, and vivo [1]. Group 2: Financial Performance - Revenue figures for Longqi Technology from 2022 to 2024 are as follows: 29.343 billion RMB, 27.185 billion RMB, and 46.382 billion RMB, while net profits were 562 million RMB, 603 million RMB, and 493 million RMB respectively [1][4]. - The company's gross profit increased over the same period, but the gross margin showed a decline, with figures of 8.1%, 9.5%, and 5.8% for the respective years [4][5]. - In Q1 2025, the company reported a revenue of 9.378 billion RMB, a year-on-year decrease of 9.27%, but net profit increased by 20.33% to 154 million RMB, indicating a recovery in profitability [4]. Group 3: Strategic Growth Initiatives - Longqi Technology has adopted a "1+2+X" strategy to align with new market demands and technological innovations, focusing on smartphones as the core, with personal computing and automotive electronics as key growth areas, and expanding into emerging consumer electronics [2]. - The AIoT and other product segments have shown continuous revenue growth, with figures of 1.887 billion RMB, 2.511 billion RMB, and 5.573 billion RMB from 2022 to 2024, increasing their revenue share from 6.5% to 12% [3][4]. Group 4: Industry Trends and Future Outlook - The rise of AI technology is expected to create significant growth opportunities for Longqi Technology, particularly in the smartphone and PC markets, where AI penetration rates are projected to reach 18% and 32% respectively in 2024 [6]. - Longqi Technology has been proactive in developing AIoT products, launching over 10 million smartwatches in 2023 and achieving significant sales in smart glasses and VR products since 2015 [6][7]. - The company is well-positioned to benefit from the anticipated growth in AI hardware products, which could enhance its performance and appeal to growth-oriented investors in the Hong Kong capital market [8].

本文的主要作者来自北京航空航天大学、北京大学和北京智源人工智能研究院。本文的第一作者为北京航空航天大学硕士生周恩申，主要研究方向为具身智 能和多模态大模型。本文的共一作者兼项目负责人为北京智源研究院研究员迟程。本文的通讯作者为北京航空航天大学副教授盛律和北京大学计算机学院研 究员、助理教授仉尚航。 机器人走出实验室、进入真实世界真正可用，远比想象中更复杂。现实环境常常杂乱无序、物体种类繁多、灵活多变，远不像实验室那样干净、单一、可 控。 想象一下，你正在餐厅吃饭，身边有个服务机器人。你对它说： 「 把第二列最远的黄色寿司盘，放到离我最近的寿司和酱油碟之间的空位上。 」 （左图） 又或者，你希望它 「 拿起最左边、饮料 logo 正对的苹果，放到最近的桌子上，并与之前的苹果排成一排、间距一致。 」 （右图） 这些听起来是我们日常再熟悉不过的指令，其实是一个典型 空间指代（Spatial Referring） 任务。简单来说，就是让机器人通过 「 最远 」 「 第二列 」 「 等间距 」 「 正对着」这 类空间关系，搞清楚要抓哪个对象、放在哪里、或者走向哪个位置。 听着简单，做起来却不容易。哪怕是目前最强大、最先 ...

Core Viewpoint - The RuiPath pathology model, developed by Ruijin Hospital in collaboration with Huawei, aims to enhance the efficiency and accuracy of pathology diagnostics in China by leveraging AI technology [1][5]. Group 1: Model Development and Features - The RuiPath model is a clinical-grade multimodal pathology model that covers 90% of the annual cancer incidence in China, addressing 19 common cancer types and hundreds of auxiliary diagnostic tasks [1][5]. - The model has achieved state-of-the-art (SOTA) performance in 7 out of 14 auxiliary diagnostic tasks tested against 12 mainstream public datasets, surpassing the performance of Harvard's UNI2 model [4]. - The model's core "visual foundation model" was developed using over one million high-quality digital pathology slides from Ruijin Hospital, utilizing Huawei's AI toolchain for annotation, training, and fine-tuning [2][4]. Group 2: Efficiency and Impact - The implementation of the RuiPath model allows pathologists to increase their daily workload from 200-300 slides to 400-500 or more, significantly improving diagnostic efficiency [5]. - The model aims to standardize digital pathology practices across hospitals in China, enabling easier deployment and reducing training costs for other institutions [5][10]. - The collaboration between Ruijin Hospital and Huawei has streamlined the model training process, allowing for the completion of the RuiPath model development with only a 16-card cluster, making it more accessible for hospitals [10][11]. Group 3: Industry Challenges and Solutions - There is a significant shortage of pathology doctors in China, with only about 20,000 available and a gap of 140,000 needed, highlighting the importance of AI solutions in addressing this challenge [5]. - The partnership has evolved through two phases: digitalization and smart pathology, focusing on data standardization and collaborative model development [7][8]. - The use of Huawei's ModelEngine has transformed the annotation process, allowing pathologists to review over 700 slides in a day, thus enhancing both efficiency and accuracy [10].

注意看，这张图画的是谁？ 如果不够明显，那再眯着眼睛看呢。没错，就是蒙娜丽莎。 这张来自日本艺术家北冈秋吉前几天创作的一张图，结果直接把一众大模型难倒了。 ChatGPT只能分辨出这是一张脸。。 Gemini则是直接识别错了人。 白交 发自 凹非寺 量子位 | 公众号 QbitAI 又一个让大模型几乎全军覆没的难题出现了。 大模型不会眯眼睛？ 为了验证这一推论，咱们也进行了一波简单实测。既然直接问「画的是谁」问不出来，那给一点提示，让它可以尝试眯着眼睛。 网友：懂了，大模型不会眯眼睛。 首先来看ChatGPT的表现。 它能准确判断出这是一张视觉错位图，也通过「眯着眼睛」识别出来了具体的脸部轮廓，结果在最后关头识别错了。 再来看看国产玩家的表现，还是挑出三个来看看。 豆包的回答与Gemini类似，能够判断出图像的风格、人脸轮廓，但是无法判断具体任务。 但选择 深度思考 之后，它深入研究了故障艺术的通道分离特性，然后从黑色轮廓中判断出这是 爱因斯坦 （？？？？ 而再让它深度思考一下，答案给不出来了。 Gemini的回答则还停留在各种颜色的条纹，也辨认出这是个侧脸的轮廓。 Grok则直接表示无法辨认，让我提供一张更 ...

中新网武汉7月6日电 (马芙蓉王驰骋)2025年湖北省海峡两岸数字赋农助力乡村振兴论坛6日在武汉举 行，来自两岸农业领域的高校专家、行业协会代表、企业代表等100多人，聚焦智慧农业分享发展经 验，寻找合作机遇。 此次论坛召开前，主办方走访在鄂台企，收集其在科技研发、农业生产方面的困难与需求，并联动专家 团队开展定向帮扶。论坛现场，华中农业大学、湖北省农业科学院等4家科研团队与4家台企代表，签署 了技术合作供需对接协议。 "这次参访交流发现湖北有丰富的农业资源及智慧农业设备。"台湾富丽农村发展基金会董事长龚世明介 绍，他看中湖北一家企业的植物智能培育设备，正洽谈采购细节。湖北"科技小院"人才培养模式，也让 他深受启发。 此次论坛是第十九届湖北·武汉台湾周交流活动之一，由湖北省科协主办。(完) （文章来源：中国新闻网） 湖北是农业大省，素有"鱼米之乡"美誉。借助数字技术、人工智能等新兴技术，湖北推进智慧农业发 展，在基础设施建设、技术应用、产业融合等方面，已取得系列成果。 此次论坛上，发布了湖北省智慧农业领域科技成果，促进供需对接。20多项成果涉及水稻、油菜、棉 花、番茄、柑橘、鱼类、生猪等农产品、水产品，涵盖生 ...

Core Insights - The article highlights the recent trends in financing events within various industries, indicating a total of 108 financing events this week, with a total amount of approximately 11.37 billion RMB, marking an increase from the previous week [4][12]. Financing Overview - This week, there were 108 financing events (excluding mergers and acquisitions), an increase of 4 from last week, with a total financing amount of approximately 11.37 billion RMB, up by 4.64 billion RMB [4]. - Among these, 28 events had financing amounts of 100 million RMB or more, an increase of 6 from last week [4]. - There were 32 public exit cases this week, a decrease of 4 from last week [4]. Financing Amount Distribution - The distribution of disclosed financing events shows 70 events, with amounts ranging as follows: - 5 events below 5 million RMB - 22 events between 5 million and 10 million RMB - 13 events between 10 million and 50 million RMB - 18 events between 50 million and 100 million RMB - 9 events between 100 million and 500 million RMB - 2 events between 500 million and 1 billion RMB - 1 event over 1 billion RMB [5]. Notable Investment Events 1. **Photovoltaic Materials**: Tongwei Co., Ltd. announced that its wholly-owned subsidiary, Yongxiang Co., completed a strategic financing round of approximately 4.916 billion RMB, with funds primarily used for repaying financial institution debts and supplementing working capital [7]. 2. **AI Large Models**: Zhizhu completed a strategic financing round of 1 billion RMB, aimed at building an open platform for model-as-a-service (MaaS) and enhancing AI infrastructure in Shanghai [8]. 3. **AI Chips**: Zhenliang Intelligent announced nearly 1 billion RMB in strategic financing, focusing on high-performance GPU development [9]. 4. **Autonomous Driving Software**: Tiantong Vision completed a D round financing of 500 million RMB, aimed at promoting Robotaxi deployment nationwide [9]. Industry Distribution - The financing events this week spanned 13 industries, with the top five being: - Information Technology: 26 events - Electronics: 25 events - Equipment Manufacturing: 14 events - Healthcare: 10 events - Materials: 9 events [12]. - In terms of financing amount, the top five industries were: - Materials: 49.16 billion RMB (due to Yongxiang's large financing) - Electronics - Information Technology - Equipment Manufacturing - Automotive [14]. Regional Distribution - The top five regions for financing events were: - Guangdong: 17 events - Jiangsu: 17 events - Zhejiang: 16 events - Shanghai: 13 events - Beijing: 13 events - These regions accounted for 70.37% of all financing events [19][20]. Financing Round Distribution - The distribution of financing rounds showed that seed and angel rounds accounted for 59 events, while strategic financing ranked third with 21 events. Early-stage financing (A round and earlier) accounted for 58.33% of the total events [23]. Investment Institutions - A total of 97 investment institutions participated this week, with notable activity from Beijing Guoguan and Jinpu Investment [26].