文 | 冯伟冯大白
美国当地时间10月21日，特斯拉在三季度财报中宣布，Optimus已在弗里蒙特工厂开始生产。 马斯克在随后的电话会上给出的外部交付目标仍是2027年底左右：机器人的手、耐用性和AI都还要继续改进。为了获取训练数据，特斯拉还计划造出1万至2万台机器人，让它们在“Optimus学院”练习各种任务。
车企要的是更高效地造车，也想把机器人卖给新的客户。它还可能先出现在门店：为汽车提供接待和讲解，同时观察人与机器人如何互动。
从生产线走向客户手里，机器人得经历任务训练、故障排查、设计修改和批量制造。特斯拉的计划正好把这段距离摆了出来：产线已经启动，接下来仍要造出上万台机器，让它们不断练习。
汽车业为何有条件走这段路？已有业务能承担早期投入，工厂提供真实任务，工程和供应链可以把改进做进下一代产品。
汽车工业一直在改写生产方式
汽车市场足够大，生产模式上的不断改善可以转变为可观的收益。按国际汽车制造商协会OICA公布的数据，2025年全球汽车产量约9640万辆。它覆盖乘用车和商用车，背后是庞大的零部件采购、工厂运行和售后需求。 对这个规模的产业来说，少报废一批零件、少占用一些库存、让新车型更快投入生产，都值得企业持续花钱研究。
已有业务能够为新项目提供资金，但能投入多久，仍受主业经营状况约束。小鹏今年一季度业绩发布时，何小鹏提出，用汽车业务的盈利和现金流支撑物理AI研发，再培育包括机器人在内的新收入来源。 此后，资金来源又有了扩展：据小鹏披露，其机器人业务于8月24日完成首轮股权融资，融资超过9亿美元。 这使开发风险可以由更多出资方共同承担。车企先提供技术、制造和应用条件，再为新业务引入外部资本，就有机会减轻主业独自承担长期投入的压力。融资支持的是继续开发，能否形成稳定收入，仍由产品的销售和使用决定。
这类投入在汽车工业里有一条清楚的历史线索，1913年，福特的移动装配线让车辆逐站完成装配，重新组织了大批量生产。 1961年，Unimate进入通用汽车工厂，接过压铸机卸料这类危险、重复的工作。 到智能制造阶段，试错又被提前到数字环境中：宝马用虚拟工厂检查新车型与产线的碰撞问题，将原先近四周的实体测试缩短为三天的虚拟仿真。
几个阶段改变的技术不同，企业面对的要求却相通：下一批车要造得更好，新车型要按时交付，生产成本还得有竞争力。丰田生产方式强调质量异常处理和按需生产，同样是让工序改进服从整体交付。 为了卖好下一代产品，车企必须持续改造生产系统，供应商也要跟着调整零件、工艺和交付安排。这种长期压力，才是汽车业追逐生产力变化的动力。
资金投入之后，整车企业还得把分散的技术和制造资源组织成产品。一个售价和交付目标，需要拆成零件性能、制造成本、质量要求和采购计划，才能让供应商决定做什么、投入多少。放到机器人业务里，车企的作用也应沿着这条关系判断：它提出什么任务，允许多高成本，准备怎样采购，供应商才有依据为此开发部件。庞大的产业链通过明确需求和订单预期被调动起来，行业体量才真正转化为工程力量。
汽车又是一种高度成熟的复杂消费品，工厂关心节拍、良率和停机，门店最先听到消费者对价格、设计和功能的疑问，售后网络则不断接触真实使用中的故障。这些反馈影响下一款车怎样定义、哪些零件要改、哪些功能值得继续投入。进入机器人业务后，汽车网络可以提供早期接触机会，机器人仍要收集属于自己的使用反馈。工厂里的搬运是否稳定、门店里的交流是否有效，直到家庭用户是否愿意每天使用，分别对应不同的产品要求。
当然，电子制造同样拥有庞大市场和精密制造能力。IFR公布的2024年工业机器人安装量数据显示，电子行业还略高于汽车业。 汽车值得被单独讨论，在于大规模生产需求、复杂机电产品开发、消费品牌与销售服务网络集中在同一产业中，并且持续经历产品更新。它有条件同时成为机器人的使用者、产品经营者和早期市场组织者。年轻车企也能参与这套体系，关键在于怎样运用这些条件，而非公司成立了多少年。
机器开始学习新的工作
汽车工厂并不缺机器人，固定机械臂擅长的焊接、喷涂等工作，早已形成成熟方案。今天还要研究人形机器人，是因为企业希望机器处理更有变化的任务：物品的位置会变，生产的型号会变，原本由人完成的取放、搬运和操作，也很难全部用一套固定动作覆盖。
视觉、语言与动作模型正在尝试把环境、指令和动作联系起来。Google DeepMind在2025年公布的端侧机器人模型，已展示通过少量示范适应新任务的研究进展。 对车企而言，值得下注的方向是：汽车换型、零件换位置之后，机器人能否通过训练继续工作，减少每次重新开发专用方案的负担。
人形结构的吸引力也与此有关，大量工位、工具和通道本来就是供人使用的，接近人的操作范围，使机器人有机会沿用部分现成环境。企业便可以尝试把机器送到任务面前，而不必每次围着设备重建一套工作场地。双足并非所有任务的最佳选择，稳定地面上的轮式设备也能完成许多工作；车企最终购买的仍是适合任务的方案。人形带来的空间适应性，要与动作能力一起证明价值。
宝马与Figure的合作，已经展示了真实任务怎样改变机器人。Figure 02在斯帕坦堡工厂承担钣金件上料，要从料架取件，再放进焊接夹具。根据Figure的项目复盘，在累计1250余小时的运行中，前臂是主要硬件故障点。第三代产品因此重做腕部电子结构，去掉该处的通信分发板和动态线缆。车间里跑出来的毛病，就这么进了下一代的图纸。
这项合作让宝马有机会在现有生产线上验证新设备，Figure则获得了实验室难以持续提供的故障与维护反馈。双方希望获得的回报不同：宝马需要更好用的生产工具，Figure要把改进做进自己的产品，服务更多客户。
车企自身近年的智能化投入，又使它们有机会进一步参与开发。小鹏在IRON中采用自研图灵芯片，同时为机器人新建专用产线。 芯片、计算设施和研发工具可以服务不同产品，机器人仍要建立自己的动作控制与任务能力。对企业而言，这种投入的意义不只是少建一套系统：原来用于汽车的技术资源，开始有了验证其他产品方向的机会。
车能识别道路，机器人还要学会抓取、接触和操作物体，产品面对的约束并不相同。但一家企业如果相信智能技术能够不断扩大可完成的物理任务，就有理由在汽车之外保留一个研发方向。等到机器人市场完全成熟再进入，意味着把早期的产品定义、使用反馈和合作关系都留给先行者。提前投入所购买的，既有技术进展，也有选择下一门生意的时间。
投入怎样回到生意里
一家企业同时推进多种机器人用途时，就得分配有限的研发人力和预算。门店希望机器人外形亲切、交流顺畅，工厂更关心负载、动作可靠性和连续运行。为接待优化的产品，未必适合进入生产线；针对工厂开发的能力，也未必能吸引消费者。管理层需要先确定优先服务谁、由哪个部门承担开发和试用成本，才能按相应的要求验收。否则，展示获得了关注，生产部门却用不上，项目很难靠几份各说各话的成绩单持续经营。
现代把Atlas的训练放在制造任务中，提供了一条可观察的路径。RMAC先围绕零件排序、物料搬运开展验证，正式工厂部署则安排在后续阶段。 训练期间的投入，是为未来的生产应用做准备。如果机器人能逐步承担重复搬运或负担较重的任务，工厂就可以重新安排部分岗位，把人用于需要判断、处理异常和维护设备的工作。改变人力结构，包含岗位与技能的调整，不能直接等同于裁掉多少工人。
生产部门验收的是整项工作的变化：训练、维护、人工接管和停机一共花了多少钱，合格产出是否更稳定。如果同一设备还能通过训练适应新零件和新工位，换型时就有机会减少专用改造。成本仍高、产出又不稳定时，工厂更有理由缩小任务范围、改用其他设备或延后采购。Figure在宝马现场积累的运行记录，回答了产品该怎么改；扩大采购还需要证明，机器人在这项任务上比已有方案更值得投入。
集团可以安排自己的工厂试用，外部客户却会拿机器人与人工、固定机械臂等方案比较价格、部署和维护成本。集团内部的一笔采购，不能证明集团整体获得新增收益。第二曲线要成立，产品还得让外部客户愿意持续付费。
家庭市场又是另一道关，消费者会看机器人能否完成搬运、取放等日常任务，是否好用、好修，值不值这个价。车企经营耐用消费品的经验可以帮它考虑交付和长期服务，购买理由仍要由机器人自己提供。
这种市场扩展也会改变产品设计。企业不能为每一家客户都从头做一台机器，否则销售越多，专门适配和服务的负担也越重。机器人需要逐渐形成可重复制造的硬件与可扩展的任务能力，使后续客户能够利用前面已经完成的开发。制造经验在这里的价值，是帮助新业务把产品更新与持续交付组织起来。订单和设计逐渐稳定之后，规模才有机会进一步降低成本。
在这门新生意成熟之前，机器人也可能先进入汽车的销售过程。墨甲公开介绍过在海外汽车营销服务场景中的应用，并披露其使用奇瑞的渠道、物流和服务网络。 汽车门店为机器人提供了可以接触消费者的场所，机器人则为门店增加了一种互动方式。
智能技术的很多投入，消费者平时看不见。能够自然交流、完成接待的机器人，会让一部分技术能力成为当场可体验的东西，也给销售人员增加介绍汽车的机会。
工厂试用、门店接待和对外销售没有统一的回报顺序。现代先围绕制造任务验证，小鹏与墨甲更早考虑面向人的服务场景。门店应用尤其要分清一次展示与长期运营：活动预算可以支持短期亮相，经销商愿不愿意持续承担维护和服务成本，才更接近这条渠道的真实需求。
商业目标开始分出不同路线
希望得到哪一种回报，决定了车企要承担多少机器人业务。自主开发产品、控股专业公司、向外部企业采购方案，都能让机器人进入汽车业，资金的去向和经营责任却不同。
现代控股波士顿动力，又让摩比斯参与Atlas执行器的开发与供应，能够在集团内协调产品开发、部件制造和内部应用。 它既要投入机器人的研发与制造，也要面对这项业务的外部市场。宝马与Figure合作，则可以围绕汽车生产需要提出要求、参与验证，由Figure负责机器人产品开发。宝马需要判断方案是否值得用，Figure还得判断怎样把改进卖给下一家客户。
小鹏选择自主开发IRON并建设专用产线，产品定义、研发、制造和交付的责任更多地落在自己的业务体系里。它能自行决定机器人怎样与已有智能技术协作，也需要为这条产品线组织销售和售后服务。按其公布的计划，机器人将先从门店和园区起步。 这些场景能提供交互与服务任务，内部试用之后，它还得建立面向外部客户的交付和维护能力。
墨甲借奇瑞渠道进入海外汽车营销场景，能够利用现有网点接触客户，借助物流和服务网络安排交付。但有汽车销售服务网络，不代表已经具备机器人服务能力。机器人企业仍要针对当地语言、任务和维修需求提供支持，让网点能够实际运营这类产品。
这种分工也给专业机器人公司和零部件企业留下了位置。汽车客户提出需求、参与验证，产品和部件供应商负责把经验变成可以再次销售的成果。整车企业可以从更好的生产方案中获益，供应商则可以面向其他客户继续经营。汽车产业推动机器人发展的作用，也会通过这类合作发生，不能只按车企自己卖出了多少台机器人衡量。
现代让Atlas先练零件排序，小鹏让IRON从门店和园区起步，宝马则让Figure在产线上找问题。它们已经把下一轮研发交给不同的现场：工厂不断暴露抓取、故障和停机问题，门店则会暴露交互、接待和维护问题。汽车业可以把资金、制造和销售网络都动员起来，先在哪个现场反复解决问题，就会先做出哪一类机器人。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen Wei Wei Weng White.
On 21 October, United States local time, Tesla announced in a quarterly newspaper that Optimus had begun production at the Frimont plant. The external delivery target given by Mask at the subsequent teleconference remains around the end of 2027: robotic hands, durability and AI will continue to improve. In order to obtain training data, Tesla also plans to build 10,000 to 20,000 robots to practice various tasks at the Optimus College.
It's about making cars more efficiently and selling robots to new customers. It may also appear in front of the door: providing hospitality and lectures to cars, while observing how people interact with robots.
From the production line to the customer, robots undergo mission training, fail-checking, design modifications and bulk manufacturing. Tesla's plan just put this distance out: the production line has been activated, and thousands of machines are still to be built and practiced.
Why does the automobile industry have to go this way? There are operations that can take on early inputs, factories provide real task, and engineering and supply chains can move improvements into the next generation.
The auto industry has been rewriting production methods.
The automobile market is large enough, and continuous improvements in production patterns can translate into significant gains. According to data published by the International Association of Automobile Manufacturers, the global vehicle production in 2025 was about 96.4 million. It covers passenger and commercial vehicles and is backed by large spare parts purchases, plant operations and after-sale requirements. For an industry of this size, the reduction of spare parts, the reduction of stocks and the faster entry of new vehicles into production are worthy of continuous research.
Business already exists to fund new projects, but for how long it can be invested, it remains subject to the business situation. When Peng performed publish in the first quarter of this year, Peng Ho offered to use the profits and cash flow of the automobile business to support the development of physical AI and to foster new sources of income, including robots. Since then, funding has expanded: According to Peng, its robotic business completed its first round of equity financing on 24 August, with over $900 million. This makes development risk more affordable. The provision of technology, manufacturing and application conditions, followed by the introduction of external capital for new operations, provides an opportunity to relieve the burden of long-term inputs on the industry alone. Financing supports continued development, and the ability to generate a stable income is still determined by the sale and use of products.
This type of input has a clear historical clue in the automobile industry, where in 1913 Ford ' s mobile assembly line allowed vehicles to be assembled on a station-by-station basis and organized large-scale production. In 1961, Unimate entered the General Motors Plant and took over hazardous and repetitive work such as the unloading of casting machines. At the intellectual manufacturing stage, the test error was advanced to the digital environment: the BMW used a virtual factory to examine the problem of collisions between the new model and the production line, reducing the physical test, which had been around it, to a three-day virtual simulation.
The technologies that have changed in several stages are not the same: the next batch of cars is better built, new models are delivered on time and production costs are competitive. Toyota production methods, which emphasize quality anomalies and demand-driven production, also make process improvements subject to overall delivery. In order to sell the next generation of products, companies must continuously adapt their production systems and suppliers must adjust spare parts, processes and delivery arrangements accordingly. This long-term pressure is the driving force behind productivity change in the automobile industry.
With the investment of funds, the whole car enterprise also had to organize its decentralized technology and manufacturing resources into products. A sales price and delivery target would require the removal of spare parts performance, manufacturing costs, quality requirements and procurement plans to allow suppliers to decide what to do and how much to invest. In robotic business, the role of the car company should also be judged along the lines of what it proposes, what it allows for, how high costs, and how it intends to procure, so that the supplier has a basis for developing components for this purpose. Large industrial chains are expected to be mobilized through clear demand and orders, so that industry volumes can truly be transformed into engineering power.
Automobiles are also a highly sophisticated and complex consumer, with factories concerned with rhythm, probabilities and downtimes, doors first hearing consumer questions about prices, design and functionality, and after-sales networks constantly exposed to failures in real use. These feedbacks affect how the next car is defined, which parts are to be modified and which functions are worth continuing input. Once in robotic business, the vehicle network can provide early exposure, and robots continue to collect their own use feedback. The stability of the handling in the plant and the effectiveness of the communication in the door shop are maintained until household users are willing to use it on a daily basis, depending on the product requirements.
Of course, electronic manufacturing also has large markets and sophisticated manufacturing capabilities. The 2024 industrial robot installation data published by IFR show that the electronics sector is slightly higher than the automobile industry. Automobiles deserve separate discussion in the context of large-scale production demand, the development of complex electromechanical products, the concentration of consumer brands and distribution service networks in the same industry and continuous product renewal. It has the conditions to become both a user of robots, a product operator and an early market organizer. Young car companies can also participate in the system, and the key is how to apply these conditions rather than how many years the company has been in existence.
The machine begins to learn new jobs.
Automobile factories do not lack robotics, fixed mechanical arms are good at welding, spraying, etc., and mature programmes have been developed. Today, human robots are also being studied because businesses want machines to deal with more dynamic tasks: things change in location, models change in production, human-made extraction, handling and operation, and it is difficult to cover them all with a fixed set of actions.
Visual, linguistic and action models are trying to link environment, command and action. Google DeepMind, published in 2025, has demonstrated the progress of research to adapt to new missions through a small demonstration. In the case of automobiles, it is worth noting whether robots can continue to work through training to reduce the burden of each redevelopment of a dedicated programme, following a remodelling of cars and a change of location of parts.
The attraction of human structures is also related to the fact that a large number of places, tools and corridors are originally available for human use, close to the human sphere of operation, giving robots the opportunity to follow part of the off-the-shelf environment. Businesses can try to bring machines to the task without having to re-establish a workspace around the equipment each time. Double foot is not the best option for all tasks, and there is much work that can be done by stabilizing the wheeled equipment on the ground; ultimately, the vehicle company is still buying a mission-appropriate option. The spatial adaptability of a human face must be demonstrated with mobility.
BMW and Figure have shown how real task changes robots. Figure 02 is responsible for the gold material at the Spatánburg plant, which is to be taken from the material shelf and placed in the welding clamp. According to Figure ' s Project Reassembly, the forearms were the main hardware failure point during the cumulative operation of more than 1250 hours. The third generation product was thus re-structured with the wrist electronics, removing the communication distribution boards and dynamic cables. That's what happened in the workshop.
This cooperation provided the BMW with the opportunity to validate new equipment on existing production lines, while Figure received feedback on failures and maintenance that the laboratory could not sustain. The returns are different: the BMW needs better means of production, and Figure wants to improve into its own products and serve more customers.
The intelligent input of the car companies themselves in recent years has given them the opportunity to become more involved in development. Peng uses a self-study Turing chip in Iron, while creating a new dedicated line for robots. Chips, computing facilities and research and development tools can serve different products, and robots still have to build their own operational control and mission capabilities. For an enterprise, the value of such input is not just to build a system less: The technical resources originally used for cars are beginning to have the opportunity to validate the direction of other products.
Cars recognize roads and robots learn to capture, contact and operate objects, and products are subject to different constraints. However, if an enterprise believes that smart technologies can expand the physical tasks that can be accomplished, it is justified to retain a research and development orientation outside the vehicle. Waiting for the full maturity of the robotics market to enter means leaving early product definitions, user feedback and partnerships to pioneers. Both technological advances and the timing of the choice of the next business are among those purchased in advance.
How do I get back into business?
When an enterprise promotes multiple robotic uses at the same time, it allocates limited R & D manpower and budget. The door shop wants the robot to be familiar and well-communicated, and the factory is more concerned with loads, operational reliability and continuous operation. It is not always appropriate to enter production lines in order to accommodate optimal products; nor is the ability to develop plants attractive to consumers. Management needs to determine who will prioritize services and which department will bear the development and pilot costs before it can meet the corresponding requirements of acceptance check. Otherwise, the demonstration received attention and the production sector did not use it, and it was difficult for the project to continue on several separate report cards.
The modern placement of Atlas training in manufacturing tasks provides an observable path. RMAC first performs validation around the sorting of parts, material handling, and the formal plant deployment is scheduled for a subsequent phase. Inputs during the training were prepared for future production applications. If robots are able to gradually undertake repetitive or more burdensome tasks, the plant can reposition some of its jobs and use people for work requiring judgement, handling anomalies and maintenance equipment. Changing the manpower structure to include job and skills adjustments cannot be directly equated with the number of workers laid off.
The production sector acceptance check is a change in the whole process: training, maintenance, manual takeovers and decommissioning costs, and more stable qualified output. If the same equipment can also be adapted to new parts and places through training, there is an opportunity to reduce specialized modifications when remodelled. When the costs remain high and the output is unstable, the plant is more justified in reducing the scope of the mandate, moving to other equipment or delaying procurement. Figure's record of operations at the BMW site answered how the product should be modified; the expanded procurement also required proof that robots were more worthy of input in this task than had already been done.
The group could arrange for its own plant to be tested, while external customers would compare the price, deployment and maintenance costs of robots with programmes such as manual, fixed machine arms, etc. An intra-group procurement cannot prove that the group as a whole receives additional proceeds. If the second curve is to be established, the product also has to make external clients willing to continue with paid.
The household market is another point where consumers look at whether robots can perform routine tasks such as handling, removal, etc., and whether they are useful, well repaired or not. The experience of a car company operating durable consumer goods can help it to consider delivery and long-term services, and the reason for purchase remains to be provided by the robot itself.
Such market expansion also changes product design. Businesses cannot make a machine for every customer from scratch, otherwise the more they sell, the greater the burden of specialized matching and servicing. Robots need to develop reusable hardware and scalable mission capabilities to enable subsequent clients to take advantage of the developments that have already been completed. The value of manufacturing experience here is to help new businesses organize product renewal and continuous delivery. Once orders and designs have stabilized, there is an opportunity to further reduce costs.
The robot may also enter the car sales process before the new business matures. Ink has publicly presented applications in overseas car marketing services and has disclosed its access, logistics and service networks using Chiri. Car door shops provide access to consumers for robots, while robots add an interactive way to the door shop.
Many inputs to smart technology are often invisible to consumers. A robot that can communicate naturally and complete its reception will make a part of technological capability an experience on the ground and will increase the opportunities for vendors to introduce cars.
There is no uniform order of return for factory trials, reception and external sales. Modernism revolves around manufacturing mission validation, and peng and ink are thinking earlier about a human-oriented service scenario. In particular, the door shop application distinguishes between a demonstration and a long-term operation: the activity budget can support short-term visibility, and distributors are willing to afford the cost of maintenance and services on a sustainable basis to be closer to the real demand of this channel.
Business targets are starting to divide different routes.
The kind of return that you want to get, determines how much robotic business the car company has to take on. Autonomous development of products, holding professional companies, and procurement programmes from outside firms allow robots to enter the automobile industry, with different funding and responsibilities.
Modern holding of Boston power and the involvement of Mobis in the development and supply of Atlas implementers allowed for the coordination of product development, component manufacturing and internal applications within the group. It has to invest in robotic research and development and face the external market for this business. The BMW, in cooperation with Figure, can request and participate in the validation of automobile production needs, with Figure in charge of robotic product development. The BMW needed to judge whether the programme was worth it and Figure had to judge how the improvements would be sold to the next customer.
Peng has chosen to develop Iron on its own and to build dedicated production lines, and responsibility for product definition, R & D, manufacturing and delivery is increasingly vested in its own business system. It makes its own decisions about how robots can work with existing smart technologies, and it also needs to organize sales and after-sale services for this product line. According to its published plan, robots will start from door stores and parks. These scenarios provide interactive and service-oriented tasks and, after internal testing, they also have to build delivery and maintenance capacity for external clients.
The Mercer Chiri channel has access to overseas vehicle marketing scenes and is able to use existing sites to reach customers and arrange deliveries through logistics and service networks. However, the existence of a network of automobile sales services does not mean that robotic services are already available. Robotic enterprises still need to support local language, mission and maintenance needs so that the site can actually operate such products.
This division of labour also leaves a place for professional robotics companies and spare parts enterprises. Automobile customers demand and participate in validation, and suppliers of products and components are responsible for translating experience into results that can be resold again. Car-wide enterprises could benefit from better production programmes and suppliers could continue to operate with other customers. The role of the automobile industry in promoting robotics also occurs through this type of cooperation, which cannot be measured solely by how many robots the car companies themselves sell.
Modernity has allowed Atlas to pre-sort the parts, Peng to allow Iron to start from the door shop and the park, and BMW to ask Figure for problems on the production line. They have handed over the next round of R & D to different sites: factories are constantly exposed to problems of capture, failure and shutdown, and shops are exposed to problems of interaction, reception and maintenance. The automobile industry can mobilize funds, manufacturing and distribution networks, and first create robotics where problems are repeated.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.