编辑｜Panda 2024 年底，硅谷和北京的茶水间里都在讨论同一个令人不安的话题： Scaling Law 似乎正在撞墙 。 那时候，尽管英伟达的股价还在狂飙，但多方信源显示，包括彼时备受期待的 Orion（原计划的 GPT-5）在内，新一代旗舰模型在单纯增加参数规模和训练数据 后，并未展现出预期的边际效益提升。另外，也有研究认为预训练所需的数据将会很快耗尽，其甚至还预测了明确的时间节点：2028 年。 来自论文 arXiv:2211.04325v2 OpenAI 和 Safe Superintelligence Inc 的联合创始人 Ilya Sutskever 当时还留下了一句意味深长的判词：「2010 年代是规模扩大的时代，现在人们又回到了奇迹和发 现的时代。」这句话在当时被许多人解读为悲观的预警，也就是单纯依靠堆砌算力和数据的预训练路线，恐怕已经触到了天花板。 直到 2025 年初，接连的惊喜打破了僵局。 那时候，OpenAI 的 o1 模型已在几个月前率先引入了强化推理，展示了模型在思考时间换取智能深度上的惊人潜力，证明了 test-time scaling（测试时间扩展）是一 条通往更高智能的可 ...

百架无人机在夜空精准勾勒流光图案，工业机械臂协同完成毫米级精密加工，智能车流控制技术有效避 免城市拥堵……这些充满未来感的场景，正依托群体智能技术逐步成为现实。然而，设备大规模协同运 作的背后，高功耗、通信延迟、突发故障等问题，始终制约着群体智能系统的可靠性与持久运行。 群体智能系统的稳定运行，离不开通信、控制、运维三大核心环节的协同支撑。其中，无线通信是群体 协同的"神经"，但持续的全量数据传输极易引发网络拥堵与能耗过高问题。 团队研发的"事件触发滤波"技术，革新了群体智能无线通信时的传统模式。团队成员、南工大副教授岳 冬冬介绍，该系统仅在关键数据变化超越设定阈值时才启动传输，其余时间保持静默。这一设计在保障 控制精度的前提下，可降低70%—80%的通信量与能耗，大幅延长无人机等移动平台的连续作业时 长。"它可让设备从'时刻待命'变为'按需发声'，做到智能节流，解决续航之困。"岳冬冬说。 控制则犹如群体智能的"行动中枢"。"设备在复杂多变的环境中运行，难免出现故障，就像编队飞行的 候鸟会突遭意外。"团队成员、南工大教授钱默抒说，面对这种情况，传统处理方式是停机维修或回 收，不仅成本高昂，在紧急救援等时效性要 ...

记者丨林典驰 编辑丨朱益民 12月26日，优必选第1000台工业人形机器人Walker S2下线，下线仪式现场，超200台Walker S2组成气势恢宏的方阵。 后续这些机器人将交付到各地，投入产业一线应用，覆盖汽车制造、智能制造、智慧物流、具身智能数据中心等重点领域。 2025年的优必选交出了硬核答卷 ，Walker S系列已成为全球进入最多车厂实训的工业人形机器人。全年优必选工业人形机器人产能已突破1000 台，交付量超过500台，并计划在2026年将年产能提升至万台规模。 一连串数字背后，没有概念炒作，没有PPT造"机器人"，只有两年来数百台机器人在新能源汽车厂、3C产线和物流仓库中日复一日的真实实 训。 作为国内最早押注人形机器人赛道的"长期主义者"，当行业在2025年突然被资本与政策点燃，周剑和他的优必选却选择踩下"理性油门"，从工 业制造的实训入手，连同客户解决真实需求，主动收缩初期泛化需求，聚焦搬运、分拣、质检三个工位。 "这三类工位是我们认为存在着人员流失高、管理难度大、工作重复性强、工位相对结构化等特点，不仅是客户迫切需要人形机器人提供能力 协同的领域，也是人形机器人现阶段更适合发挥能力的领 ...

Core Viewpoint - UBTECH has successfully launched its 1000th industrial humanoid robot, Walker S2, marking a significant milestone in its production capabilities and commitment to practical applications in various industries [1][2]. Production and Delivery - The Walker S series has become the most widely adopted industrial humanoid robot in global automotive training, with an annual production capacity exceeding 1000 units and a delivery of over 500 units in 2025, aiming to scale to 10,000 units by 2026 [2][4]. - The company emphasizes real-world training and application, focusing on high-demand areas such as handling, sorting, and quality inspection, which are characterized by high employee turnover and repetitive tasks [3][8]. Market Strategy - UBTECH's strategy involves a phased approach: starting with industrial manufacturing, followed by commercial services, and eventually household applications, ensuring a sustainable business model that addresses real customer needs [5][11]. - The company has secured significant orders, including a record single bid of 264 million yuan, primarily from sectors like automotive manufacturing and smart logistics [4][6]. Technological Advancements - Continuous investment in humanoid robot technology, including hardware control, AI, and visual servo systems, has enabled the development of autonomous operational capabilities [5][12]. - The introduction of group intelligence technology allows for collaborative operation among humanoid robots, enhancing efficiency and adaptability in real-world scenarios [12][13]. Cost Management - UBTECH aims to reduce production costs by 20%-30% annually, targeting a unit cost of under $20,000 as production scales up, leveraging advancements in technology and supply chain integration [16][14]. - The company has achieved a 90% localization rate for core components, enhancing its competitive edge in the global market [15]. Industry Context - The humanoid robot industry is expected to face a talent shortage, with a projected demand gap of nearly 30 million workers in key manufacturing sectors by 2025, positioning humanoid robots as a viable solution [9][10]. - Despite the rapid advancements in the industry, UBTECH acknowledges that widespread adoption of humanoid robots in everyday life will take an estimated 5-10 years, necessitating ongoing investment and development [20].

Core Viewpoint - UBTECH has successfully launched its 1000th industrial humanoid robot, Walker S2, which will be deployed in key sectors such as automotive manufacturing, smart manufacturing, and logistics [2][3] Production and Delivery - The company has achieved an annual production capacity of over 1000 units and plans to increase this to 10,000 units by 2026, with over 500 units expected to be delivered in 2025 [2][3][5] - The Walker S series has become the most widely used humanoid robot in training across various automotive factories [2] Market Strategy - UBTECH focuses on practical applications in industrial settings, specifically targeting high turnover, high management difficulty, and repetitive tasks [3][8] - The company has adopted a three-step strategy: starting with industrial manufacturing, followed by commercial services, and finally home companionship [4] Order Acquisition - The company has secured several large orders, including a record single order of 264 million yuan, primarily from the automotive and smart manufacturing sectors [3][4] - The demand for humanoid robots is driven by labor shortages in repetitive and hazardous jobs, with a projected talent gap of nearly 30 million in China's manufacturing sector by 2025 [7] Technological Development - UBTECH has invested in a full-stack technology approach, including hardware control, AI, and visual servo systems, to enhance the capabilities of its robots [4][10] - The introduction of group intelligence technology allows for collaborative work among humanoid robots, significantly improving efficiency [10][11] Cost Management - The company anticipates a 20%-30% reduction in manufacturing costs annually, aiming to keep the cost per unit below $20,000 at scale [14] - UBTECH is focusing on optimizing design, manufacturing processes, and supply chain integration to achieve cost reductions [12][14] Industry Outlook - The humanoid robot industry is expected to see significant growth, with potential market sizes reaching hundreds of billions to trillions [17] - Despite rapid advancements, the widespread adoption of humanoid robots in everyday life is projected to take 5-10 years [17]

Core Insights - The manufacturing cost of humanoid robots is expected to decrease by 20%-30% annually, with a target cost of under $20,000 per unit at a production scale of 10,000 units [15] - The company has achieved significant milestones, including the production of its 1,000th industrial humanoid robot, Walker S2, and plans to increase annual production capacity to 10,000 units by 2026 [1][3] - The company focuses on practical applications in industrial settings, particularly in sectors like automotive manufacturing, smart manufacturing, and logistics, rather than on flashy demonstrations [3][8] Production and Delivery - The company has delivered over 500 units of the Walker S series and aims to exceed 1,000 units in monthly production by 2025 [3][4] - The company has secured several large orders, including a record single order worth 264 million yuan, primarily from the industrial sector [3][4] Technology and Innovation - The company emphasizes the importance of a robust technology stack, including hardware control, AI, and visual servoing, to enhance the capabilities of humanoid robots [4][10] - Innovations such as the Co-Agent technology and group intelligence are being developed to improve collaborative efficiency among humanoid robots [10][11] Market Demand and Strategy - The company identifies high turnover, management challenges, and repetitive tasks in certain job roles as key areas where humanoid robots can provide value [2][8] - The company has adopted a phased strategy, focusing first on industrial applications before expanding into commercial services and home companionship [4][8] Industry Context - The humanoid robot industry is experiencing rapid growth, with a projected market potential reaching hundreds of billions to trillions of yuan [18] - The company acknowledges the need for a sustainable business model that can withstand market fluctuations and capital withdrawal [3][18] Supply Chain and Cost Management - The company has achieved a 90% localization rate for core components, enhancing its competitive edge in the global market [14] - Efforts to control costs include optimizing design, exploring different manufacturing methods, and improving supply chain integration [11][12]

（来源：中国妇女网） 转自：中国妇女网 原标题：从月壤到"月宫"！中国科学家"解锁"智造月球科研站 阳光取代了窑火，月壤作为原始建材，智能机器人充当建筑工人……这些正是我国航天领域最前沿的探 索方向之一：月球原位自主智造。 12月19日，由深空探测实验室承办的以"地外资源开发利用技术前沿与发展战略——太空采矿与深空制 造"为主题的中国工程院工程科技学术研讨会在安徽省亳州市举行，多位院士专家详解了"月宫"建造"黑 科技"。 随着嫦娥六号完成月背采样返回，中国探月工程"绕、落、回"三步走战略圆满收官。如何利用月球本身 资源建设可持续的科研站，成为下一阶段的关键课题。 东华大学科研团队依据嫦娥五号取回的真实月壤，在实验室通过高温熔融和真空牵引技术，成功制备出 直径仅10至20微米的超细月壤连续纤维。中国科学院院士、东华大学教授朱美芳说，团队已成功研发适 应月球高真空、低重力环境的自动成纤装备，为未来月面原位制造复合材料开辟了新可能。 "无论是打印成砖，还是拉制成纤，目的都是将月球上最丰富的表层物质——月壤，转化为可用的工程 材料。"深空探测实验室总工程师史平彦表示，多条技术路线并行探索，是为了应对月球极端环境的严 ...

（来源：湖州日报） 科研人员用抛物面镜将太阳光聚焦数千倍，产生超过1300摄氏度的高温，再通过一根柔性的光纤远距离 传输聚光太阳能，就像一支精准的"光笔"，结合3D打印技术，将月壤材料打印出结构坚实的砖体或任 意形状的构件。 "未来月球科研站的建设，核心是'月球原位取材、集群协同智造、自主智能作业'。"中国工程院院士、 哈尔滨工业大学党委书记陈杰说，其目标是转化利用月球的"土"并建成月球的"家"，最大限度降低对地 球补给的长距离依赖，实现地外基地的智能建造、自主运维和可持续拓展。 除了将月壤高温熔融打印成结构件，我国科学家还探索了将月壤制成高性能纤维的新方法。 东华大学科研团队依据嫦娥五号取回的真实月壤，在实验室通过高温熔融和真空牵引技术，成功制备出 直径仅10至20微米的超细月壤连续纤维。中国科学院院士、东华大学教授朱美芳说，团队已成功研发适 应月球高真空、低重力环境的自动成纤装备，为未来月面原位制造复合材料开辟了新可能。 "无论是打印成砖，还是拉制成纤，目的都是将月球上最丰富的表层物质——月壤，转化为可用的工程 材料。"深空探测实验室总工程师史平彦表示，多条技术路线并行探索，是为了应对月球极端环境的严 苛 ...

转自：北京日报客户端 阳光取代了窑火，月壤作为原始建材，智能机器人充当建筑工人……这些正是我国航天领域最前沿的探 索方向之一：月球原位自主智造。 19日，由深空探测实验室承办的以"地外资源开发利用技术前沿与发展战略——太空采矿与深空制造"为 主题的中国工程院工程科技学术研讨会在安徽省亳州市举行，多位院士专家详解了"月宫"建造"黑科 技"。 除了将月壤高温熔融打印成结构件，我国科学家还探索了将月壤制成高性能纤维的新方法。 东华大学科研团队依据嫦娥五号取回的真实月壤，在实验室通过高温熔融和真空牵引技术，成功制备出 直径仅10至20微米的超细月壤连续纤维。中国科学院院士、东华大学教授朱美芳说，团队已成功研发适 应月球高真空、低重力环境的自动成纤装备，为未来月面原位制造复合材料开辟了新可能。 "无论是打印成砖，还是拉制成纤，目的都是将月球上最丰富的表层物质——月壤，转化为可用的工程 材料。"深空探测实验室总工程师史平彦表示，多条技术路线并行探索，是为了应对月球极端环境的严 苛挑战，找到最优解决方案。 月球表面呈现为一个集极端温差、高真空、强辐射以及带电月尘于一体的复杂环境，任何制造设备首先 要解决长期可靠运行的难题。 ...

" 当Transformer走到尽头，面对的是参数和规模的极限，端侧群 体智能或许是AI发展的另一条出路。 " 作者丨张夏宁 编辑丨胡敏 以下是邹佳思的现场演讲内容，雷峰网作了不改变原意的编辑及整理： 大家好，今天我想探讨一个与大家生活密切相关的主题——设备端的智能。 12月12日， 第八届 GAIR 全球人工智能与机器人大会 于深圳正式拉开帷幕。 本次大会为期两天，由GAIR研究院与雷峰网联合主办，高文院士任指导委员会主席，杨强院士与朱晓蕊教 授任大会主席。 作为AI产学研投界标杆盛会，GAIR自2016年创办以来，始终坚守"传承+创新"内核，是AI学界思想接力的 阵地、技术交流的平台，更是中国AI四十年发展的精神家园。过去四年大模型驱动AI产业加速变革，岁末 年初GAIR如约而至，以高质量观点碰撞，为行业与大众呈现AI时代的前沿洞见。 本次峰会之上，RockAI CMO邹佳思以"摆脱Transformer的束缚，让智能重新定义硬件"为主题，为参会 者们带来了一场精彩纷呈的演讲。 邹佳思通过提出一个生活场景的假设作为开场，描绘了一幅充分开发端侧智能后的日常生活图景。在设想 中，家庭场景中智能设备的联动无需云 ...