2025年8月18日清晨8点20分，上海浦东新区的柏油路面已经蒸腾起了热浪。地表 61℃（实测气温37℃）的 高温下，一台全尺寸人形机器人迈过最后一个减速带，完成了连续 24 小时的自主行走。 镜头扫过一旁更换了 73 人次的摄影师团队，再切回机器人平稳从容的步态。 这场被称为 "夏日 CityWalk"的 直播，不仅创下全球首次人形机器人高温户外极限挑战的纪录，更在直播收官时抛出重磅消息：智元机器人 六大产品线同步登陆智元商城与京东商城。 从实验室里的 "摔跤常客"到街头 24 小时不关机的 "硅基行者"，从单一技术 Demo 到覆盖工业、服务、科研 的产品矩阵，智元的这一步跨越，或许正标志着人形机器人行业从 "概念热炒" 迈向 "实用化落地"的关键拐 点。 从24小时到720小时的极限挑战 这场直播的戏剧性，藏在一组对比数据里：实测37℃的日间高温、低至 22℃的凌晨温差，柏油、砖石、积 水等 7 种路面材质，锥桶、减速带、突发行人等 12 类障碍 —— 这不是实验室里的理想环境，而是机器人真 正走向商业场景必须面对的 "真实世界考题"。 "去年机器人走几步就摔，是实验室产品；今年量产机寿命达几千小时，成 ...

Core Viewpoint - The article discusses the development of the RLVMR framework by Tencent's Hunyuan AI Digital Human team, which aims to enhance the reasoning capabilities of AI agents by rewarding the quality of their thought processes rather than just the outcomes, addressing inefficiencies in long-horizon tasks and improving generalization abilities [4][26]. Group 1: Challenges in Current AI Agents - Many AI agents succeed in tasks but rely on luck and inefficient trial-and-error methods, leading to a lack of effective reasoning capabilities [2]. - The low-efficiency exploration problem arises as agents often engage in meaningless actions, resulting in high training costs and low reasoning efficiency [2]. - The generalization fragility issue occurs because strategies learned through guessing lack a logical foundation, making them vulnerable in new tasks [3]. Group 2: RLVMR Framework Introduction - RLVMR introduces a meta-reasoning approach that rewards good thinking processes, enabling end-to-end reinforcement learning for reasoning in long-horizon tasks [4][6]. - The framework allows agents to label their cognitive states, enhancing self-awareness and tracking their thought processes [7]. - A lightweight verification rule evaluates the quality of the agent's thinking in real-time, providing immediate rewards for good reasoning and penalizing ineffective habits [8]. Group 3: Experimental Results - The RLVMR-trained 7B model achieved a success rate of 83.6% on the most challenging L2 generalization tasks in ALFWorld and ScienceWorld, outperforming all previous state-of-the-art models [11]. - The number of actions required to solve tasks in complex environments decreased by up to 28.1%, indicating more efficient problem-solving paths [13]. - The training process showed faster convergence and more stable strategies, significantly alleviating the issue of ineffective exploration [13]. Group 4: Insights from RLVMR - The introduction of a reflection mechanism allows agents to identify problems and adjust strategies rather than blindly retrying, leading to a significant reduction in repeated actions and an increase in task success rates [19]. - Rewarding good reasoning habits establishes a flexible problem-solving framework that enhances generalization capabilities in unseen tasks [20][21]. - The two-phase training process of cold-start SFT followed by reinforcement learning aligns with cognitive principles, suggesting that teaching agents how to think before allowing them to learn from mistakes is more efficient [22][24]. Group 5: Conclusion and Future Outlook - RLVMR represents a paradigm shift from outcome-oriented to process-oriented training, effectively addressing the challenges of low-efficiency exploration and generalization fragility in long-horizon tasks [26]. - The ultimate goal is to develop AI agents capable of independent thinking and rational decision-making, moving beyond mere shortcut-seeking behaviors [26][27].

闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 蛋白质模型的GPT时刻来了！ 清华大学智能产业研究院（AIR）周浩副教授课题组联合上海人工智能实验室发布了 AMix-1 ： 首次以 Scaling Law、Emergent Ability、In-Context Learning和Test-time Scaling的系统化方法论 来构建蛋白质基座模型。 这为通往蛋白质的通用智能构建起了新的技术范式。让停留在BERT时代、缺乏可扩展性和通用性的蛋白质基座领域实现了向GPT时代的跨 越。 就像NLP领域的ChatGPT一样，AMix-1不再局限于某一种蛋白质任务，而是能举一反三、自主学习。 而在GPT时代，系统化的讨论逐渐显现，通用智能的爆发也因此开始。 然而，在蛋白质基座领域，几乎没有贯彻这条智能涌现的路径，一系列工作同样停留在BERT时代， 在"预训练+任务微调"这一范式下前行， 缺乏对可扩展性、通用性和涌现能力的系统化讨论 。 | Model | | Scaling Law | Emergent Ability | In-Context Learning | Test-time Scaling | | ...

Group 1: Future Industry Development in Anhui - Anhui is accelerating the layout of a future industry system focusing on innovation, quality project introduction, and talent cultivation, with an aim to form a competitive and self-controlled industry by 2027, targeting a scale of over 200 billion yuan and 500 billion yuan by 2030 [1][2] - The province is concentrating on seven key areas including quantum technology, aerospace information, general intelligence, low-carbon energy, life sciences, advanced materials, and future networks, along with third-generation semiconductors and blockchain [1][2] - The future industry in Anhui is characterized by high technological content, long transformation cycles, significant R&D investment, and high policy demand [1][2] Group 2: Quantum Technology and Nuclear Fusion - Anhui has established a strategic action plan for the commercial application of nuclear fusion energy, aiming for a three-step development strategy involving experimental, engineering, and commercial reactors [2][3] - The province has nearly 60 nuclear fusion energy enterprises, covering the entire industry chain from superconducting wire production to main equipment manufacturing [3] - Quantum technology is being applied in various fields, with over 70 quantum industry chain enterprises in Hefei, focusing on quantum computing, communication, and precision measurement [4][5] Group 3: Talent Development and Youth Involvement - The influx of young talent into future industries is notable, with companies like Zhixiang Future actively recruiting professionals from diverse backgrounds to enhance their AI capabilities [5][6] - Local enterprises are implementing training programs to develop skilled workers, with a focus on integrating local youth into the workforce [6][7] - The aerospace sector in Hefei is also expanding, with initiatives aimed at nurturing young scientists and providing them with practical applications in the industry [6][7] Group 4: Industry Integration and Innovation - Companies in Anhui are increasingly interconnected, with products becoming essential components in future industry supply chains, such as the IC carrier board produced by Chip聚德科技 [7][8] - The development of ultra-thin flexible glass by 中建材玻璃新材料研究院集团 showcases the province's innovation capabilities, with a significant proportion of the research team being young professionals [8][9] - The government is supporting future industries through the implementation of the "Anhui Province Future Industry Development Action Plan," which includes measures for technology innovation and industry clustering [9]

Group 1 - Tesla CEO Elon Musk predicts that AI intelligence will surpass individual human intelligence in less than two years and exceed the total human intelligence in about five years [2] - Musk emphasizes the current AI capabilities have surpassed most humans but not the top individuals or specialized teams, indicating a trajectory of "accelerating returns" driven by improvements in computing power, algorithms, and data [2] - The AI industry is rapidly transforming the world, with breakthroughs in large models enabling machines to possess language, vision, and reasoning capabilities, leading to trillion-dollar applications in areas like autonomous driving and smart manufacturing [3] Group 2 - The US and China are leading the global AI race, holding over 80% of AI patents and 90% of unicorn companies, with the US excelling in foundational research and hardware ecosystems, while China focuses on application-driven innovation [3] - As of Q1 2024, China's AI core industry scale is nearing 600 billion RMB, with a total of 478 large AI models released, ranking second globally after the US [6] - Experts suggest that AI technologies, particularly large models, are crucial for driving high-quality economic development in China, advocating for increased investment in foundational research to create a virtuous cycle between AI research and application [6]

Core Viewpoint - The article discusses the emergence of World Labs, a startup founded by AI expert Fei-Fei Li, focusing on developing the next generation of AI systems with "spatial intelligence" and world modeling capabilities. This shift signifies a new direction in AI development beyond traditional language models [2][3]. Group 1: Company Overview - World Labs was founded in 2024 by Fei-Fei Li and has quickly raised approximately $230 million in funding, achieving a valuation of over $1 billion, making it a new unicorn in the AI sector [2]. - The company has attracted significant investment from major players in the tech and venture capital space, including a16z, Radical Ventures, NEA, Nvidia NVentures, AMD Ventures, and Intel Capital [2]. Group 2: Importance of World Modeling - Fei-Fei Li emphasizes the importance of world modeling, which refers to AI's ability to understand the three-dimensional structure of the real world, moving beyond mere language processing [9][10]. - The concept of world modeling is likened to how humans perceive and interact with their environment, integrating visual, spatial, and motion information to create a comprehensive understanding of the world [10][12]. Group 3: Key Technologies for World Modeling - Several key technologies are being explored to enable AI to understand and reconstruct three-dimensional worlds, including: - Neural Radiance Fields (NeRF), which allows AI to reconstruct a 3D world from 2D images [17]. - Gaussian Splatting, which enhances rendering speed and efficiency for real-time applications [19]. - Diffusion Models, which improve AI's ability to understand and generate three-dimensional content [20]. - Multi-view data fusion, enabling AI to integrate information from various angles to form a complete understanding of objects [21]. - Physics simulation and dynamic modeling, allowing AI to predict and understand the movement and interaction of objects in the real world [23]. Group 4: Applications of World Modeling - The applications of world modeling technology are extensive, including: - In the gaming industry, AI can automatically generate realistic 3D environments from images or videos [25]. - In architecture, AI can quickly create detailed spatial structures, significantly reducing design time [26]. - In robotics, enhancing robots' spatial understanding allows them to navigate and interact with their environment more effectively [26]. - Digital twins can be created for factories, buildings, and cities, enabling simulations for testing and optimization [27]. Group 5: Challenges Ahead - Despite the promising direction of world modeling, several challenges remain: - Data availability is crucial; AI requires extensive and diverse real-world data to learn effectively [31]. - Computational power is a significant barrier, as many current technologies demand high resources, making large-scale deployment challenging [32]. - Generalization ability is limited; AI models often struggle to adapt to unfamiliar environments [33]. Group 6: Future Vision - Fei-Fei Li envisions a future where AI not only sees and reconstructs the world but also participates in it, enhancing human capabilities rather than replacing them [42][43]. - The ultimate goal of AI development is to achieve General Artificial Intelligence (AGI), which requires spatial perception, dynamic reasoning, and collaborative abilities [46][47].

这场争论的价值，或许正是在于撕开技术理想主义的面纱，让行业在狂热中看清现实。 近日，一场被一些业内人士视作是"中国具身智能路线之争"的学术争论，引起了机器人行业的讨论。 一位讨论参与者是许华哲，来自清华大学交叉信息研究院（简称"清华叉院"），是中国科技顶尖人才培养机构的助理教授，同时也是机器人企业星海图的联 合创始人。另一位则是亚洲第一个获得IEEE T-RO最佳论文奖的周博宇，他是获得该机器人领域顶级期刊奖项的"亚洲第一人"，同时也是南方科技大学的助 理教授、博士生导师。知乎平台中，他们围绕"机器人领域特殊任务研究是否有价值"等话题展开了讨论。 整场讨论的原点，是许华哲在知乎上发表了一篇名为《具身智能需要从ImageNet做起吗？》的文章。 许华哲在其中提到，传统机器人学有相当一部分的研究重点在于"特别"的机器人或者"特别"的任务。比如一个蛇形机器人、一个老鼠机器人，或是让机器人 去包饺子、抖落衣服。这类"特殊任务研究"的任务对科学虽然有用，但对"推动具身智能的发展几乎没有用处"。 "我认为这种观点显然不对。"在《具身智能：一场需要谦逊与耐心的科学远征》中，周博宇直接指出，具身智能本身是跨学科产物，它的发展 ...

Core Viewpoint - The establishment of the world's first "Humanoid Robot Intelligence Grading" standard (T/CIE 298-2025) aims to enhance the evaluation framework for humanoid robots, transitioning the industry from a "function-oriented" approach to "intelligent evolution" [1][2][6]. Group 1: Standard Development - The new standard was developed by leading organizations including the Beijing Humanoid Robot Innovation Center and various enterprises and research institutions, addressing the lack of evaluation standards in the humanoid robot industry [1][2]. - The standard introduces a "four-dimensional five-level" evaluation framework, creating a grading system from L1 to L5 for intelligent capabilities [2][3]. Group 2: Evaluation Framework - The evaluation framework consists of four core capability dimensions: Perception and Cognition (P), Decision Learning (D), Execution Performance (E), and Collaboration Interaction (C) [2][3]. - The standard provides 22 primary indicators and over 100 technical clauses, along with a common safety baseline and typical application scenarios, serving as a reference for product design and performance benchmarking [3]. Group 3: Capability Descriptions - Perception and Cognition: Humanoid robots should acquire, process, and understand environmental and self-state information, enabling reasoning and knowledge application [3][4]. - Decision Learning: Robots must utilize methods like large models and reinforcement learning to achieve precise perception, logical reasoning, and dynamic decision-making [4]. - Execution Performance: Robots should support precise control of joint movements and complex task execution, including mobility and dynamic balance [4]. - Collaboration Interaction: Robots must communicate and collaborate safely and efficiently with humans and other intelligent agents [5]. Group 4: Future Implications - The implementation of the grading standard is expected to facilitate the transition from "demonstrative intelligence" to "general intelligence" in humanoid robots, enabling large-scale applications across various sectors such as special operations, logistics, industrial manufacturing, education, and healthcare [6][8]. - The standard aims to provide a unified technical language and evaluation tools for the industry, promoting market order and guiding technological development [6][8].