具身智能正从技术验证迈向小批量商用，数据卡点成为业界公司着力突破的方向。记者从天奇股份获悉，其运营的江苏省具身智能机器人工业数据采集与实训中心，已累计沉淀近2万小时、PB级具身智能多模态数据，搭建36个一级应用场景、163个二级细分采集场景。这家汽车装备企业，正尝试将长期积累的产线经验，转化为支撑工业具身智能落地的数据与技术能力。
Embodied intelligence is moving from technical validation toward small-batch commercialization, with data bottlenecks becoming a focus for industry players. Reporters learned from Tianqi Shares that the Jiangsu Embodied Intelligence Robot Industrial Data Collection and Training Center it operates has accumulated nearly 20,000 hours and PB-level embodied intelligence multimodal data, building 36 first-level application scenarios and 163 second-level细分 collection scenarios. This automotive equipment enterprise is attempting to convert long-accumulated production line experience into data and technical capability supporting industrial embodied intelligence deployment.
政策正在为这一转变提供牵引。今年6月，工信部、国务院国资委联合印发通知，明确提出到2026年底，围绕工业、服务、特种等真实场景，形成百个以上高价值应用场景，带动形成万台级规模落地能力。工业制造是这条路径的重要入口。料箱搬运、零部件分拣等任务边界相对明确、作业频次高，便于核算收益，也更适合通过部署、作业、数据回流、模型迭代的闭环持续提升能力。天奇股份相关人士今年5月曾表示，哪怕一个订单十几二十台，在工业场景已经是很大的进步。从前期客户反馈看，约三年回本，是工业客户普遍能够接受的投资预期。
Policy is providing traction for this shift. In June this year, MIIT and SASAC jointly issued a notice explicitly proposing that by end-2026, around real scenarios including industrial, service, and special, over 100 high-value application scenarios be formed, driving ten-thousand-unit-scale deployment capability. Industrial manufacturing is an important entry point for this path. Tasks including container handling and parts sorting have relatively clear boundaries and high operation frequency, making revenue calculation easier and better suited for continuous capability improvement through the closed loop of deployment, operation, data return, and model iteration. A Tianqi Shares executive said in May this year that even an order of a dozen or twenty units is already significant progress in industrial scenarios. Based on early customer feedback, about three-year payback is an investment expectation industrial customers generally accept.
天奇股份的差异化资源，首先来自汽车装备业务积累。不少工业场景相对封闭，外部企业难以获取数据，天奇则依托对产线的理解，在采集与实训中心复刻工业环境，并根据客户需求搭建特定场景。熟悉生产流程，才能确定需要训练哪些任务、记录哪些动作，以及如何检验机器人是否满足工艺要求。目前，具身中心构建了真机、UMI无本体采集、EGO第一人称视角采集三大融合数采路径，采集内容覆盖深度图像、点云、关节运动、力觉反馈、控制指令及任务语义等标注信息。针对工业数据碎片化、复用难的问题，天奇股份自研DataMatrix数据采集管理系统，打通采集、清洗、质检和资产交付流程。
Tianqi Shares' differentiated resources first come from automotive equipment business accumulation. Many industrial scenarios are relatively closed, making data difficult for outside enterprises to obtain; Tianqi, relying on production line understanding, replicates industrial environments at the collection and training center and builds specific scenarios based on customer needs. Familiarity with production processes is required to determine which tasks to train, which actions to record, and how to verify whether robots meet process requirements. Currently, the embodied center has built three integrated data collection paths—real machine, UMI bodyless collection, and EGO first-person view collection—with content covering depth images, point clouds, joint motion, force feedback, control commands, and task semantics annotation information. Targeting industrial data fragmentation and reuse difficulty, Tianqi Shares self-developed the DataMatrix data collection management system, connecting collection, cleaning, quality inspection, and asset delivery processes.
根据天奇股份中报，上半年长城汽车人形机器人物流应用项目已完成验收，丰田通商、东风猛士、武汉富士康等具身智能应用项目已完成产品交付或技术验证，应用覆盖整车、汽车零部件、轮胎、锂电池和3C制造等领域。9月，天奇股份定向增发获批，预计融资总额约9.77亿元，其中约2.61亿元投向面向汽车行业应用的机器人具身智能系统研发中心建设项目。天奇股份表示，通过募投项目开展具身智能数据采集管理，定向开发柔性装配、分拣等垂类模型，开发具身智能产线应用解决方案，以具身智能加赋能汽车智能装备，打造一条业务发展的第二增长曲线。
According to Tianqi Shares' interim report, in the first half the Great Wall Motor humanoid robot logistics application project completed acceptance; Toyota Tsusho, Dongfeng Mengshi, Wuhan Foxconn and other embodied intelligence application projects completed product delivery or technical validation, with applications covering complete vehicles, auto parts, tires, lithium batteries, and 3C manufacturing. In September, Tianqi Shares' private placement was approved, with expected total financing of about 977 million yuan, of which about 261 million yuan goes to the robot embodied intelligence system R&D center construction project for automotive industry applications. Tianqi Shares stated it will use the fundraising project to conduct embodied intelligence data collection management, develop vertical models including flexible assembly and sorting, develop embodied intelligence production line application solutions, and empower automotive intelligent equipment with embodied intelligence plus, creating a second growth curve.
这2万小时数据是阶段性积累，但能不能转化成稳定作业能力和交付效率，让客户算清机器人的经济账，还需要订单和财报来验证。前面讨论过，机器人行业当前的核心矛盾不是能不能做出来，而是能不能稳定用、能不能规模化、能不能赚钱。天奇股份这条路能不能走通，取决于两件事，一是这些数据能不能真正训练出在真实产线上稳定作业的模型，二是天奇银河的商业化能力能不能把这些模型变成可复制的解决方案。
These 20,000 hours of data are staged accumulation, but whether they can convert into stable operation capability and delivery efficiency, letting customers calculate the robot's economic账, still needs orders and financial reports to verify. Earlier we discussed the robot industry's current core contradiction is not whether it can be made but whether it can be stably used, scaled, and profitable. Whether Tianqi Shares' path works depends on two things—whether this data can truly train models that operate stably on real production lines, and whether Tianqi Galaxy's commercialization capability can turn these models into replicable solutions.
