人形机器人进厂正从单点概念验证转向产线常态化运行。近日，财联社记者前往汽车轮毂轴承龙头兆丰股份位于杭州的工厂实探。在该公司与乐聚智能合作部署的车间内，15台夸父人形机器人已投入常态化上料作业，经过一年磨合，累计运行超过约1000小时，完成超6万件零件生产，作业成功率稳定在98%以上。高工机器人产业研究所预测，2026年国内人形机器人出货量有望从2025年的约1.8万台增至6.25万台，工业制造是各家厂商切入的首选领域。

Humanoid robots entering factories are shifting from single-point proof of concept to normalized production line operation. Recently, Caixin reporters visited the Hangzhou factory of automotive wheel bearing leader Zhaofeng Co., Ltd. In the workshop deployed in cooperation with Leju Intelligent, 15 Kuafu humanoid robots have entered normalized loading operations; after a year of running-in, they have cumulatively operated over about 1,000 hours, completed over 60,000 parts production, with operation success rate stable above 98%. GGII predicts 2026 domestic humanoid robot shipments could rise from about 18,000 units in 2025 to 62,500 units, with industrial manufacturing the首选 field for manufacturers.

不过，记者在车间看到，夸父机器人单次上料耗时35至40秒，而产线熟练工人上料仅需约15秒；与此同时，此次上岗的15台机器人全部采用轮式底盘加人形双臂构型。既然当前作业速度尚不及熟练工人与固定机械臂，制造企业为何还要引入人形机器人。在汽车制造与3C电子车间，人形机器人目前承接的岗位集中在上下料、搬运及分拣等工序。GGII所长蔡炳贞分析，汽车线束装配及焊接、喷涂、打磨等强工艺环节，预计还需8年以上才可能由人形机器人承担；现阶段具备落地条件的均为偏弱工艺环节，其中柔性上下料是需求最迫切的切入点之一。

However, reporters saw in the workshop that Kuafu robots take 35 to 40 seconds per loading, while skilled production line workers need only about 15 seconds; meanwhile, all 15 robots on duty use a wheeled chassis plus humanoid dual-arm configuration. Since current operation speed still lags skilled workers and fixed robotic arms, why do manufacturers still introduce humanoid robots. In automotive and 3C electronics workshops, humanoid robots currently take positions concentrated in loading/unloading, handling, and sorting. GGII Director Cai Bingzhen analyzed that strong process links including automotive wiring harness assembly, welding, painting, and grinding are expected to require over 8 more years before humanoid robots can undertake them; currently deployable links are all weaker process links, with flexible loading/unloading one of the most urgent entry points.

为何不用10秒节拍的固定机械臂。蔡炳贞解释，在中小企业及频繁切换工序的产线中，传统机械臂更换工具后无法直接适配，仍依赖大量人工上下料；此外，涉及软包等柔性材质处理时，普通机械臂难以替代人工，人形机器人部署更快、泛化性更好。以兆丰股份为例，该公司总经理孔辰寰表示，固定产线适用于大批量、单一品种生产，而公司部分海外售后订单属小批量、多品种，最小起订量仅250至300套，对高频换产诉求极高，人形机器人能弥补传统刚性连线的不足。一线用工结构的变化，也是制造企业引入机器人的考量。兆丰股份董事长孔爱祥坦言，现在不是工资高低的问题，是年轻人普遍不愿意干车间的体力劳动，工厂办到明年、后年可能就会遇到用工瓶颈，没人干就只能训练机器人来分担。

Why not use fixed robotic arms with 10-second cycle times. Cai Bingzhen explained that in SMEs and production lines with frequent process switching, traditional robotic arms cannot directly adapt after tool changes, still relying on大量 manual loading; additionally, when handling flexible materials like soft bags, ordinary robotic arms struggle to replace humans, and humanoid robots deploy faster with better generalization. Taking Zhaofeng as an example, General Manager Kong Chenhuan said fixed lines suit large-batch, single-variety production, while some of the company's overseas after-sales orders are small-batch, multi-variety with minimum order quantities of only 250 to 300 sets, demanding high-frequency changeovers; humanoid robots can弥补 traditional rigid line shortcomings. Changes in frontline labor structure are also a consideration. Zhaofeng Chairman Kong Aixiang candidly said: "It's not a wage issue now—young people generally don't want to do physical workshop labor. The factory may hit a labor bottleneck next year or the year after; if nobody does it, we can only train robots to share the load."

在兆丰股份车间内，这15台夸父机器人已适配15个工位、覆盖10余种零件，按工人指令完成取料、转运、上料，可稳定抓取内径80至120毫米的零件。孔爱祥告诉记者，机器人进车间经历了漫长磨合。机器人刚来时手拿不住，光手部抓取这个零件就做了约1.5万次训练，反复调整力度和算法模型。乐聚智能常务副总裁柯真东补充道，在实验室里把成功率做到95%以上并不难，但车间地面有油污，伴随人员穿行与光照变化，团队必须不断修正数据集，与自动化产线节拍进行对齐，才能将产线成功率稳定在98%以上。

In Zhaofeng's workshop, these 15 Kuafu robots have adapted to 15 workstations covering over 10 part types, completing material picking, transfer, and loading per worker instructions, stably grasping parts with inner diameters of 80 to 120 mm. Kong Aixiang told reporters robots went through lengthy running-in. "When the robots first arrived, their hands couldn't hold things—just hand grasping this part required about 15,000 training iterations, repeatedly adjusting force and algorithm models." Leju Intelligent Executive Vice President Ke Zhendong added that achieving over 95% success in the lab is not hard, but workshop floors have oil stains, with personnel walking through and lighting changes; the team must continuously correct datasets and align with automated line cycle times to stabilize line success above 98%.

解决了抓得住的问题后，作业节拍成为机器人能否推广的关键。乐聚智能CEO常琳表示，目前机器人的准确度和平均无故障工作时间均已达标，下一步是将单件上料时间压缩至与生产节拍相匹配的15秒。孔爱祥透露，双方将优化算法以减少弯腰、伸腰等冗余动作，推进双臂单次抓取十四五个零件，同时将推进芯轴零件抓取；预计半年到一年内单次上料节拍有望达到15秒左右，投资回报周期预计大约在3年。除提升单工位节拍外，兆丰股份还在探索跨时段的多技能包调度。孔辰寰透露，公司一条产线有26至28个工位，切换型号时需3名工程师在开工前花2至3小时逐台调校设备。如果让机器人在早晨5、6点的非开工时间，通过工人指令提前做完二十几个工位的点位调校，每天就多出2至3小时生产时间，直接转化为新增产能。

After solving grasping, operation cycle time became key to whether robots can scale. Leju Intelligent CEO Chang Lin said current robot accuracy and mean time between failures have met standards; next is compressing single-piece loading time to 15 seconds matching production cycle. Kong Aixiang revealed both parties will optimize algorithms to reduce redundant bending and stretching motions, advancing dual-arm single-grasp of 14 to 15 parts, while advancing core shaft part grasping; within half a year to a year, single loading cycle is expected to reach about 15 seconds, with investment payback estimated at about 3 years. Beyond improving single-workstation cycle time, Zhaofeng is exploring cross-time multi-skill package scheduling. Kong Chenhuan revealed one production line has 26 to 28 workstations; when switching models, 3 engineers need 2 to 3 hours before work to calibrate equipment one by one. If robots at 5 or 6 a.m. non-working hours, through worker instructions, complete point calibration for over twenty workstations in advance, each day gains 2 to 3 hours of production time, directly converting to new capacity.

针对为何此次方案全部采用轮式机器人，柯真东说，我们认为具身智能的核心是数据驱动人形机器人产业落地，无所谓下半身为人形还是轮臂，核心是通过数据训练让机器人具备更多的场景技能泛化。基于行业现阶段的情况，轮臂能更快完成工业批量落地，双足人形能更快完成室外复杂地形巡检批量落地。随着小脑模型性能的不断提升，人形依然是最终方向。蔡炳贞表示，轮式机器人在工厂中的更大价值在于打通跨车间、跨工作站的非结构化流转，将搬运与分拣合二为一；至于双足机器人，中大型工厂地面平整，使用双足必要性不高，只有无法改造厂房的小型企业或非结构化场景才需要双足，预计双足在工厂大规模常态化部署至少还需5年。

On why all solutions used wheeled robots, Ke Zhendong said: "We believe embodied intelligence's core is data-driven humanoid robot industrial deployment—it doesn't matter whether the lower body is humanoid or wheeled-arm; the core is training robots through data to have more scenario skill generalization. Based on the industry's current stage, wheeled-arm can faster complete industrial batch deployment; bipedal humanoid can faster complete outdoor complex terrain inspection batch deployment. As cerebellum model performance continuously improves, humanoid remains the ultimate direction." Cai Bingzhen said wheeled robots' greater value in factories lies in connecting cross-workshop, cross-workstation unstructured flows, combining handling and sorting into one; as for bipedal robots, medium and large factories have flat floors, making bipedal unnecessary; only small enterprises unable to renovate factories or unstructured scenarios need bipedal, with large-scale normalized bipedal factory deployment requiring at least 5 more years.

对人形机器人行业而言，批量进厂除了分担产线工序，另一作用是获取生产环境中的作业数据。在蔡炳贞看来，将车间收集的真机数据与仿真数据结合，部署到本体上微调，再将作业数据回流训练模型，跑通这一研发数据闭环，是机器人向高难度工艺进阶的基础。高质量产线真机数据的稀缺，正成为制约具身大模型泛化能力的共性瓶颈。开放原子开源基金会副秘书长李博表示，训练通用具身基础模型通常需要千万小时级别的真机交互数据，而全球合规可用的同类数据仅约50万小时。

For the humanoid robot industry, batch factory entry beyond sharing production processes serves another purpose—obtaining operational data in production environments. In Cai Bingzhen's view, combining workshop-collected real-machine data with simulation data, deploying to bodies for fine-tuning, then returning operational data to train models—running through this R&D data closed loop—is the foundation for robots advancing to high-difficulty processes. Scarcity of high-quality production line real-machine data is becoming a common bottleneck constraining embodied large model generalization. OpenAtom Foundation Deputy Secretary-General Li Bo said training general embodied foundation models typically requires tens of millions of hours of real-machine interaction data, while globally compliant usable同类 data is only about 500,000 hours.

此次兆丰股份与乐聚智能便将15台夸父机器人在产线积累的1000小时上料真机回流数据，向开放原子开源基金会旗下的OpenLET社区开源。常琳拆解了具身模型落地的三个阶段。第一阶段是基础模型预训练，由头部大模型厂商投入算力。第二阶段是垂域模型中训练，仅需填入约1000小时训练场真机数据，即可提升场景理解能力。第三阶段则是决定商业化交付的后训练环节，开发团队先利用100至200小时真机数据开展模仿学习，可将准确率推升至95%；而最后3%至5%的精度跨越，则必须依赖产线真机强化学习。OpenLET技术委员会主任王松补充道，仿真数据只能把模型训到30至60分，训练场真机数据能带到80分左右，只有依靠包含油污、光照变化及异常摆放的产线回流数据，才能把模型成功率推向98%以上。

This time Zhaofeng and Leju Intelligent open-sourced 1,000 hours of loading real-machine return data accumulated by 15 Kuafu robots on the production line to the OpenLET community under the OpenAtom Foundation. Chang Lin broke down embodied model deployment's three stages. First is foundation model pretraining, with top large model makers investing compute. Second is vertical domain model training, requiring only about 1,000 hours of training ground real-machine data to improve scenario understanding. Third is the post-training stage determining commercial delivery—development teams first use 100 to 200 hours of real-machine data for imitation learning, pushing accuracy to 95%; the final 3% to 5% precision leap must rely on production line real-machine reinforcement learning. OpenLET Technical Committee Director Wang Song added simulation data can only train models to 30–60 points; training ground real-machine data brings them to about 80; only production line return data including oil stains, lighting changes, and abnormal placements can push model success above 98%.

从本体厂商看，能否跑通产线常态化部署与数据闭环，也关系到资本市场的审视。今年5月，乐聚智能创业板IPO申请获深交所受理。招股书显示，2025年公司营收约2.58亿元，其中夸父系列销售577台，贡献收入约1.78亿元；同年归母净亏损6977.94万元，预计2028年扭亏。此次IPO公司拟募资26亿元，其中23.7%计划投向数据集建设。蔡炳贞表示，今年以来，资本市场对人形机器人整机本体及关节模组的投资明显减少，资金正集中投向灵巧手、具身大模型、触觉传感器等最后一公里灵巧操作领域；与此同时，二级市场上市审核趋严，近期甚至有头部人形机器人企业放出新一轮融资信息，实际上暂缓了IPO进程，缺乏实际落地场景支撑的第二梯队企业推进IPO的可能性已非常低。

From the body manufacturer perspective, whether it can run through normalized production line deployment and data closed loops also relates to capital market scrutiny. In May this year, Leju Intelligent's ChiNext IPO application was accepted by the Shenzhen Stock Exchange. The prospectus shows 2025 revenue of about 258 million yuan, of which Kuafu series sold 577 units contributing about 178 million yuan; same-year net loss attributable to parent was 69.7794 million yuan, with turnaround expected in 2028. This IPO plans to raise 2.6 billion yuan, with 23.7% planned for dataset construction. Cai Bingzhen said since this year capital market investment in humanoid robot complete machine bodies and joint modules has明显 decreased, with funds concentrating on dexterous hands, embodied large models, tactile sensors, and other "last mile" dexterous manipulation fields; meanwhile, secondary market listing review has tightened—recently even a leading humanoid robot enterprise released new funding information, effectively postponing IPO; second-tier enterprises lacking actual deployment scenario support have very low IPO prospects.