In the era of humanoid robots, what is lacking is not funding or conviction, but the compelling reasons that would make factories willing to place orders. Before the arrival of ChatGPT, all narratives concerning mass production were merely anticipatory.

作者丨千浔  编辑丨九黎

人形机器人距离真正的量产还剩多远?宇树科技创始人王兴兴对这个问题给出的答案是:等一个类ChatGPT的时刻到来。泛化能力是绕不过去的门槛,而那个时刻还没有到来。

数据比口号诚实。2026年上半年，中国市场卖出约2.3万台人形机器人，其中约三成流进科研院所，三成涌上展会舞台，三成多漂洋过海出了国。真正走进工厂生产线的，不到5%。家用场景，全球加起来不满一千台。这个行业目前最大的客户，是它自己。

资本市场是另一套叙事。今年一季度机器人赛道融资超过300亿元，追平了2025年全年的融资规模；行业预测全年出货量从6万台到20万台不等，市场呈现出如烈火烹油般的热烈态势。但是王兴兴那句话之所以被反复引用，正是因为它戳中了融资PPT与真实账本之间的裂痕：硬件跑得太快，软件跟不上，商业化卡在中间。

成本：降价是真的，"能用"是另一回事

过去一年，人形机器人的价格跌幅超出所有人预期。

宇树科技的G1型号已降至8.5万元起售，松延动力的Bumi则直接杀至9998元，其价格与一台高配iPhone相等。在二手市场上，2024年购入价三四十万的工程样机如今只需三五万便可打包拉走。

降价动力来自两条线。一是供应链复用，核心零部件大量来自新能源车和消费电子的成熟产线，宇树G1整机物料成本仅4万元出头，在降价之后毛利率仍有40%。二是国产替代加速，三大核心部件国产化率升至75%到90%，特斯拉的数据最直观：不含中国供应链的Optimus物料成本13.1万美元，依托后降至4.6万美元，近3倍剪刀差。

But there is one issue that has been deliberately obscured: low-priced products and industrial-grade products are fundamentally different species.

跌破10万元的机型大多进行工程降级处理，将极端环境适应能力砍掉，同时单臂负载限制在2公斤左右，续航也仅维持在1到2小时。而真正能进工厂干活的工业级机器人，其BOM成本仍在30万元区间，算上产线改造和运维体系，落地总投入要50万以上。

尴尬的分层由此形成：消费级产品便宜了，但只能跳舞翻跟头；工业级产品能干活了，但买不起也回不了本。

A worker earns an annual salary of between 60,000 and 80,000 yuan. Factories have an expected payback period of less than two years for such robots. An industrial-grade humanoid robot costing 300,000 to 500,000 yuan, even when calculated according to Unitree’s disclosed productivity reaching 30 to 40 percent of human output, has a static payback period of around five years, not accounting for depreciation and equipment obsolescence caused by technological iteration.

文宏杰的判断值得认真琢磨。一旦误差成本降低到低于人工成本的50%，那么工厂对机器人的需求就会从项目制转变为自来水制，市场的天花板也将瞬间打开。反过来说，在这个临界点到来之前，工厂的采购逻辑不会发生变化，它仍会先通过小批量试用来验证效果，然后再决定是否扩大规模。而这个先试试的阶段，可能比所有人都预期的都要长。

进厂在加快，稳定才是真正的考验。

If cost is the issue of affordability in purchase, then the scenario question amounts to whether the purchase is worth making.

行业里存在一个共识，人形机器人商业化的起点是真正进厂打工。优必选Walker进入蔚来以及极氪产线，智元远征在富临精工工厂中搬运物料，银河通用与百达精工签署了超过一千台的订单。看起来，工业落地已在发生。

但细节经不起推敲。智元与富临精工的合作中，近百台机器人所承担的工作为物料搬运以及接收指令，并与AMR协同作用把周转箱搬到另一个位置。这本质上就是AMR与机械臂的协同就能完成的事，为什么要用人形机器人？

答案很现实：目前人形机器人能稳定胜任的场景，恰好是传统工业机器人已经覆盖的领域。

真正能体现人形机器人价值的场景，是双足移动、双手灵巧操作以及环境的自适应，然而这恰恰是它最难以胜任的领域。浙江大学机器人研究院院长朱世强说得直截了当：行业虽然热闹，却显得尴尬，其根本原因在于大脑多模态大模型以及灵巧手均尚未实现真正的突破。

A set of data better illustrates the core issue. Embodiment intelligence achieves a success rate of 89.4 percent in simulated environments. Upon entering real-world scenarios, the success rate drops to 12 percent. Although factories are more structured than households, variables such as light changes, differences in workpiece batches and occasional interferences continue to cause the strategies in controlled demos to fragment.

更深层的问题在于数据负循环。
没有大规模商用，就没有高质量真机数据；
没有数据，泛化能力就上不去；
泛化能力上不去，场景就受限；
场景受限，就更没法扩大规模。

Unitree founder Wang Xingxing states that the industry must await the ChatGPT moment. However, ChatGPT relies on nearly infinite text data from the internet, whereas the data acquisition cost for humanoid robots is extremely high. Tesla hired more than 50 people for several months to gather data so that Optimus could learn to sort batteries. The data loop of embodiment intelligence turns far slower than that of large language models.

普华永道判断，到2030年工业机械臂、轮式机器人和人形机器人将形成清晰的分工并长期共存。问题在于，人形机器人的“特定场景”到底有多大，目前没有人能给出令人信服的答案。

生态：硬件在贬值，软件才是真正的壁垒

If the cost and scenarios represent the “visible challenges,” then the developer ecosystem serves as the underestimated variable.

In the first half of 2026, leading enterprises began to shift their competitive focus from hardware to software. Zhiyuan Technology released Lingqu OS, which serves as an Android-like operating system for robots, after investing more than 2 billion yuan over five years. The Beijing Humanoid Robot Innovation Center open-sourced a complete set of technical stack comprising the embodiment, control framework, and world model. The OpenLoong open-source community has gathered more than 20,000 developers.

Logic is clear: hardware is becoming a homogenized standard product, and the real barrier lies in the software layer. Zhang Chi of New Ding Capital judges that the future winners will be two types of enterprises, one pushing hardware costs to the extreme, the other focusing on software for the brain, and the latter will very likely become the highest-value link in the industry, similar to autonomous driving suppliers in the new energy vehicle sector.

但生态建设目前面临两个硬约束。

First, the generality of cross-body data. The question of whether data collected from one body can be applied to another body remains without an answer. Data cannot be reused across bodies, so every enterprise must accumulate data from scratch, and the industry's "data flywheel" cannot turn.

第二，开发者门槛偏高。智元号称做了机器人版剪映，但工业场景开发远比模仿一个动作复杂。目前的生态繁荣更多是科研和极客层面的，距离千行百业开发者还很远。

In the humanoid robot industry, efforts are underway to replicate the smartphone model of hardware plus operating system plus application ecosystem. Yet the complexity involved far exceeds that of a mobile phone. This is because robots must not only process information but also interact with the physical world, and the uncertainty in the latter represents a difference on the order of magnitude.

量产在逼近，商业才是真正的终点

Three dimensions viewed together reveal that the cost side shows genuine price reductions, but a structural gap exists between low prices and actual usability; on the scenario side, stable landings coincide exactly with domains already covered by traditional robots, while differentiated value remains in the demo stage; on the ecosystem side, the direction is correct, yet software maturity lags far behind hardware cost reductions.

硬件的量产前夜已经到来，但商业化的量产前夜还没有。前者意味着造得出也造得便宜，而后者则意味着卖得动、用得起、回得了本。两者之间，可能还隔着两到三年。

真正值得跟踪的不是出货量数字，而是三件事：工业级产品的回本周期能否压缩到2年以内，具身大模型的泛化能力能否突破仿真到现实的落差，以及跨本体的数据标准能否在行业层面形成共识。

In the three matters, any one of them achieving a substantial breakthrough would carry far greater significance than yet another company announcing ten-thousand-unit capacity. Before the arrival of mass production, what is truly scarce is not the capacity itself, but the reason that would make such capacity meaningful.

数据来源：

智元APC2026合作伙伴大会（2026年）

中邮证券《宇树G1人形机器人拆解报告》

中邮证券于2026年3月30日发布《宇树G1人形机器人拆解报告》，对宇树科技的核心在售产品G1基础版进行了全面的硬件拆解。报告系统梳理了G1在售价、成本构成、供应链布局以及技术特点等多个维度，明确指出了其成本控制与轻量化设计的突出表现，同时也揭示了行业在技术路线选择和场景适配方面的整体特征。

G1基础版整机BOM成本预计为4.16万元，其中核心关节模组成本占比较高，达到2.75万元。叠加0.3万元的加工费用后，营业成本合计4.46万元。以税前售价7.52万元计算，预估毛利率达到40.7%。而EDU高配版本凭借更高的配置水平，毛利率进一步提升至63.5%至66.7%，显示出企业在成本端具备较强的竞争力。

报告同时对供应链布局进行了梳理，核心零部件大量采用成熟的新能源车和消费电子产线产出的器件，体现了供应链复用的优势。硬件本身并不构成高壁垒，真正突出的竞争力在于运控算法与世界模型的融合，能够支撑机器人实现高动态运动与复杂操作。

整体来看，该报告为观察人形机器人商业化路径提供了清晰的视角，硬件降本已具备现实基础，而软件壁垒仍将是未来突破的关键节点。

OFweek《2026年人形机器人行业洞察报告》

腾讯新闻《三大核心部件国产化提速》，2026年6月30日

澎湃新闻报道称，价格出现大幅下挫，网友们表示要开始动手了。

In the era of humanoid robots, what remains lacking is not funding or conviction, but the compelling reasons that would make factories willing to place orders. Before the arrival of ChatGPT, all narratives concerning mass production were merely anticipatory.

As of mid-2026, the distance to true mass production for humanoid robots remains substantial. Unitree founder Wang Xingxing provided the industry with a direct answer to this question: it will occur only when a moment equivalent to the arrival of ChatGPT arrives. Generality represents a threshold that cannot be bypassed, and that moment has yet to arrive.

Data remains more honest than slogans. In the first half of 2026, China’s market sold approximately 23,000 humanoid robots, of which about 30 percent flowed into research institutes, 30 percent entered exhibition stages, and more than 30 percent were exported overseas. In fact, fewer than 5 percent entered factory production lines, and the total for household scenarios worldwide remains below 1,000 units. The industry’s largest customer at present is still itself.

Capital markets offer another narrative. In the first quarter of 2026, financing in the robot sector exceeded 30 billion yuan, matching the full-year total for 2025; industry predictions for the year range from 60,000 to 200,000 units, producing a market that appears to be in a state of intense heat. Yet the statement from Wang Xingxing is cited repeatedly precisely because it points to the gap between financing PPT slides and actual accounts: hardware advances too quickly while software lags, and commercialization remains stalled in the middle.

Over the past year, price declines for humanoid robots exceeded expectations.

The Unitree G1 model is now offered at a starting price of 85,000 yuan, while Suoyuan Dynamics’ Bumi drops directly to 9,998 yuan, a price equivalent to a high-end iPhone. On the second-hand market, engineering prototypes purchased at 300,000 to 400,000 yuan in 2024 can now be bundled for 30,000 to 50,000 yuan.

Price declines stem from two lines. First, supply-chain reuse: core components are sourced extensively from mature production lines in the new-energy vehicle and consumer-electronics sectors, with the Unitree G1 whole-machine material cost at just over 40,000 yuan; even after price cuts, gross margins remain around 40 percent. Second, accelerated domestic substitution: the localization rate for the three core components has risen to 75 to 90 percent, as evidenced most clearly by Tesla data: without the Chinese supply chain, Optimus component cost was 131,000 USD, falling to 46,000 USD after reliance on it, producing a nearly threefold cost reduction.

But there is one issue that has been deliberately obscured: low-priced products and industrial-grade products are fundamentally different species.

Models that have dropped below 100,000 yuan have undergone extensive engineering downgrades, including removal of extreme-environment adaptability and limitation of single-arm load to around 2 kilograms, with endurance maintained only for 1 to 2 hours. In contrast, industrial-grade robots capable of performing actual factory work maintain BOM costs in the 300,000-yuan range; adding production-line modifications and maintenance systems, total landed cost exceeds 500,000 yuan.

This creates an awkward division: consumer-grade products have become affordable but can only perform dances and backflips; industrial-grade products can work but are too expensive to purchase or recover their costs.

A worker earns an annual salary between 60,000 and 80,000 yuan. Factories require a payback period of less than two years for such robots. An industrial-grade humanoid robot costing 300,000 to 500,000 yuan, even when calculated according to Unitree’s disclosed productivity reaching 30 to 40 percent of human output, has a static payback period of around five years, not accounting for depreciation and equipment obsolescence caused by technological iteration.

Wen Hongjie’s assessment is worth careful consideration. Once error costs fall below 50 percent of labor costs, factories will shift demand for robots from project-based to utility-based, and the market ceiling will open instantaneously. Conversely, before this critical point arrives, factories will continue to purchase logic that first uses small-batch trials to validate effects before deciding on scale-up. This trial phase may extend longer than most people expect.

Entry into factories is accelerating, but stability is the true test.

If cost constitutes the issue of affordability in purchase, then the scenario question amounts to whether the purchase is worth making.

A consensus exists in the industry that the commercialization starting point for humanoid robots is genuine factory deployment. UBTech’s Walker has entered production lines at NIO and Li Auto, Zhiyuan Remote Control has moved material in Fulim精工 factories, and Galaxy Universal has signed orders exceeding 1,000 units with Baida精工. It appears that industrial landing is already occurring.

Yet details do not withstand scrutiny. In the Zhiyuan and Fulim精工 cooperation, nearly 100 robots handle material movement and instruction reception while coordinating with AMR to transport turnover boxes to another location. This is in essence the kind of task already covered by AMR combined with robotic arms, so why use humanoid robots?

The answer is straightforward: the scenarios in which humanoid robots can currently perform stably are exactly those already covered by traditional industrial robots.

The true value of humanoid robots lies in bipedal locomotion, dexterous hand operations, and environmental adaptability, yet these are precisely the domains in which it is most difficult to succeed. Zhu Shiqiang, president of the Zhejiang University Robotics Institute, stated the matter directly: although the industry appears lively, it seems awkward, with the fundamental reason lying in the fact that multimodal large models for brains and dexterous hands have yet to achieve genuine breakthroughs.

A set of data better illustrates the core issue. Embodiment intelligence achieves a success rate of 89.4 percent in simulated environments. Upon entering real-world scenarios, the success rate drops to 12 percent. Although factories are more structured than households, variables such as light changes, differences in workpiece batches, and occasional interferences continue to cause the strategies in controlled demos to fragment.

A deeper issue lies in the data negative loop.

No large-scale commercialization means no high-quality real-machine data; without data, generality cannot improve; without generality, scenarios are limited; and with scenarios limited, further scale-up becomes even more difficult.

Unitree founder Wang Xingxing states that the industry must await the ChatGPT moment. However, ChatGPT relies on nearly infinite text data from the internet, whereas the data acquisition cost for humanoid robots is extremely high. Tesla hired more than 50 people for several months to gather data so that Optimus could learn to sort batteries. The data loop of embodiment intelligence turns far slower than that of large language models.

PricewaterhouseCoopers judges that by 2030 industrial robotic arms, wheeled robots, and humanoid robots will form a clear division of labor and coexist for the long term. The problem is that the size of humanoid robots’ “specific scenarios” has not yet been given by anyone in a convincing answer.

Ecosystem: hardware is depreciating, software is the true barrier

If cost and scenarios represent the visible challenges, then the developer ecosystem serves as the underestimated variable.

In the first half of 2026, leading enterprises began to shift their competitive focus from hardware to software. Zhiyuan Technology released Lingqu OS, which serves as an Android-like operating system for robots, after investing more than 2 billion yuan over five years. The Beijing Humanoid Robot Innovation Center open-sourced a complete set of technical stack comprising the embodiment, control framework, and world model. The OpenLoong open-source community has gathered more than 20,000 developers.

The logic is clear: hardware is becoming a homogenized standard product, and the real barrier lies in the software layer. Zhang Chi of New Ding Capital judges that the future winners will be two types of enterprises, one pushing hardware costs to the extreme, and the other focusing on software for the brain, and the latter will very likely become the highest-value link in the industry, similar to autonomous driving suppliers in the new-energy vehicle sector.

First, the generality of cross-body data. The question of whether data collected from one body can be applied to another body remains without an answer. Data cannot be reused across bodies, so every enterprise must accumulate data from scratch, and the industry’s data flywheel cannot turn.

Second, developer threshold remains high. Zhiyuan claims to have built a robot version of CapCut, but development in industrial scenarios is far more complex than imitating a single action. Current ecosystem prosperity exists more at the research and enthusiast levels and remains far from thousands of industry developers.

In the humanoid robot industry, efforts are underway to replicate the smartphone model of hardware plus operating system plus application ecosystem. Yet the complexity involved far exceeds that of a mobile phone. This is because robots must not only process information but also interact with the physical world, and the uncertainty in the latter represents a difference on the order of magnitude.

Viewed together, these three dimensions reveal that on the cost side genuine price reductions have occurred, but a structural gap exists between low prices and actual usability; on the scenario side, stable landings coincide exactly with domains already covered by traditional robots while differentiated value remains in the demo stage; on the ecosystem side, the direction is correct yet software maturity lags far behind hardware cost reductions.

Hardware is approaching the eve of mass production, but the eve of commercialization mass production has not yet arrived. The former means that it is possible both to produce and to produce cheaply, while the latter means that it can be sold, used, and paid back. Between the two, two to three years may still remain.

What is truly worth tracking is not shipment-volume figures but three matters: whether the payback period for industrial-grade products can be compressed to within two years, whether the generality of embodied large models can break through the gap from simulation to reality, and whether cross-body data standards can form industry-wide consensus.

In these three matters, any one substantial breakthrough would carry far greater significance than yet another company announcing ten-thousand-unit capacity. Before the arrival of mass production, what is truly scarce is not the capacity itself, but the reason that would make such capacity meaningful.

Sohu, “The Embarrassment of Humanoid Robots: Hard to Make, Difficult to Sell, What Is Sold Is Still for Display,” 21 July 2026

腾讯新闻，《全球百分之八十的人形机器人由国内进行制造，但七成的“它们”还在台上进行演戏》，2026年7月27日

- END -

In the pre-mass-production night for humanoid robots: the three pillars of cost, scenarios, and ecosystem, how many more steps remain before commercialization?

来源：人形机器人量产前夜：成本、场景、生态，商业化还差几步？ | OFweek机器人网