受访者常年在一线负责科技产业商业化，此前是一家物理 AI 基础设施公司的联合创始人。过去一年，他向几乎所有能坐下来聊的具身智能公司反复问同一个问题，你们到底怎么赚钱。为什么一直问这个问题，因为他做商业化出身，客户就是这些做本体、做方案、做模型的公司。这个位置的好处是能见到足够多的公司，而且是坐下来谈生意的那种见，不是展会上握个手。坏处是没法只当旁观者，客户如果拿不到产业侧的收入，上游的采购能力迟早会收缩，没有例外。
The interviewee has long been responsible for commercialization on the front lines of the tech industry, previously co-founder of a physical AI infrastructure company. Over the past year, he has repeatedly asked almost every embodied intelligence company he could sit down with the same question—how exactly do you make money. Why keep asking this? Because he comes from a commercialization background, and his customers are these body, solution, and model companies. The advantage of this position is seeing enough companies—the kind of seeing where you sit down to talk business, not shake hands at an expo. The downside is not being able to remain a bystander—if customers cannot get industry-side revenue, upstream procurement capacity will eventually contract, no exceptions.
差别有多大，完全是两套说法。你在展会上跟一家公司聊，听到的是一套。当你跟他成了朋友，晚上坐下来喝一杯，听到的是另一套。私下交流里，大家反复提到的并不是机器人完全干不了活，而是另一个更准确的问题，在经过充分训练、环境和物体都相对固定的任务里，机器人已经能把动作做得很流畅，但只要物品、光照、摩擦力或摆放姿态发生变化，成功率就可能明显下降。这不是对整个行业的否定，而是目前从演示走向稳定作业时，必须面对的能力边界。公开场合通常展示的是最佳结果，私下讨论的则是结果能否在不确定条件下被持续复现。
How big is the difference? Completely two sets of narratives. At an expo, you hear one set. When you become friends and sit down for a drink at night, you hear another. In private conversations, what people repeatedly mention is not that robots cannot work at all, but a more accurate problem: in tasks with sufficient training and relatively fixed environments and objects, robots can already perform movements smoothly; but as soon as objects, lighting, friction, or placement posture change, success rates can drop noticeably. This is not a denial of the entire industry, but a capability boundary that must be faced when moving from demonstration to stable operation. Public occasions usually show best results; private discussions are about whether results can be continuously reproduced under uncertain conditions.
问题不是动作做不出来。动作大部分都能做出来。你去展台上看，抓取、装配、搬运，都在动。如果任务足够确定，通过强化学习和工程调优，动作可以做到流畅而精细。真正困难的是条件发生变化之后，机器人还能不能维持相近的成功率。所以比精度不够更准确的说法，是不确定环境下的稳定性还不够。这里面当然包含定位精度，但不止定位精度。还包括，公差能不能守住，节拍能不能跟上产线，连续跑几百次的一致性会不会漂，换了一批来料还能不能干，光照和物体表面摩擦力发生变化后能不能适应，连着干三个班次会不会累积误差。演示成功一次，与在这些变化里稳定复现，是两个量级的工程问题。
The problem is not that movements cannot be made. Most movements can be made. Go to a booth and you see grasping, assembly, handling—all moving. If the task is sufficiently certain, through reinforcement learning and engineering tuning, movements can be smooth and fine. The truly difficult part is whether the robot can maintain similar success rates after conditions change. So a more accurate statement than "insufficient precision" is "insufficient stability under uncertain environments." This certainly includes positioning precision, but not only positioning precision. It also includes: can tolerances be maintained, can cycle times keep up with production lines, will consistency drift over hundreds of consecutive runs, can it still work with a new batch of incoming materials, can it adapt after lighting and surface friction change, will errors accumulate over three consecutive shifts. Succeeding once in a demonstration and stably reproducing under these variations are engineering problems of two different orders of magnitude.
而且这里有个更根本的问题，行业里其实心里都清楚，但不太愿意摆在台面上讲，人形在干活这件事上，跟传统工业自动化的差别还是很大的。传统工业自动化，是在一个高度结构化的环境里，用专用机构把一件事做到极致的精度和节拍。它的整个技术路线，就是拿灵活性换确定性。而人形、或者说通用具身，走的是反方向，拿泛化能力做主张，我什么都能干一点。但产业客户在产线上要买的，恰恰是确定性。所以你会看到一个很尴尬的错位，你的核心卖点，是他不需要的那个东西，而他真正需要的那个东西，恰恰是你为了换取卖点而牺牲掉的。
And there is a more fundamental problem here that the industry knows internally but is reluctant to put on the table: humanoid robots differ greatly from traditional industrial automation in "working." Traditional industrial automation, in a highly structured environment, uses dedicated mechanisms to achieve extreme precision and cycle time for one thing. Its entire technical route trades flexibility for certainty. Humanoid, or general-purpose embodied, goes the opposite direction—claiming generalization capability, "I can do a bit of everything." But what industrial customers buy on production lines is precisely certainty. So you see an awkward mismatch: your core selling point is what he does not need; what he truly needs is exactly what you sacrificed to trade for the selling point.
今年的公开信息基调不是这个。工信部说全年整机产量有望破十万台，十五五规划把具身智能纳进国家未来产业，头部公司出货、订单、盈利的消息一条接一条。这些数据反映了真实进展，产能在形成，政策在推进，订单也在增长。但量产元年和商业化元年不是一件事情。量产元年说的是供给侧，我能造出来、造得快、造得便宜。这件事今年确实在发生。商业化元年说的是需求侧，有人用它挣到钱了，或者用它省下了钱，而且这笔账算得过来。这两件事之间没有必然联系。更要紧的是，他进入工业场景，也不代表他在商业化上能打平。
The tone of public information this year is not this. MIIT says full-year complete machine output is expected to break 100,000 units; the 15th Five-Year Plan incorporates embodied intelligence into national future industries; leading companies' shipment, order, and profit news come one after another. These data reflect real progress: capacity is forming, policy is advancing, orders are growing. But "mass production first year" and "commercialization first year" are not the same thing. Mass production first year is about the supply side: I can build it, build it fast, build it cheap. This is indeed happening this year. Commercialization first year is about the demand side: someone has made money with it, or saved money with it, and the math works out. There is no necessary connection between these two things. More importantly: entering an industrial scenario does not mean breaking even commercially.
很多人看到人形机器人进了某汽车厂产线就认为商业化闭环了。进场只是拿到一张入场券，它证明设备已经进入验证阶段，不能证明它创造的价值已经高于成本。一台设备进了工厂，可能是在干活，也可能是在做联合创新项目、在做示范工位、在给双方的品牌部门提供素材。这三种情况，在新闻稿里长得一模一样。所以看这些消息的时候，只关心一个问题，这笔钱是从客户的哪个口袋出的。如果是从研发预算、创新预算、战略合作预算出的，那它是一次投资行为，只有从生产运营的成本预算出的，才是真的商业化。
Many people see a humanoid robot entering an automaker's production line and think the commercialization loop is closed. Entry is just getting an admission ticket—it proves the equipment has entered the validation stage, not that the value it creates exceeds its cost. A device entering a factory may be working, or may be doing a joint innovation project, a demonstration workstation, or providing material for both parties' brand departments. These three situations look identical in press releases. So when looking at this news, only one question matters: which pocket of the customer is this money coming from. If it comes from R&D budget, innovation budget, or strategic cooperation budget, it is an investment behavior; only from production operation cost budget is it true commercialization.
如果落实到一个具体工业工站，今天部署一台具身机器人，企业到底在为什么付钱。不能只看本体价格，也不能只看机器人是否完成了动作。一个工站真正需要投入的人力、财力至少包括，现场改造、数据采集、示教或遥操作、模型训练与调优、安全验证、与 MES 或产线控制系统对接，以及后续现场运维。在当前技术水平下，很多项目的主要成本不在机器人本体，而在部署机器人所需要的人。一个工站背后往往要同时投入算法、工程、产品和现场实施人员。机器人最终即使能够完成任务，也要继续追问三个问题，需要多少人工兜底，换一个工站要不要重新做一遍，持续运维成本能否低于它替代或新增的价值。如果每落一个工站都要重新投入一支团队，它可以是有价值的示范项目，但还不能算可规模复制的产品。
If it comes down to a specific industrial workstation, what exactly is an enterprise paying for when deploying an embodied robot today. You cannot look only at body price, nor only at whether the robot completed the movement. The manpower and financial resources a workstation truly requires at minimum include on-site modification, data collection, teaching or teleoperation, model training and tuning, safety validation, integration with MES or production line control systems, and subsequent on-site operations. At current technology levels, the main cost of many projects is not the robot body but the people needed to deploy the robot. Behind one workstation, algorithm, engineering, product, and on-site implementation personnel often must be invested simultaneously. Even if the robot ultimately completes the task, three questions must still be asked: how much manual fallback is needed, must it be redone for a different workstation, and can continuous operations cost be lower than the value it replaces or adds. If every workstation requires re-investing a team, it can be a valuable demonstration project but not yet a replicable product.
小米汽车工厂的公开案例可以说明这个问题。其人形机器人在自攻螺母上件工站连续自主运行 3 小时，初期双侧安装成功率为 90.2%，满足最快 76 秒的产线节拍，经过数月迭代后，成功率提升至 98%，接近人工约 99% 的水平。这组数据证明技术在快速进步，也说明一个工站从能做到接近可用需要持续迭代。但公开信息没有披露该工站投入了多少研发与现场人员、产线改造成本、人工接管率、设备利用率和全生命周期成本。因此它能够证明技术可行性，暂时还不足以单独证明商业 ROI。这正是商业分析与技术演示的区别，不仅要看成功率提高了多少，还要看为提高这几个百分点投入了什么，以及同样的投入能否复制到下一个工站。
Xiaomi Auto's public case illustrates this problem. Its humanoid robot ran autonomously for 3 consecutive hours at a self-tapping nut loading workstation, with initial bilateral installation success of 90.2%, meeting the fastest 76-second line cycle time; after months of iteration, success rose to 98%, approaching manual labor's roughly 99% level. This data proves technology is progressing rapidly and shows that a workstation going from "can do" to "near usable" requires continuous iteration. But public information does not disclose how much R&D and on-site personnel were invested, line modification costs, manual intervention rate, equipment utilization, and full lifecycle cost. So it can prove technical feasibility but is not yet sufficient to prove commercial ROI alone. This is exactly the difference between commercial analysis and technical demonstration: not only how much success improved, but what was invested to improve those percentage points, and whether the same investment can be replicated to the next workstation.
那么大家现在在做什么。搞科研。就接触到的团队而言，不少仍以研究工作为主，指标是能力边界推进了多少，而不是产品在客户现场稳定运行了多少个班次。而支撑这个状态的，是一个几乎全行业共享的期待，大家都在等具身智能的 GPT 时刻。等能力泛化到某个阈值，突然之间一个模型什么场景都能上，商业化就自然发生了。所以现在不用太纠结商业化，现在应该 all in 把能力做上去，先熬到那一天。这个判断，不现实。
So what is everyone doing now. Doing research. Based on teams encountered, many still focus on research work: the metric is how much the capability boundary advanced, not how many shifts the product ran stably at a customer site. Supporting this state is an expectation shared almost industry-wide—everyone is waiting for embodied intelligence's GPT moment. Waiting for capability to generalize past some threshold, and suddenly one model can do any scenario, and commercialization happens naturally. So no need to worry too much about commercialization now; go all in on capability, and endure until that day. This judgment is unrealistic.
先说一个前提，语言模型与物理智能面对的数据条件不同。大规模语言模型能够快速扩展，至少受益于三点，互联网文本规模足够大，语言能够被离散成 token，训练又可以利用上下文做大规模自监督学习。这使模型可以在统一目标下持续扩大数据和参数规模。物理世界并非没有逻辑，恰恰相反，它受明确的物理规律约束。难点在于观测连续、多模态且不完全，接触动力学复杂，传感器存在噪声，而且机器人的动作会改变下一时刻的状态。数据采集还受到设备、安全、场地和人工操作成本约束，无法像文本一样低成本复制。
First a premise: language models and physical intelligence face different data conditions. Large-scale language models can scale rapidly, benefiting from at least three things: internet text is large enough, language can be discretized into tokens, and training can use context for large-scale self-supervised learning. This lets models continuously expand data and parameter scale under a unified objective. The physical world is not without logic—on the contrary, it is constrained by clear physical laws. The difficulty is that observations are continuous, multimodal, and incomplete, contact dynamics are complex, sensors have noise, and robot actions change the next moment's state. Data collection is also constrained by equipment, safety, site, and manual operation costs, and cannot be replicated at low cost like text.
真实数据当然重要，但不能只看采集了多少小时、多少条轨迹。当前不少具身数据来自遥操作或动作捕捉，人的动作被记录下来，再用于训练机器人。这里至少要再问两层，采集的数据是否覆盖了决定部署成败的那些不确定性，同一批数据迁移到不同本体、不同工位时，还剩下多少有效信息。如果采集内容主要是标准条件下的成功动作，它能帮助机器人复现动作，却不一定能教会机器人处理光照变化、材质差异、抓取滑移、来料偏差和异常中断。数据量的增长不等于有效训练信息按相同比例增长。所以对数据够多、泛化自然出现这条推理同样持保留态度。关键不是有多少数据，而是这些数据有没有覆盖真正影响生产结果的边界条件。
Real data is certainly important, but you cannot look only at how many hours or trajectories were collected. Many embodied data currently come from teleoperation or motion capture, with human movements recorded and then used to train robots. Here at least two more layers must be asked: does the collected data cover the uncertainties that determine deployment success; how much effective information remains when the same batch of data migrates to different bodies and workstations. If the collection content is mainly successful actions under standard conditions, it can help robots reproduce actions but not necessarily teach robots to handle lighting changes, material differences, grasp slip, incoming material deviation, and abnormal interruptions. Data volume growth does not equal effective training information growing at the same proportion. So I am equally reserved about the reasoning that "enough data and generalization naturally appears." The key is not how much data, but whether the data covers the boundary conditions that truly affect production results.
第二层，是长时序任务的误差累积。物理动作存在先后约束，不能先拧螺丝再对位，也不能先合模再放件。机器人每一步的感知与执行偏差都会进入下一步状态，任务越长、异常分支越多，组合复杂度和恢复难度越高。语言输出出现偏差通常可以重新生成，物理执行错误则可能直接造成废件、停线或安全事故。第三层，是架构路线尚未统一。很多 VLA 使用预训练视觉—语言模型作为基础，但动作生成并没有统一路线，RT-2、OpenVLA 将连续动作量化为离散 token，复用语言模型的词表和输出层进行自回归预测，π0 则增加独立的 action expert，通过 flow matching 从噪声逐步生成连续的 action chunk，另有模型采用扩散策略或分层控制。这些都是公开论文中明确披露的客观架构差异。它们说明行业已经有多条可行路线，但尚未收敛为统一方案。接触动力学、连续控制、实时性、安全约束与异常恢复，也很难由单一模型架构全部解决。
The second layer is error accumulation in long-horizon tasks. Physical actions have sequential constraints: you cannot tighten a screw before alignment, nor close the mold before placing the part. Each step's perception and execution deviation enters the next state; the longer the task and the more abnormal branches, the higher the combinatorial complexity and recovery difficulty. Language output deviations can usually be regenerated; physical execution errors can directly cause scrap, line stoppage, or safety incidents. The third layer is that architecture routes are not yet unified. Many VLAs use pretrained vision-language models as a foundation, but action generation has no unified route: RT-2 and OpenVLA quantize continuous actions into discrete tokens, reusing the language model's vocabulary and output layer for autoregressive prediction; π0 adds an independent action expert, using flow matching to gradually generate continuous action chunks from noise; other models use diffusion policies or hierarchical control. These are objective architectural differences explicitly disclosed in public papers. They show the industry has multiple feasible routes but has not yet converged to a unified solution. Contact dynamics, continuous control, real-time performance, safety constraints, and abnormal recovery are also hard for a single model architecture to solve entirely.
把这些因素叠加起来，判断是，通用泛化能力短期内取得大幅突破并非不可能，但不应被当作商业规划中的唯一前提。这意味着具身智能商业化至少存在两条不同的路径。第一条是等待通用能力突破，继续押注模型、数据和本体能力，等到一个类似 GPT 时刻的拐点，再由通用能力向场景扩散。第二条是渐进式场景迭代，不等通用能力成熟，而是先找出现有能力已经够用的窄场景，在真实交付中解决稳定性、部署成本和 ROI，再把这个场景里得到的数据、工程经验和产品能力迁移到相邻场景。这两条路线没有谁在今天就能被证明一定正确。判断更偏向第二条。如果通用突破很快到来，等待当然是更高效的策略。但如果短期内不会到来，押注等待的每一天都在消耗现金，而且团队积累的是研究能力，不是交付能力、场景理解和客户关系。等待不是一个中性动作，它同样是在选择能力积累的方向。
Stacking these factors, the judgment is: a major breakthrough in general generalization capability in the short term is not impossible, but should not be treated as the sole premise in commercial planning. This means embodied intelligence commercialization has at least two different paths. The first is waiting for general capability breakthroughs, continuing to bet on models, data, and body capabilities, waiting for a GPT-moment inflection point, then diffusing general capability to scenarios. The second is incremental scenario iteration, not waiting for general capability to mature, but first finding narrow scenarios where existing capability is already sufficient, solving stability, deployment cost, and ROI in real delivery, then migrating the data, engineering experience, and product capability gained from this scenario to adjacent scenarios. Neither route can be proven certainly correct today. My judgment leans toward the second. If a general breakthrough comes soon, waiting is certainly a more efficient strategy. But if it does not come in the short term, every day betting on waiting consumes cash, and the team accumulates research capability, not delivery capability, scenario understanding, and customer relationships. Waiting is not a neutral action—it is also choosing the direction of capability accumulation.
