Core Viewpoint - Yunji, founded in 2016, is a leading company in the home robotics sector, recognized as one of China's top 50 technology robotics companies for three consecutive years and listed in the Hurun Global Unicorns for five years [2] Company Overview - Yunji has established a presence in over 20 provinces and 70 cities in China, with more than 500 authorized retail outlets [2] - The company is accelerating its global brand expansion, with products available in over 50 countries and regions, including North America, Europe, Japan, South Korea, Australia, and Southeast Asia [2] - In Europe, Yunji has partnered with over 5,000 offline stores, and in North America, its products are available in nearly 300 major retail locations such as Best Buy and Costco [2] - The global user base has surpassed 4 million, and by Q2 2025, Yunji's global shipment of vacuum robots reached 524,200 units, capturing an 8.5% market share, ranking fifth globally [2] Recruitment and Talent Acquisition - Yunji is actively recruiting experts in embodied VLA algorithms, reinforcement learning, imitation learning, and robotic arm planning algorithms, offering open salary discussions for qualified candidates [2]

Core Viewpoint - The article emphasizes the advantages of the Imeta-Y1 robotic arm, highlighting its user-friendly features and cost-effectiveness for beginners and researchers in the field of embodied intelligence [5][6][9]. Group 1: Workflow Improvement - Before using Imeta-Y1, users spend 70% of their time on hardware communication and sensor calibration, facing challenges in code adaptation between simulation and real machines [1]. - After adopting Imeta-Y1, users can quickly simulate and validate algorithms in Gazebo, deploy verified programs to real machines with one click, and efficiently iterate from inspiration to physical action [2][20]. Group 2: Product Features - Imeta-Y1 is designed as a lightweight, high-cost-performance robotic arm, suitable for students, educators, and novice developers [5][6]. - It offers a complete open-source toolchain and code examples, supporting both Python and C++ interfaces, and is compatible with ROS1 and ROS2 [7][21]. - The arm features high-precision motion control, low power consumption, and an open hardware architecture, facilitating seamless integration from simulation to real machine [9][10]. Group 3: Technical Specifications - The robotic arm has a weight of 4.2 kg, a rated load of 3 kg, and 6 degrees of freedom, with a working radius of 612.5 mm and a repeat positioning accuracy of ±0.1 mm [11][22]. - It operates at a voltage of 24V and communicates via CAN, with a compact design suitable for embedded AI and robotic learning platforms [10][11]. Group 4: Development Support - The product provides a full-process toolchain for data collection, model training, and inference deployment, supporting multi-modal data fusion and compatibility with mainstream frameworks like TensorFlow and PyTorch [20][39]. - Users can validate algorithm logic in simulation and deploy to physical devices, significantly reducing development risks and debugging costs [25][39]. Group 5: Customer Support - The company offers 24-hour rapid response for customer support, ensuring users do not face delays in their learning and development processes [7][22]. - There are bulk purchase discounts available, and the company supports project development and training based on the product [22].

Core Insights - The article discusses the advancements in robotics, particularly focusing on the Lumo-1 model developed by Stardust Intelligence, which aims to enhance robots' reasoning and action capabilities, allowing them to perform complex tasks without explicit programming [7][9][11]. Group 1: Lumo-1 Model Overview - Lumo-1 is an end-to-end VLA model designed to integrate reasoning and action in robotics, enabling robots to understand task intentions and execute them seamlessly [7][9]. - The model demonstrates superior performance in multi-step tasks, fine manipulation, and generalizable actions compared to previous models like π0 and π0.5, especially in out-of-distribution scenarios [9][11]. Group 2: Training Phases - The training of Lumo-1 consists of three phases: 1. Embodied VLM pre-training on selected visual-language data to develop spatial understanding and trajectory inference [15]. 2. Cross-ontology joint training to enhance instruction following and spatial reasoning capabilities [16]. 3. Real-world reasoning-action training using the Astribot S1 robot to learn executable action patterns [16][18]. Group 3: Reasoning and Action Alignment - Lumo-1 incorporates structured reasoning, allowing the robot to break down tasks into sub-tasks and understand the relationships between actions and instructions [22][30]. - The model employs reinforcement learning for reasoning-action alignment, calibrating the discrepancies between high-level reasoning and low-level actions, which significantly improves task success rates and generalization capabilities [27][28]. Group 4: Performance Metrics - Lumo-1 outperforms mainstream models in six out of seven multimodal benchmark tests, demonstrating its robust multimodal perception and reasoning abilities without compromising its core functionalities [29]. - The model's ability to adapt to various environments and tasks, such as adjusting arm positions for different container heights and recognizing handwritten menus, showcases its impressive generalization capabilities [29].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 明明在实验室里表现完美的机器人，为何一到真实场景就掉链子？核心症结在于传统静态数据采集模式的局限：一条训练轨迹只能对应一个固定空间配置，物体位 置、目标落点、相机视角全是 "死的"。要让机器人适应不同场景，就得没完没了采集海量数据，不仅耗时耗力，还会陷入数据稀疏的困境。 作者丨 Huanqian Wang等 编辑丨具身智能之心 如今，北京智源人工智能研究院、清华大学、东南大学等机构的团队，凭借一项名为 MOVE（MOtion-Based Variability Enhancement，基于运动的可变性增强） 的创新技术，精准破解了这一行业痛点，让机器人真正学会 "举一反三"！ >> 点击进入→ 具身智能之心 技术交流群 技术资源已公开，欢迎行业伙伴探索 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 核心思路：让数据"动起来"，一条轨迹顶N条用 家里让机器人拿杯水，换个杯子摆放位置就失灵；工厂里机器人抓取零件，摄像头角度稍调就 "抓空"；仓库中分拣货物，货架高度变了就成了 "断线木偶"…… 在 ...

Core Viewpoint - The article discusses the challenges and advancements in the VLA (Variable Learning Algorithm) models, emphasizing the importance of real machine data and practical experience in achieving effective results in embodied intelligence applications [2][4]. Group 1: Data Collection - Data collection methods primarily include imitation learning and reinforcement learning, with remote operation, VR, and full-body motion capture being key techniques [6][7]. - Ensuring high-quality data and effective data collection is crucial, particularly in the context of sim2real applications [7]. Group 2: VLA Training - Prior to real machine deployment, simulation debugging is essential, especially when real machine data is insufficient, making frameworks like Mujoco and Isaac Gym important [9]. - Training techniques are critical, with challenges in fine-tuning models and achieving good results with small data sets; models like π0 and π0.5 require high attention to detail and experience [9][10]. Group 3: VLA Model Deployment - After training, models need to undergo a "slimming" process due to their typically large parameter sizes, which poses challenges for deployment on edge chips; techniques like quantization and distillation are necessary [11]. Group 4: Educational Initiatives - The article introduces a practical course aimed at helping students effectively learn VLA, covering various aspects such as hardware, data collection, algorithms, evaluation, simulation, and deployment [12][14]. - The course is designed for individuals seeking to enter the embodied intelligence field, including students and professionals transitioning from traditional CV, robotics, or autonomous driving sectors [24].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 编辑丨具身智能之心 出发点与背景介绍 与聊天机器人或搜索引擎不同，具身智能体必须具备实时运行能力。智能体动作与外部环境间的反馈闭环决定了其必须拥有快速响应性——就像人类运动员一 样，智能体无法在外界环境不断变化的过程中" 停下来思考 "。但随着前沿模型的规模持续扩大，这一要求的实现难度也越来越高。这一点在机器人学习领域表 现得尤为明显：如今，参数规模达数十亿的视觉-语言-动作模型（VLAs）正被越来越多地用于高频率控制机器人，以完成各类灵巧操作任务。而当模型推理延 迟达到数十至数百毫秒时，如何生成既平滑又具备响应性的动作轨迹，就成了一项不小的挑战。 实时动作分块技术整合了动作分块、流匹配和推理时补绘等技术思路，为解决上述问题提供了一套方案。在该技术框架下，动作块的预测过程是异步进行的 ——即在当前动作块执行的同时，就开始生成下一个动作块。为保证动作块之间的连续性，每次生成新动作块时，都会基于此前已预测的动作构成的固定前 缀，并对剩余部分进行补绘。但遗憾的是，实时动作分块所采用的推理时补绘方法会引入额外计算开销，进而产生延迟，在一定程度上削弱了实时执行框架的 ...

Core Viewpoint - Liang Wenfeng has been recognized as one of the top ten scientists of 2025 by the journal Nature for his significant contributions to the AI field through the DeepSeek model, which has disrupted traditional AI paradigms [3][4][8]. Group 1: DeepSeek Model and Its Impact - The DeepSeek model has dramatically reduced costs in the AI industry while enhancing the global visibility of domestic large models [9]. - DeepSeek demonstrates that high-performance models do not necessarily require vast amounts of data, parameters, or servers to achieve top-tier capabilities [10]. - The recent release of the DeepSeek V3.2 series model has achieved the highest evaluation level among current open-source models in the Agent domain [11][12]. Group 2: Liang Wenfeng's Background - Liang Wenfeng, born in 1985, excelled academically, becoming a top student in his high school and later studying electronic information engineering at Zhejiang University [14][15]. - He transitioned into quantitative investment in 2008, capitalizing on the emerging trend of quantitative trading in China, and his team quickly grew their proprietary funds to over 500 million yuan [17]. - In 2021, his firm, Huanfang Quantitative, became one of the largest quantitative private equity firms in China, managing over 100 billion yuan [19]. Group 3: Recognition of Other Researchers - Mengran Du, another researcher recognized by Nature, discovered the deepest known animal ecosystem on Earth, contributing to the understanding of extreme life and carbon cycling in deep-sea environments [25][28]. - Du's research has been published in prestigious journals and she has participated in numerous deep-sea explorations, enhancing the scientific community's knowledge of deep-sea ecosystems [33].

Core Insights - The article discusses the current state of mass production of embodied robots, highlighting the commitments and developments from various companies in the industry [1][5]. Group 1: Company Developments - Hyundai Motor has committed to deploying tens of thousands of Atlas robots in its manufacturing and logistics operations, addressing production capacity challenges by integrating automotive manufacturing expertise to support Boston Dynamics in scaling robot production [2][4]. - UTree Technology has not disclosed specific order volumes for the year but anticipates annual revenue exceeding 1.2 billion [7]. - ZhiYuan Robotics announced the cumulative production of 5,000 robots, with applications across entertainment, manufacturing, logistics, and scientific research [8]. - UBTECH Robotics secured a significant order worth 264 million yuan for its Walker S2 robot, which can autonomously change batteries, and has established contracts for various industrial applications [10][11]. - Tesla's Optimus robot is positioned as a core future asset, with a target of producing 5,000 units by the end of December 2025 and scaling up to 100,000 units by the end of 2026 [14]. Group 2: Order and Production Capacity - UBTECH's Walker series has achieved a cumulative order volume of 1.3 billion yuan, with a production capacity of 300 units per month, expecting to exceed 500 units in deliveries by 2025 [12]. - The partnership between Shenzhen Huizhi and ZhiPing aims to deploy over 1,000 embodied intelligent robots in logistics and manufacturing processes over three years [15]. - Star Dust Intelligent announced a strategic cooperation for a thousand-unit order of humanoid robots, focusing on industrial applications and leveraging AI technology for enhanced operational capabilities [17][19]. - Songyan Power expects to surpass 2,500 units in orders for bio-inspired and educational robots, with total order value exceeding 100 million yuan [20]. - Original Force Unlimited signed a strategic cooperation agreement worth 260 million yuan with a cultural tourism group [22]. Group 3: Market Trends and Future Outlook - The article indicates a growing trend in the deployment of humanoid robots across various sectors, including industrial, manufacturing, and logistics, with expectations for expansion into more niche markets such as 3C and automotive [19]. - The capital market performance of companies like Zhongqing Robotics shows significant investment interest, with plans to deliver 2,000 units over three years and collaborations with major firms like JD and NVIDIA [24]. - Leju Robotics has ramped up its delivery pace from hundreds to nearly a thousand units, with a target of 2,000 units for the year [25].

点击下方 卡片 ，关注" 具身智能 之心 "公众号 作者丨 Ruizhe Liu等 编辑丨具身智能之心 本文只做学术分享，如有侵权，联系删文 >> 点击进入→ 具身智能之心 技术交流群 更多干货，欢迎加入国内首个具身智能全栈学习社区 ： 具身智能之心知识星球 (戳我) ， 这里包含所有你想要的。 本文第一作者为香港大学InfoBodied AI实验室博士生刘瑞哲，合作者包括周佩、罗谦（同属忆生科技）和孙力。通讯作者为香港大学数据科学研究院及电机电子工 程系助理教授杨言超，以及阿里巴巴达摩院研究员岑俊和宋奕兵。InfoBodied AI实验室在CVPR、ICML、NeurIPS、ICLR等顶会持续发表代表性成果，与国内外知 名高校，科研机构广泛开展合作。 1 机器人为何需要「概念」？ 机器人操作模型常在训练环境表现优异，却在分布外场景失败。例如，能稳定完成"将杯子放入容器"的策略，仅需改变物体颜色、调整位置或增加隔板，就可能彻 底失效。 港大与阿里达摩院联合提出的HiMaCon指出： 泛化失败的根源不在于策略学习不足，而在于缺乏"操作概念"这一认知层。 人类执行任务时，会自然形成"对齐物体"、"抓取目标"、"规 ...

Core Viewpoint - The ATEC 2025 competition represents a significant step towards achieving general embodied intelligence in robotics, challenging teams to develop robots that can operate autonomously in real-world environments rather than controlled settings [28][30][31]. Group 1: Competition Overview - The ATEC 2025 competition took place in a real-world outdoor setting, featuring diverse terrains such as bridges, hills, and stairs, which posed significant challenges for the participating robots [36][38]. - The event aimed to test the robots' abilities to adapt to unpredictable environments, moving beyond the controlled conditions typically seen in robotic competitions [30][31]. - The competition included four main tasks: garbage sorting, autonomous watering, orienteering, and bridge crossing, each designed to assess the robots' multi-modal perception and decision-making capabilities [47][50]. Group 2: Technical Challenges - The competition highlighted three major technical challenges for robotics: environmental perception and cognition, intelligent decision-making and response, and hardware and computational limitations [52][60][69]. - Environmental perception involves not just recognizing objects but understanding their context and condition, which is complicated by real-world factors like contamination and deformation [59]. - Intelligent decision-making requires robots to adapt to dynamic environments, making real-time decisions based on changing conditions, which is currently a significant limitation in robotic capabilities [62][64]. Group 3: Team Performance and Insights - Out of 396 teams, only 13 advanced to the finals, with the winning team, wongtsai, scoring 434 points through effective preparation and innovative strategies [78][79]. - Teams reported that the greatest challenge in outdoor environments was the unpredictability of conditions, which often led to difficult decisions regarding whether to operate autonomously or remotely [80][81]. - The competition fostered a spirit of collaboration and innovation among participants, with many expressing optimism about the future of robotics in real-world applications [93][96].