Workflow
具身智能
icon
Search documents
质疑VLA模型、AI完全不够用?有从业者隔空回应宇树王兴兴
Di Yi Cai Jing· 2025-08-11 11:33
传统的人形机器人面临感知局限、决策断层、泛化瓶颈三大核心挑战。 在世界机器人大会上,宇树CEO王兴兴一口气提了不少"非共识"。他对 VLA (Vision-Language-Action视 觉-语言-动作)模型持怀疑态度, 认为这属于"相对傻瓜式架构";他还说机器人行业对数据关注度有点太 高了,包括灵巧手在内的硬件虽然不够好但够用,行业最大的问题在于具身智能的AI完全不够用。 王兴兴的观点在业内持续引发讨论。今日世界机器人大会上,记者留意到,国家地方共建人形机器人创 新中心首席科学家江磊近20分钟的演讲中,3次提到了王兴兴。 对于王兴兴关于"硬件足够用、大模型不够用"的观点,江磊分享了与阿里、华为等企业交流的体 会:"我们是选不到一个很好的身体",并坦承今天行业确实还用不上全参数模型,机器人的大脑、小 脑、肢体需要深度协同;王兴兴质疑VLA并尝试用视频生成驱动机器人任务,江磊承认"感知-认知-决 策-执行的闭环尚未闭合",呼吁重构VLA模型,寻求新的解决范式;王兴兴还提到,机器人在RL(强 化学习)的Scaling law(尺度定律)是非常值得做的方向,江磊认同表示,强化学习跟模仿学习都需要 进入Scalin ...
杭州,大动作!热门赛道迎利好
证券时报· 2025-08-11 11:28
面对当前研发训练需求和未来具身智能机器人的广泛应用需求,草案要求加强网络和算力基础设施规划布局, 建立多元智算供给服务体系,提升算力资源利用效率,降低算力使用成本。 杭州具身智能机器人产业迎利好。 杭州市司法局8月11日起公开征求《杭州市促进具身智能机器人产业发展条例(草案)》(以下简称"草 案")。该草案提出,根据具身智能机器人产业的技术特性与发展规律,针对不同智能化等级、应用场景分类 的具身智能机器人产品及服务,探索构建分级分类监管体系与沙盒监管机制,建立包容审慎、鼓励创新的监管 生态,激发各类主体创新活力。 当前是具身智能机器人技术密集创新和高速发展的战略机遇期、黄金窗口期,杭州凭借良好的产业基础和创新 生态,在以人形机器人、四足机器人为代表的智能机器人整机研发制造以及生成式人工智能、空间智能等人工 智能技术创新方面已形成先发优势。据统计,杭州已集聚机器人相关企业超200家,在智能机器人运动控制算 法、多模态交互等关键领域处于国内领先水平。 草案要求,将具身智能机器人产业纳入杭州市政策支持框架,通过制度设计增强政策合力,进一步规范产业高 质量发展路径。 超前布局具身智能机器人基础设施 具身智能机器人产业 ...
清华创客,批量南下深圳
21世纪经济报道· 2025-08-11 10:48
Core Insights - The article highlights the significant influence of Tsinghua University alumni in the Chinese internet and AI sectors, with a notable concentration of entrepreneurial activity in Shenzhen [1][12] - As of the end of 2024, there are 97 companies listed on the Science and Technology Innovation Board, Beijing Stock Exchange, and Hong Kong Stock Exchange, with a total market capitalization of 3.8 trillion yuan, showcasing the financial impact of the "Tsinghua system" [12] - Tsinghua University is recognized as the leading institution in cultivating AI entrepreneurs in China, with a focus on various fields including large models, autonomous driving, AI chips, and embodied intelligent robots [1][12] Tsinghua Alumni and AI Entrepreneurship - Zhizhu AI, founded in 2019, is the first among the "Big Six" AI companies to initiate an IPO process, with strong ties to Tsinghua University through its leadership [2] - The "Big Six" AI companies are predominantly founded by Tsinghua alumni, indicating the university's substantial contribution to China's AI landscape [2] - The shift from "All in AI" to "AI for All" and "AI in All" reflects the evolving nature of AI integration into daily life, with Tsinghua entrepreneurs actively participating in this transition [2] Shenzhen as a Hub for Tsinghua Alumni - Shenzhen is emerging as a central point for Tsinghua alumni, driven by the concentration of alumni and potential collaboration opportunities with local enterprises and professors [3] - The synergy between AI software and hardware is particularly strong in Shenzhen, where many AI companies are looking to leverage the city's hardware advantages [3][4] - The robotics sector in Shenzhen has seen rapid growth, with over 74,032 companies as of last year, marking a 39.64% increase from the previous year [4] Robotics Industry Growth - The robotics industry in Shenzhen is characterized by a significant presence of Tsinghua alumni, with many companies founded by graduates from the university [4][6] - Notable companies include Dahuang Robotics and Xingji Guangnian, both founded by Tsinghua alumni, showcasing the university's influence in the robotics sector [4][6] - The emergence of new startups, including Zero Point Robotics, indicates a trend of younger Tsinghua graduates entering the robotics field [5] Technology Transfer and Innovation - Tsinghua University leads in technology transfer among Chinese universities, with a reported contract amount of 3.45 billion yuan from 120 approved projects in 2023 [8][9] - The university's alumni network plays a crucial role in facilitating technology transfer and collaboration, enhancing the innovation ecosystem [9] - The establishment of a supportive alumni network fosters trust and collaboration among entrepreneurs, which is vital for startup success [9] Investment and Funding Landscape - Zero Point Robotics recently completed significant funding rounds, highlighting the active investment environment surrounding Tsinghua alumni ventures [10][11] - Investment firms associated with Tsinghua alumni, such as Lihua Science and Technology, are instrumental in supporting startups through various funding stages [10][11] - The interconnectedness of technology, industry resources, and capital within the "Tsinghua system" creates a multiplier effect for innovation and entrepreneurship [10]
4D空间智能:AI如何一步步「看懂」时空结构?一篇综述解析通往四维世界的五大层次
机器之心· 2025-08-11 10:40
4D 空间智能重建 是计算机视觉领域的核心挑战,其目标在于从视觉数据中还原三维空间的动态演化过程。这一技术通过整合静态场景结构与时空动态变化,构建 出具有时间维度的空间表征系统,在虚拟现实、数字孪生和智能交互等领域展现出关键价值。 当前研究主要围绕两大技术维度展开: 基础重建层面聚焦深度估计、相机定位、动态点云等底层视觉要素的精准提取;高阶理解层面则致力于解析场景组件的时 空关联与物理约束。 这种多维度的空间建模能力正成为新一代人工智能发展的基础设施——无论是构建具身智能的环境认知体系,还是训练具备物理常识的世界模型,高保真的 4D 空 间表征都发挥着基石作用。 值得注意的是,前沿研究正从单纯的几何重建转向对 场景物理属性 和 交互逻辑 的建模,这种转变使得空间智能不仅能呈现视觉真实的动态场景,更能支撑智能 体与虚拟环境的拟真交互。 为了填补关于 4D 空间智能重建分析的空白,南洋理工大学 S-Lab、香港科技大学以及德州农工大学的研究者们全面调研了该领域的发展和最前沿的研究方法,撰 写了综述论文,对 400 余篇代表性论文进行了系统归纳和分析。 Paper:Reconstructing 4D Spatial ...
25自由度+20N指尖力,WRC最强仿人灵巧手横空出世
机器人大讲堂· 2025-08-11 10:36
在人形机器人赛道加速狂奔的当下,核心零部件的性能与量产能力始终是制约行业落地的关键瓶颈。 其中, 机器人灵巧手作为人形机器人的核心功能部件 ,在其中占据了重要地位,目前也是人形机器人板块技术迭代 最快的部件。 中商产业研究院数据显示, 2024 年灵巧手市场容量为 76.01 万只,随着灵巧手功能持续完善,叠加下游人 形机器人的放量,预计 2025 年灵巧手市场容量将达 86.18 万只,对应市场规模为 19.21 亿美元 。而 根 据 QYResearch 预测, 2030 年全球机器人多指灵巧手市场规模将超过 50 亿美元, 2025-2031年CAGR 为68.7% 。 2025 年世界机器人大会( WRC )上, 机器人大讲堂注意到, 中国初创企业曦诺未来 就 带来 了 三 款 突破性 产品 ,包括 全球首款全自研量产的 25 自由度腱绳传动灵巧手 , 扭矩密度最 高 可 达 322Nm/kg 的关节模组 和 推力密度最 高 可达 5.2N/g 的 微型 电缸 , 硬核参数领先国际顶尖水平,为行业提供了一 套从技术原理到工程落地的完整解决方案。 ▍ 8 个月崛起:技术积累的 " 厚积薄发 " 全球首款 ...
AH股市场周度观察(8月第2周)-20250811
ZHONGTAI SECURITIES· 2025-08-11 10:35
A-Share Market - The overall market showed an increase last week, with small-cap stocks performing significantly better than mid and large-cap stocks. The CSI 2000 index rose by 3.54%, and the CSI 1000 index increased by 2.51%, while the ChiNext index saw a modest rise of less than 0.5% [5] - The average daily trading volume reached 1.70 trillion, reflecting a week-on-week decline of 6.26% [5] - The recent market recovery has favored small-cap stocks, with sectors such as non-ferrous metals, machinery, and national defense performing well, while the computer and pharmaceutical sectors lagged [5] - The Shanghai government released the "Implementation Plan for the Development of Embodied Intelligence Industry," aiming for a core industry scale to exceed 50 billion by 2027, which, along with the World Robot Conference, has attracted investment in the machinery sector [5] - Following the recent Politburo meeting, which aligned with expectations, the overall monetary policy remains stable, and the "anti-involution" narrative is moderate. The A-share market is expected to continue a structural upward trend driven by valuation recovery under a dual easing fiscal and monetary environment [5] Hong Kong Market - The Hong Kong market experienced a rebound last week, with the Hang Seng Index rising by 1.43% and the Hang Seng Tech Index increasing by 1.17%. The materials sector, particularly metals, saw significant gains, with weekly increases exceeding 10%, while healthcare and consumer discretionary sectors faced pressure [6] - The rebound in the Hong Kong market was primarily driven by the materials sector, especially metals. The fifth round of price increases for coking coal was fully implemented, and the ongoing "anti-involution" trend boosted the coal sector. Additionally, rising gold prices contributed to substantial gains in gold stocks [6] - Companies in the metals sector, such as China Hongqiao, reported better-than-expected interim results, further supporting the rise in the metals sector [6] - Market sentiment has improved, with the Hang Seng Index showing relative strength, mainly driven by cyclical sectors. The ongoing "anti-involution" trend is expected to sustain the recovery of cyclical sectors. Furthermore, the current low valuation of the internet social services sector in Hong Kong presents potential upside, and with rising AI capital expenditures and increased support for technological innovation policies, leading companies in the Hang Seng Tech sector have medium to long-term growth potential [6]
机器人大会应用及降本双提速,智能化再成能力瓶颈
Investment Rating - The report maintains a positive outlook on the automotive industry, indicating a "Buy" rating for the sector [1]. Core Insights - The report highlights the potential of robotics in reducing costs and expanding applications in various fields, including home companionship, healthcare, logistics, and industrial sectors. It suggests that the robotics market will experience a surge similar to the early days of autonomous driving [4]. - The report emphasizes the importance of companies with strong scene implementation capabilities and suggests focusing on high-end markets and strong alpha enterprises such as NIO, JAC, Li Auto, and Xiaomi [4]. - The automotive market is expected to see numerous new model launches and deliveries in August, driven by supply and strong alpha performance [4]. Industry Updates - In the week of July 28 to August 3, 2025, retail sales of passenger cars reached 462,000 units, a month-on-month increase of 5.24% but a year-on-year decrease of 1.28%. Traditional energy vehicles sold approximately 217,000 units, while new energy vehicles sold 245,000 units, with a penetration rate of 53.03% [4]. - The automotive industry recorded a total transaction value of 523.07 billion yuan, with the automotive index rising by 2.68% during the week [4][22]. - The report notes that traditional and new energy raw material price indices have both increased recently, indicating rising costs in the industry [4]. Market Situation - The automotive industry index outperformed the Shanghai Composite Index, ranking 7th among all sectors in terms of weekly growth [22]. - A total of 259 automotive stocks rose, while 31 fell, with the largest gainers being Zhongma Transmission, Huami New Materials, and Aerospace Science and Technology [28]. Key Events - The 2025 World Robot Conference showcased advancements in embodied intelligence and robotics, highlighting the potential for industrial applications [5]. - BYD announced an OTA upgrade for its Song L EV model, enhancing its intelligent driving capabilities [10]. - The launch of the Geely Galaxy A7 fills a gap in the mid-size car market, with a price range of 81,800 to 117,800 yuan [17][18]. Investment Recommendations - The report recommends focusing on domestic strong alpha manufacturers such as Li Auto, NIO, Xiaomi, and Xpeng, as well as companies involved in intelligent trends and state-owned enterprise reforms [4]. - It suggests monitoring component manufacturers with strong performance growth and robotics capabilities, recommending companies like Fuyao Glass and New Spring [4].
重磅发布!刚刚,利好传来!
券商中国· 2025-08-11 10:31
机器人赛道迎来利好政策。 8月11日,杭州市司法局发布公开征求《杭州市促进具身智能机器人产业发展条例(草案)》(以下简称《条 例》)意见的公告,通过制度设计规范发展秩序、优化产业生态、强化要素保障,全面推动具身智能机器人产 业实现高质量发展。 值得一提的是,《条例》提到,促进具身智能机器人在工业制造、农业生产、医疗健康、教育培训、特种作 业、公共安全等重点领域场景的应用推广。 重磅发布 8月11日消息,杭州市司法局发布公开征求《杭州市促进具身智能机器人产业发展条例(草案)》)意见的公 告,将具身智能机器人产业纳入全市政策支持框架,通过制度设计增强政策合力,进一步规范产业高质量发展 路径。 《条例》指出,面对当前研发训练需求和未来具身智能机器人的广泛应用需求,要求加强网络和算力基础设施 规划布局,建立多元智算供给服务体系,提升算力资源利用效率,降低算力使用成本。 《条例》明确核心技术主攻方向,聚焦具身智能"大脑"(具身智能大模型)、"小脑"(运动控制系统)及"本 体"(机器人核心部件和整机)三大核心以及专用芯片等"卡脖子"环节,具体来看: 在"大脑"研发方面,支持开展具身智能模型的研发和应用,提升多模态感知融合 ...
对话越疆科技首席科学家:未来10年,每个家庭都会有台机器人
证券时报· 2025-08-11 10:13
在2025世界机器人大会(下称"WRC2025")上,越疆科技发布具身智能人形机器人DOBOTAtom第二代产品。越疆早期凭借创新的四轴机械臂迅速占领市场, 此后随着市场需求的变化,越疆将战略重心转向更高附加值的六轴协作机器人领域,2023年起正式启动人形机器人研发。 越疆做人形机器人是基于哪些考量?人形机器人的生态未来会如何发展?带着这些问题,证券时报记者在WRC2025期间专访了越疆首席科学家郎需林。 从工业机器人到人形机器人 证券时报:越疆大概什么时候开始研发人形机器人?为什么考虑做人形机器人?越疆人形机器人的特色和亮点在哪儿? 郎需林: 我们开始的不算早,但我们选对了路径。越疆从2023年起正式启动人形机器人研发,但我们更早就在做底层准备——从控制器、驱动器到关节模组的自 研能力,为人形机器人铺设了底层"脊椎"。 之所以进场,是因为我们看到AI模型涌现带来的控制、认知、交互能力突破,让"人形"不再是外形模仿,而是"人类行为的具身再现"。 我们的Atmo系列强调"可量产、可部署",具备闭环控制系统、轻量化关节模组,以及适配VLA等大模型的多模态感知与任务泛化能力,这让它不仅能走路,还能工 作。 证券时报 ...
WRC整理床铺机器人背后模型曝光!端到端双系统全身智能VLA,仅凭少量微调就能get任务
量子位· 2025-08-11 10:12
Core Insights - The article discusses the launch of the G0 model by Xinghai Map, which enables robots to autonomously and continuously perform tasks like bed-making with minimal fine-tuning [1][3] - The G0 model integrates a large-scale open-world dataset with a dual-system architecture to enhance the robot's adaptability in unstructured environments [3][4] Dataset Overview - The Galaxea Open-World Dataset is the first high-quality dataset collected in real human environments, covering 50 types of settings and totaling 500 hours of high-quality mobile operation data [7][9] - The dataset includes over 150 tasks, 1600+ objects, and 58 skills, emphasizing diversity and complexity in task distribution [10][12] - Data collection utilized a unified robotic platform, ensuring consistent action space and perception inputs across various tasks and scenarios [8][9] Model Architecture - The G0 model employs a dual-system structure, separating high-level reasoning and low-level action execution into G0-VLM and G0-VLA modules [18][19] - G0-VLM processes natural language task instructions and breaks them down into executable subtasks, while G0-VLA executes actions at a high frequency [19] Training Strategy - The training of the G0 model is divided into three phases: cross-platform pre-training, single-platform pre-training, and task-specific post-training [21][23] - This phased approach enhances the model's performance and adaptability to complex skills, particularly in long-sequence tasks [21][23] Performance Evaluation - The G0 model outperformed the π0 benchmark model in various tasks, demonstrating significant advantages in object manipulation and language instruction following [22][30] - The model achieved over 50% improvement in accuracy compared to baseline models, particularly after specialized training [32] Conclusion - The combination of the Galaxea Open-World Dataset and the G0 dual-system VLA model provides a scalable and high-fidelity approach for training and deploying embodied intelligence [33] - The open-sourcing of data and models aims to bridge technological gaps and accelerate the transition of embodied intelligence from laboratory innovations to societal applications [33]