据报道，为机器人开发灵巧操作能力的格物通智已完成近亿元天使轮融资，投资方包括鸿富资产、匹克投资、杭州资本、杭州人才集团，以及产业投资机构和上市公司。格物通智于2026年成立于杭州，主要研究如何让机器人理解物体及操作中的物理变化，并在与真实环境的交互中改进动作。公司将世界模型与物理交互学习结合，面向产业场景开发机器人大脑。创始人兼CEO蔡佳曾任职于华为2012实验室及车BU 2030实验室，参与华为ADS辅助驾驶系统研发和量产交付。本轮资金将用于模型预训练、硬件及机器人本体研发、人才引进和真实场景数据采集。
Reportedly, Gewu Tongzhi, which develops dexterous manipulation capabilities for robots, has completed a nearly 100-million-yuan angel round, with investors including Hongfu Assets, Peak Investment, Hangzhou Capital, Hangzhou Talent Group, as well as industrial investment institutions and listed companies. Gewu Tongzhi was founded in Hangzhou in 2026, mainly researching how to let robots understand objects and physical changes during manipulation, and improve actions through interaction with real environments. The company combines world models with physical interaction learning, developing robot brains for industrial scenarios. Founder and CEO Cai Jia previously worked at Huawei's 2012 Lab and Vehicle BU 2030 Lab, participating in Huawei ADS assisted driving system R&D and mass production delivery. Funds will be used for model pretraining, hardware and robot body R&D, talent recruitment, and real-scenario data collection.
格物通智这个定位，和前面讨论的机器人大脑方向是同一个赛道。前面讨论过自变量机器人的WALL-B、星尘智能的Lumo、跨维智能的UNITAS、若愚科技的若愚·九天，这些都是机器人大脑方向的不同技术路线。格物通智的差异点在于，它把世界模型与物理交互学习结合，重点研究机器人理解物体及操作中的物理变化。这个方向和前面讨论的灵巧手、触觉传感器、六维力检测是同一个问题的不同层面。灵巧手负责执行，触觉传感器负责感知，机器人大脑负责理解物理变化并做出决策。格物通智选择的是理解这一层。
Gewu Tongzhi's positioning is the same track as the robot brain directions discussed earlier. Earlier we discussed AgiBot's WALL-B, Astribot's Lumo, DexForce's UNITAS, and Ruoyu Technology's Ruoyu Jiutian—all different technical routes in the robot brain direction. Gewu Tongzhi's differentiation lies in combining world models with physical interaction learning, focusing on robots understanding objects and physical changes during manipulation. This direction operates at different levels of the same problem as the previously discussed dexterous hands, tactile sensors, and six-axis force detection. Dexterous hands handle execution, tactile sensors handle perception, and robot brains handle understanding physical changes and making decisions. Gewu Tongzhi chooses the "understanding" layer.
创始人蔡佳的背景值得单独看。他曾任职于华为2012实验室及车BU 2030实验室，参与华为ADS辅助驾驶系统研发和量产交付。这个背景和前面讨论的初光跃迁创始人张颖有相似之处，张颖是地平线HSD核心技术负责人，带队攻坚过三代智驾方案。蔡佳和张颖都是从自动驾驶领域转向具身智能。前面讨论过自动驾驶顶级工程人才正在系统性转向Physical AI，机器人研发开始Agent化、基础设施化。蔡佳这个案例是同一个趋势的又一个注脚。自动驾驶验证过的感知、推理、执行方法论，正在被迁移到机器人操作场景。
Founder Cai Jia's background deserves separate attention. He previously worked at Huawei's 2012 Lab and Vehicle BU 2030 Lab, participating in Huawei ADS assisted driving system R&D and mass production delivery. This background resembles the previously discussed Chuguang Yueqian founder Zhang Ying—Zhang was Horizon HSD core technology lead, leading three generations of intelligent driving solutions. Both Cai Jia and Zhang Ying transitioned from autonomous driving to embodied intelligence. Earlier we discussed top autonomous driving engineering talent systematically turning to Physical AI, with robot R&D becoming Agent-ized and infrastructure-ized. Cai Jia's case is another footnote to the same trend. The perception, reasoning, and execution methodologies validated in autonomous driving are being migrated to robot manipulation scenarios.
公司正与纺织服装、光模块及汽车零部件领域客户开展概念验证，预计于2026年底完成首个概念验证项目交付。这三个领域的组合值得注意。纺织服装是柔性物体操作，光模块是精密装配，汽车零部件是多品种小批量。这三个场景对机器人的要求各不相同，但共同点是都需要机器人理解物理变化。前面讨论过FPC软排线高精度装配、显卡电源线柔性抓取、多规格钣金件高精度上料，这些都是物理变化复杂的操作场景。格物通智选择这三个领域做概念验证，说明它想用不同场景来检验自己的物理交互学习能力。
The company is conducting proof-of-concept with customers in textiles and apparel, optical modules, and auto parts, expecting to complete its first proof-of-concept project delivery by end of 2026. The combination of these three fields is noteworthy. Textiles and apparel involve flexible object manipulation; optical modules involve precision assembly; auto parts involve multi-variety small-batch production. These three scenarios have different robot requirements, but their common point is that all require robots to understand physical changes. Earlier we discussed FPC flexible cable high-precision assembly, graphics card power cable flexible grasping, and multi-specification sheet metal parts high-precision loading—all manipulation scenarios with complex physical changes. Gewu Tongzhi choosing these three fields for proof-of-concept indicates it wants to use different scenarios to test its physical interaction learning capability.
