就像石油是工业时代的命脉，数据正在成为 AI 时代具身智能爆发的核心燃料。借鉴自动驾驶从数据匮乏到规 模化落地的演进路径， 2026 年，随着 算法突破、算力突围、硬件成熟之后，数据，这个曾被低估的世界级 难题 ，如今已成为决定机器人企业竞争力的关键变量 。 此外，当前具身智能行业已形成明确共识：人形机器人有望成为未来主流落地形态。基于人形机器人实现的数 据集，无疑更容易 契合行业标准与发展趋势，数据价值长期保值 。 从长期价值 来看，基于人形机器人采集 的数据， 随着人形机器人产业化落地，其采集的数据可直接用于量产机型的模型迭代，形成数据采集 - 模型 优化 - 产品落地 - 更多数据采集的正向循环 。 ▍ 具身智能数据采集企业有哪些？ （ 1 ）星海图数据采集系统 星海图构建了一套从物理硬件到数据采集，再到数据管理、标注与处理的完整闭环平台（ EDP ）。其核心逻 辑是通过标准化的机器人本体在真实世界中执行任务，系统性地采集高质量、多模态数据，并利用完备的工具 链进行管理、标注，最终用于模型训练与迭代。 数据采集系统以 R1 Lite 通用双臂移动操作平台与 R1 Pro 高性能仿人机器人作为标准化的数据 ...

Core Viewpoint - The article discusses the challenges and complexities faced by beginners in implementing VLA (Vision-Language Alignment) models, emphasizing the need for practical experience and effective training methods to achieve successful deployment in real-world applications [2][4]. Group 1: Challenges in VLA Implementation - Many students report difficulties in achieving effective results with open-source models like GR00T and PI0, despite low training loss in simulations [2][4]. - The transition from simulation to real-world application (sim2real) poses significant challenges, particularly in data collection and model training [6][7]. - Beginners often struggle with the intricacies of data collection, model training, and deployment, leading to frustration and lack of progress [4][10]. Group 2: VLA Model Components - Data collection methods for VLA primarily include imitation learning and reinforcement learning, with a focus on high-quality data acquisition [6]. - Training VLA models typically requires simulation debugging and fine-tuning, especially when real-world data is limited [7]. - Deployment of VLA models necessitates optimization techniques such as model compression to ensure efficient performance on edge devices [9]. Group 3: Educational Initiatives - The article introduces a practical course aimed at helping students effectively learn VLA, covering various aspects such as hardware, data collection, algorithms, and real-world experiments [10][12]. - The course is designed for individuals seeking to enter the field of embodied intelligence, providing hands-on experience and project support [22][25]. - The course will commence on December 30, 2025, and includes a comprehensive curriculum to enhance participants' skills in VLA [23][26].

Group 1: Market Performance - The latest price of the Robot ETF (562500) is 0.972 yuan, showing a slight increase of 0.31% [1] - Among the 73 constituent stocks, 43 stocks rose while 30 stocks fell, with some strong stocks increasing by over 3% [1] - Trading activity remains active with stable transaction volume, and the overall operation is stable [1] Group 2: Technological Developments - Yushu Technology launched a humanoid robot data collection training solution based on the G1-D wheeled robot, featuring high-performance components and comprehensive training tools [2] - The G1-D robot has a height range of approximately 1260-1680mm and is equipped with high-definition cameras [2] - The G1-D comes in two versions, with the flagship version offering additional features such as a mobile chassis and a maximum load capacity of 3kg per arm [2] Group 3: Industry Trends - CITIC Securities reports that by 2025, various regions will invest in data collection factories, with companies like Zhiyuan taking a leading role [3] - Local governments are increasingly partnering with manufacturers to establish data collection factories, with a focus on long-term technical support [3] - The domestic labor cost for data collection is significantly lower than in North America, providing a competitive advantage for scaling data collection operations [3] - It is projected that by 2025, over 700,000 hours of real data will be produced domestically, with nearly 8,000 data collection units expected by 2028 [3]

Core Viewpoint - Tesla is leveraging a data collection team to train its Optimus robot, focusing on human-like actions through extensive video data collection, which has implications for the future of robotics and AI integration [2][3][4]. Group 1: Data Collection Methodology - Tesla's data collection involves employees performing repetitive tasks for up to 8 hours, collecting at least 4 hours of usable video footage per shift [2]. - The company has shifted from using motion capture suits to camera-based data collection, which allows for larger scale data gathering [2][3]. - The physical demands on data collectors are significant, with reports of injuries due to the weight of equipment and prolonged use of headsets [3]. Group 2: Workforce and Production Goals - At its peak, Tesla had over 100 employees dedicated to data collection for the Optimus project [3]. - Elon Musk has set an ambitious target of producing 1 million units of Optimus annually, with the robot business projected to account for 80% of Tesla's value in the future [3]. Group 3: Data Types and Industry Trends - The industry recognizes the importance of diverse training data, with real data considered "golden data" for training effectiveness, despite its higher costs [4]. - A hybrid approach combining real and simulated data is becoming the standard in the robotics sector, aiming to enhance robots' environmental perception and multitasking capabilities [4]. - The data collection systems market is projected to exceed $2.4 billion by 2025, with a compound annual growth rate of approximately 5.2% from 2026 to 2035 [4]. Group 4: Future of Robotics Training - There are indications that future robot training may become "AI-driven," with Tesla recently announcing the use of self-developed world models for training Optimus [5]. - Current methods in the industry, such as world models and simulation training, have limitations in achieving generalization capabilities, indicating a need for further exploration in embodied intelligence learning methods [5].

Core Insights - The competition for high-quality data has intensified as AI companies mature, leading to the emergence of firms like Mercor, Surge, and notably, Scale AI founded by Alexandr Wang [1] - Investors are increasingly interested in companies with innovative data collection strategies, as evidenced by the recent $15 million Series A funding for Datacurve, led by Mark Goldberg's Chemistry fund [2][3] Funding and Investment - Datacurve previously secured $2.7 million in seed funding, with participation from former Coinbase CTO Balaji Srinivasan [3] - The recent funding round attracted investments from employees of DeepMind, Vercel, Anthropic, and OpenAI, indicating strong interest from key players in the AI sector [2] Business Model and Strategy - Datacurve employs a "bounty hunter" mechanism to attract skilled software engineers to gather difficult datasets, having paid out over $1 million in rewards to date [4] - The company emphasizes user experience over monetary compensation, aiming to create a consumer-grade product rather than a traditional data annotation pipeline [5] Market Trends - The demand for data is growing exponentially in both quantity and quality due to the increasing complexity of AI models, which require targeted and strategic data collection [6] - Datacurve's model is adaptable and can be applied across various sectors, including finance, marketing, and healthcare, as it builds infrastructure for post-training data collection [7]

Group 1: Cross-Border Office Solutions - Core demand for stable access and data security in remote collaboration, file transfer, and video conferencing [1] - Node selection for Asia includes Hong Kong (CN2 line optimization, <50ms latency) and Singapore (30ms latency), suitable for cross-border e-commerce and gaming acceleration [2] - Node selection for Europe and America involves multiple US nodes (e.g., New York, Los Angeles) with SD-WAN technology to reduce network costs by 30% [2] - Bandwidth configuration for lightweight business requires 3-10Mbps dedicated bandwidth, while medium to large businesses need 50-200Mbps to support high concurrency [3] Group 2: Data Collection Solutions - Core demand for high anonymity and stability to avoid anti-scraping mechanisms and obtain target data [7] - Preference for residential IPs due to high anonymity and lower risk of being flagged as "associated IPs" compared to data center IPs [8] - Dynamic IP rotation supports time-based or task-based switching to prevent triggering anti-scraping measures [9] - Multi-node redundancy with 2-3 nodes in target markets to lower latency through BGP Anycast [10] - Dedicated bandwidth ensures high-speed data transfer, especially for large volume data collection [10] Group 3: Overseas Testing Solutions - Core demand for simulating target market environments and verifying business compliance [14] - Use of target market IPs to simulate real user environments, enhancing pricing and advertising effectiveness [14] - Compliance verification requires selecting service providers certified by GDPR and CCPA to ensure localized data storage [15] - High concurrency support for testing with ≥5000 threads and bandwidth ≥1Gbps to meet large-scale pressure testing needs [15] - Real-time monitoring tools to detect response speed and packet loss, triggering alerts for stability [16]

Core Insights - The investment and financing heat in the embodied intelligence industry continues to rise, with a consensus that data collection is a critical breakthrough for advancing from L1 specific task intelligence to L2 combined task intelligence and beyond [1][4] Data Collection as a Key Variable - The core goal of embodied intelligence is to enable robots to possess common sense understanding, allowing them to deduce operational logic based on past experiences when faced with unfamiliar objects and tasks, which relies on high-quality, multi-modal interaction data [4][6] - Achieving human-eye-level 3D perception requires constructing a dataset of over 1 billion entries, highlighting the industry's urgent need for efficient and high-quality data collection solutions [3][6] Current Development Stage - Leading domestic companies are still in the early L1 development stage, capable of performing single-position tasks in specific environments, while the π0.5 model has achieved over 60% accuracy in long-range tasks in real home environments, nearing L2 levels [6][12] - The pre-training effect is crucial for the advancement of embodied intelligence technology, directly dependent on the "quantity" and "quality" of data [6][12] Four Core Data Collection Solutions - The selection of data collection solutions in embodied intelligence is fundamentally about balancing "cost, precision, and generalization capability" [7][28] - **Remote Operation**: High precision but high cost, with a complete setup exceeding 200,000 yuan, making it a significant financial burden [8][12] - **Simulation**: Low cost but suffers from distribution shift issues, making it less effective in real-world applications [14][16] - **UMI Multi-modal Sensor Fusion**: A cost-effective choice for SMEs, providing a balance between cost and precision, but limited in full-body motion capture capabilities [19][21] - **Video Learning**: Led by Tesla, this low-cost exploration method captures employee task execution videos, significantly reducing costs compared to remote operation [22][24] Industry Trends and Future Directions - The future trend in data collection for embodied intelligence will likely involve the integration of multiple solutions to achieve a balance of cost, precision, and scale [28] - The ultimate goal is to achieve an "autonomous data loop," where robots can independently complete tasks, collect data, and optimize models without human intervention [28]

Core Viewpoint - The founder and CEO of Yushu Technology, Wang Xingxing, emphasized that both data and model architecture are crucial for the development of the robotics industry, countering the notion that the main limitation is insufficient data [1] Data Utilization - The current core issue regarding data is the difficulty in determining the standards for high-quality data, including how to collect it and the necessary scale for collection, which remains ambiguous [1] - There is a call to improve the utilization rate of data as a means to enhance the industry's growth potential [1]

Core Viewpoint - The article discusses the transition from autonomous driving to embodied intelligence, highlighting the similarities and differences in algorithms and tasks between the two fields [1]. Group 1: Algorithm and Task Comparison - Embodied intelligence largely continues the algorithms used in robotics and autonomous driving, such as training and fine-tuning methods, as well as large models [1]. - There are notable differences in specific tasks, including data collection methods and the emphasis on execution hardware and structure [1]. Group 2: Community and Learning Resources - A full-stack learning community named "Embodied Intelligence Heart" has been established to share knowledge related to algorithms, data collection, and hardware solutions in the field of embodied intelligence [1]. - Key areas of focus within the community include VLA, VLN, Diffusion Policy, reinforcement learning, robotic arm grasping, pose estimation, robot simulation, multimodal large models, chip deployment, sim2real, and robot hardware structure [1].

Core Insights - The article emphasizes the importance of joining the "Embodied Intelligence Heart Knowledge Planet," a comprehensive community for learning and sharing knowledge in the field of embodied intelligence, which is rapidly growing in popularity and demand [1][16][85]. Community Features - The community offers a variety of resources including video content, written materials, learning pathways, Q&A sessions, and job exchange opportunities, aiming to create a robust platform for both beginners and advanced learners in embodied intelligence [1][2][17]. - It has established a job referral mechanism with multiple leading companies in the embodied intelligence sector, facilitating direct connections between job seekers and employers [10][17]. Learning Resources - The community has compiled over 30 technical pathways, covering various aspects of embodied intelligence, such as data collection, algorithm deployment, and simulation [2][16]. - It provides access to nearly 40 open-source projects and 60 datasets related to embodied intelligence, significantly reducing the time needed for research and development [16][30][36]. Networking and Collaboration - The community hosts roundtable discussions and live broadcasts to share insights on the latest developments in the embodied intelligence industry, fostering collaboration among members [4][76]. - Members can freely ask questions and receive guidance on career choices and research directions, enhancing the collaborative learning environment [78]. Industry Insights - The community includes members from renowned universities and leading companies in the field, ensuring a diverse range of expertise and perspectives [16][20][21]. - It provides summaries of industry reports and research papers, keeping members informed about the latest trends and applications in embodied intelligence [23][26].