机器人已经能做很多动作。旁边的人想问的是另外几件事。Robots can already do a lot of moves. The people standing next to them want to know other things.
站在人形机器人旁边的陌生人，怎么知道它要做什么，怎么让它停下，出了问题找谁？研究关注公开资料如何说明机器人的动作、安全边界、控制方式、数据去向和责任归属。If you are standing next to a humanoid robot, how do you know what it is about to do, how do you stop it, and who do you ask when something goes wrong? The research examines how public materials explain robot actions, safety boundaries, control modes, data handling and accountability.
假想客户：智元（AgiBot）。研究和验证用的都是公开资料；决策和建议是推断。Hypothetical client: AgiBot. Research and validation use public sources only; the decisions and recommendations are inferences.
桌面研究有三个发现Desk research gave three findings
这是桌面研究，没有一手访谈。This is desk research. There are no first-hand interviews.
智元出货已超过宇树，但按 IDC，近七成出货还在科研、展演和数据采集AgiBot now out-ships Unitree, but by IDC’s count nearly 70% of shipments still go to research, shows and data collection
单位：台。横轴从 5,000 开始。Units shipped. The axis starts at 5,000.
来源：SAG 与 Counterpoint；日期与链接见附录证据。Sources: SAG and Counterpoint; dates and links in the evidence appendix.
科研、展演和数据采集占出货量，从 2025 年的 83.8% 降到 2026 年上半年的 69%，仍然是最大的一块。Research, shows and data collection as a share of shipments fell from 83.8% in 2025 to 69% in H1 2026. It is still the largest share.
来源：宇树 X 官方账号页面、小红书高赞笔记；日期与链接见附录证据。Source: Unitree official X account page, top-liked Xiaohongshu posts; dates and links in the evidence appendix.
评论在问安全、真假和隐私Comments ask about safety, whether it is real, and privacy
来源：小红书评论抽样；帖子与读取日期见附录证据。Source: Xiaohongshu comment sample; posts and read dates in the evidence appendix.
样本是按点赞排序的首屏评论，每个帖子约 10 到 15 条。样本小，置信度低，只能说这些话题出现过，说不出占多少。The sample is the first screen of comments sorted by likes, about 10 to 15 per post. It is small and confidence is low: these topics appeared, but how common they are is unknown.
文字条款里写得出承诺。现场看得见的行为承诺，本次核对的公开材料中未见。The written terms can state a commitment. A commitment you can see in the moment, such as in lights, motion or sound, was not found in the public material checked.
来源：对公开页面的定义与评分；依据与方法见附录评分。Source: definitions and ratings of public pages; evidence and methods in the ratings appendix.
N/E 表示证据不足，不是 0 分。N/E means there was not enough evidence to score. It is not a zero.
评分说明About the ratings
这些分数是判断，不是测量：经一轮锚点澄清，未经访谈校准，第二轮不是盲评。These scores are judgments, not measurements: the anchors were clarified once, they have not been calibrated with interviews, and the second round was not blind.
四条趋势Four trends
以下四条趋势来自公开资料的桌面研究，没有一手访谈。These four trends come from desk research on public sources, with no first-hand interviews.
出货放量、价格下探，但口径差异很大Shipments are up and prices are down, but the counts differ by agency
出货在放量，价格在下探；但“全球出货”是多少，取决于哪家机构、怎么数。Shipments are growing and prices are falling, but how many “global shipments” there are depends on which agency counts, and how.
置信度：中Confidence: Medium
来源：IFR（2026-09）、IDC（经央视，2026-01-25）、Counterpoint（2026-08-20）；链接见附录证据。Sources: IFR (2026-09), IDC (via CCTV, 2026-01-25), Counterpoint (2026-08-20); links in the evidence appendix.
细节：事实、口径与推断Details: facts, scope and inference
这些是全行业的出货数字，不是公司数字。IFR 与 IDC 的口径不同：IFR 只统计高于 140 厘米的商用机型；IDC 和 Counterpoint 含科研、展演和数据采集。三个数字时期也不同，不能相加，也不能互相比较。These are industry-wide shipment figures, not company figures. IFR and IDC count differently: IFR counts only commercial models taller than 140 cm, while IDC and Counterpoint include research, demos and data collection. The periods differ too, so the figures cannot be added or compared.
- 2026 上半年，IDC 估算全球出货约 2.5 万台，Counterpoint 估算超 22,000 台，并预测 2026 全年超 50,000 台。（IDC，经澎湃，2026-10-01；Counterpoint，2026-08-20）For H1 2026, IDC estimates about 25,000 units shipped worldwide and Counterpoint more than 22,000, with a 2026 full-year forecast above 50,000. (IDC via The Paper, 2026-10-01; Counterpoint, 2026-08-20)
- 2025 年，IFR 统计全球售出约 7,000 台全尺寸人形机器人（高于 140 厘米，面向商用/专业，不含研发与娱乐）；IDC 估算约 1.8 万台。（IFR，2026-09；IDC，经央视，2026-01-25）For 2025, IFR counts about 7,000 full-size humanoids sold worldwide (above 140 cm, commercial or professional, excluding R&D and entertainment); IDC estimates about 18,000. (IFR, 2026-09; IDC via CCTV, 2026-01-25)
- 宇树 G1 官网标价 8.5 万元起；租赁日租标价从 2025 年初最高超 2 万元，降到 2026 年 10 月部分商家的 300 元起。（宇树 G1 官网，2026-10-28；新浪财经，2026-10-09）Unitree lists the G1 from CNY 85,000. Daily rental listings fell from above CNY 20,000 at the peak in early 2025 to CNY 300 and up at some merchants in October 2026. (Unitree G1 page, 2026-10-28; Sina Finance, 2026-10-09)
- 高盛估算，制造成本从每台约 5 万–25 万美元降到约 3 万–15 万美元（2023 到 2024 年的数据，较旧，只说明方向）。（高盛，2024-02-27）Goldman Sachs estimates manufacturing cost fell from about USD 50,000–250,000 per unit to about USD 30,000–150,000 (2023 to 2024 data, now old, so it shows direction only). (Goldman Sachs, 2024-02-27)
推断Inference推断Inference每次引用出货量都写明“机构、口径、日期”，并把“商用/专业”与“科研/展演”分开写。价格叙事少讲“便宜”，改讲“总拥有成本”（购买、维护、租赁），因为租赁价格已在下降，单讲低价容易变成价格战。Whenever shipments are quoted, state the agency, the scope and the date, and keep commercial or professional use apart from research or demos. Talk less about “cheap” and more about total cost of ownership (purchase, maintenance, rental): rental prices are already falling, and a pure low-price message can turn into a price war.
落地场景在慢慢转移，规模化重复采购尚未出现Use cases are shifting slowly, and repeat bulk buying has not appeared
有数据支持的，是展演和科研向工业与服务缓慢移动；家庭更多是期待。The data supports a slow move from demos and research toward industry and services. The home is mostly still an expectation.
置信度：中Confidence: Medium
来源：IDC（经澎湃，2026-10-01）、Counterpoint（2026-08-20）、The Robot Report（2026-10-26，转述 GXO 财报会）；链接见附录证据。Sources: IDC (via The Paper, 2026-10-01), Counterpoint (2026-08-20), The Robot Report (2026-10-26, relaying the GXO earnings call); links in the evidence appendix.
细节：事实、口径与推断Details: facts, scope and inference
IDC 只有两个时点，期长也不同（全年与上半年），只能说方向，不能说“持续下降”。SAG 的分类不同，估算工业和商业应用已超过七成。IDC gives only two points in time, covering different periods (a full year and a half year), so it shows direction only, not a steady decline. SAG classifies differently and puts industrial and commercial uses above 70%.
- IDC：科研教育、性能展示、政府（数据采集）合计占出货的比例，2025 全年 83.8%，2026 上半年 69%。只有两个时点，期长不同。（IDC，经澎湃，2026-10-01）IDC: research and education, demonstration and government (data collection) together made up 83.8% of shipments in full-year 2025 and 69% in H1 2026. Only two points in time, covering different periods. (IDC via The Paper, 2026-10-01)
- Counterpoint：2026 上半年服务与导览约占 19%，智能制造约 13%，仓储物流约 5%。（Counterpoint，2026-08-20）Counterpoint: in H1 2026, services and guided tours were about 19% of shipments, smart manufacturing about 13% and warehousing and logistics about 5%. (Counterpoint, 2026-08-20)
- SAG：工业和商业应用 2026 上半年超过 70%，一年前约 50%。分类与 IDC、Counterpoint 不同，不能直接对比。（SAG，2026-08-10）SAG: industrial and commercial applications were more than 70% of shipments in H1 2026, up from about 50% a year earlier. Its categories differ from IDC’s and Counterpoint’s, so the figures cannot be compared directly. (SAG, 2026-08-10)
- IFR 指出多数应用“仍属专用，常需要人工遥操作”，“缺乏强有力的商业案例”是工业场景的主要限制。（IFR，2026-09）IFR says most applications are still specialised and often need human teleoperation, and that the lack of a strong business case is the main limit in industrial use. (IFR, 2026-09)
- GXO（Agility 的客户）CEO 在 2026 年 8 月的财报会上称，已做 45 个人形机器人试点，尚未实现 ROI，还要大约两年。这是对财报会发言的转述，未取得原始记录。（The Robot Report，2026-10-26）GXO, an Agility customer, said on its August 2026 earnings call that it has run 45 humanoid pilots, has not yet achieved ROI and needs about two more years. This is relayed; the original call record was not obtained. (The Robot Report, 2026-10-26)
- 家庭：TechWalker 2025 年报道称傅利叶 GR-3“并非为家庭场景设计”，傅利叶 2026 年 CES 新闻稿则把家庭列入适用环境；优必选 U1 定位不承担家务；高盛 2024 年基准预测，2030 年超 25 万台出货，“几乎全部为工业用途”。（傅利叶 GR-3，TechWalker，2025-08-06；傅利叶 CES 新闻稿，2026-01-07；优必选 U1，IT之家，2026-06-30；高盛，2024-02-27）Home: TechWalker reported in 2025 that Fourier’s GR-3 is not designed for home settings, while Fourier’s 2026 CES release lists homes among its settings; UBTech positions the U1 as not doing housework; Goldman Sachs’s 2024 base case is more than 250,000 shipments in 2030, almost all industrial. (Fourier GR-3, TechWalker, 2025-08-06; Fourier CES release, 2026-01-07; UBTech U1, ITHome, 2026-06-30; Goldman Sachs, 2024-02-27)
推断Inference推断Inference对外只承诺“已有客户在跑的那个场景”，并写明小时数、任务类型、是否遥操作；家庭场景用“探索中/试点”，而不是“即将”。在页面和发布内容里，把“试点”和“常态运营”分开标。Promise publicly only the scenario a customer is already running, and state the hours, the task type and whether it is teleoperated. For homes, say “exploring” or “pilot”, not “coming soon”. In pages and releases, label pilots and routine operation separately.
2027 年前后有规则可依，服务和家庭场景仍是空白Rules arrive around 2027, but services and homes still have no humanoid standard
安全从加分项变成准入项；但 ISO 25785-1 只覆盖工业场景，商场和家庭没有人形专用的 ISO 标准。Safety is moving from a bonus to a condition of entry. But ISO 25785-1 covers only industrial settings, so malls and homes have no humanoid-specific ISO standard.
置信度：高Confidence: High（ISO、工信部、欧盟机械法规的日期）(the ISO, MIIT and EU Machinery dates)置信度：中Confidence: Medium（(EU) 2026/1744 的细节，转引自英国政府说明备忘录，欧盟原文未见）((EU) 2026/1744 details, quoted from the UK government explanatory memorandum; EU original not seen)
来源：ISO、工信部、新华社、中国经济网、欧盟委员会；(EU) 2026/1744 转引自英国政府说明备忘录（2026-08-03），欧盟原文未见；日期与链接见附录证据。Sources: ISO, MIIT, Xinhua, China Economic Net, European Commission; (EU) 2026/1744 quoted from the UK government explanatory memorandum (2026-08-03), EU original not seen; dates and links in the evidence appendix.
细节：事实、口径与推断Details: facts, scope and inference
空心圆是 ISO，实心圆是中国，方块是欧盟。ISO/CD 25785-1 仍是委员会草案，最终内容和发布时间都可能变化。(EU) 2026/1744 的内容和 2028-08-02 的期限转引自英国政府说明备忘录（2026-08-03），欧盟原文未见。Hollow circles are ISO, solid circles are China, squares are the EU. ISO/CD 25785-1 is still a committee draft, and its final content and release date may change. The (EU) 2026/1744 details and the 2028-08-02 deadline are quoted from the UK government explanatory memorandum dated 2026-08-03; the EU original was not seen.
- ISO/CD 25785-1（动态稳定工业移动机器人，含腿式）：2026-05-08 登记委员会草案，2026-07-08 征求意见截止，尚未发布；适用范围是公众被排除或限制进入的工业环境。（ISO/CD 25785-1，2026-05-08）ISO/CD 25785-1 (dynamically stable industrial mobile robots, including legged): committee draft registered on 2026-05-08, comments closed 2026-07-08, not yet published; it applies to industrial environments from which the public is excluded or restricted. (ISO/CD 25785-1, 2026-05-08)
- 中国：首批 6 项人形机器人推荐性国家标准 2026-09-28 发布，2027-04-01 起实施；工信部 2026-08-24 公开征求标准体系建设指南意见，目标到 2028 年完成至少 100 项关键标准，征求稿写明“目前国内外暂无正式发布的人形机器人标准”。（中国经济网，2026-10-16；同花顺财经，2026-10-23；工信部科技司，2026-08-24；新华社，2026-08-24）China: the first 6 recommended national standards for humanoids were published on 2026-09-28 and take effect on 2027-04-01. On 2026-08-24 MIIT opened a standards-system guide for comment, aiming for at least 100 key standards by 2028; the draft says no humanoid standard has been formally published at home or abroad. (China Economic Net, 2026-10-16; Tonghuashun Finance, 2026-10-23; MIIT, 2026-08-24; Xinhua, 2026-08-24)
- 欧盟：《机械法规》(EU) 2023/1230 自 2027-01-20 起强制适用。（欧盟委员会 Machinery 页面，2026-10-30）EU: the Machinery Regulation (EU) 2023/1230 applies on a mandatory basis from 2027-01-20. (European Commission, Machinery page, 2026-10-30)
- (EU) 2026/1744：转引自英国政府说明备忘录（2026-08-03），欧盟原文未见。备忘录称，该法规把机械移出 AI 法案的直接适用清单，委员会须在 2028-08-02 前通过授权法案，把 AI 安全要求写进机械法规。以上内容均引自该备忘录，欧盟原文未见。（英国政府说明备忘录，2026-08-03）(EU) 2026/1744: quoted from the UK government explanatory memorandum (2026-08-03); the EU original was not seen. The memorandum says the regulation moves machinery out of the AI Act’s directly applicable list, and the Commission must adopt a delegated act by 2028-08-02 to bring AI safety requirements into the Machinery Regulation. All of this is quoted from that memorandum; the EU original was not seen. (UK Government explanatory memorandum, 2026-08-03)
推断Inference推断Inference把“符合什么标准、测过什么”写成品牌页的固定内容（标准号、实施日期、测试范围），不要只写“安全”。面向欧洲的产品，先核对 2027-01-20 的合规路径；面向商场或家庭的说法，明确写“不适用 ISO 25785-1（仅工业场景）”，不要暗示已获该标准认证。Make “which standard it meets and what was tested” fixed content on the brand page (standard number, effective date, test scope), not just “safe”. For Europe, check the compliance route for 2027-01-20 first. For mall or home claims, say plainly that ISO 25785-1 does not apply (industrial settings only), and do not imply certification to it.
证明点转向运行数据，但行业总判断证据不足Proof is moving to operating data, but the industry-wide claim is thinly supported
个别公司和机构开始用运行小时数和完成量做证明；中国社媒上，头部内容仍以展演为主。A few companies and institutions now prove value with operating hours and completed work, while top content on Chinese social media is still mostly demos.
置信度：中Confidence: Medium（事实部分）(the facts)置信度：低Confidence: Low（“整个行业叙事正在转向”的总判断）(the claim that the whole industry’s narrative is shifting)
来源：The Robot Report（2026-10-26）、Allianz GI（2026-09-02）；社媒抽样见上文发现 1 和 2；链接见附录证据。Sources: The Robot Report (2026-10-26), Allianz GI (2026-09-02); the social media sample is in findings 1 and 2; links in the evidence appendix.
细节：事实、口径与推断Details: facts, scope and inference
单个事实可查，置信度为中；但“整个行业的叙事正在转向”这个总判断，只有个别案例和机构观点，没有对品牌内容的系统统计，所以标为低置信度，不给比例。Each fact can be checked and is rated medium. The claim that the whole industry’s narrative is shifting rests on single cases and an institutional view, with no systematic count of brand content, so it is rated low and no share is given.
- Agility（美国）对外的证明点是运行数据：Digit 累计超 65,000 小时真实运行，在 GXO 单点搬运超 10 万个料箱，并披露客户扩展部署。这是公司自述，经 The Robot Report 转述。（The Robot Report，2026-10-26）Agility (US) leads with operating data: Digit has over 65,000 hours of real operation and has moved over 100,000 totes at one GXO site, and it reports customers expanding deployments. This is the company’s own account, relayed by The Robot Report. (The Robot Report, 2026-10-26)
- Allianz Global Investors 建议关注运行时间、任务完成率、人工干预频率、维护成本和复购，而不是“杂技视频”。这是机构观点，不是统计数据。（Allianz GI，2026-09-02）Allianz Global Investors advises watching uptime, task completion rate, human intervention frequency, maintenance cost and repeat purchase, rather than acrobatics videos. This is an institutional view, not statistics. (Allianz GI, 2026-09-02)
- IFR 指出多数应用仍需遥操作。本页的社媒抽样见上文发现 1 和 2：头部内容仍以展演为主，评论里有真实性怀疑和安全联想。（IFR，2026-09）IFR notes that most applications still need teleoperation. This page’s social media sample is in findings 1 and 2: top content is still mostly demos, and comments voice doubts about authenticity and associations with injury. (IFR, 2026-09)
推断Inference推断Inference每个主打视频配一行“怎么拍的”（自主还是遥操作、是否一条过、是否在客户现场）。在展演之外，固定输出一类“运行记录”内容：任务、小时数、失败与恢复，数字必须有出处。Add one line to every flagship video on how it was shot (autonomous or teleoperated, one take or not, on a customer site or not). Besides demos, publish a regular type of operating-record content: task, hours, failures and recovery, with every number sourced.
中国人形机器人已经证明“它能”；接下来要回答“它会怎样做、为谁做、出了问题怎么办”。Chinese humanoids have proved they can; next comes how they act, for whom, and who responds when things go wrong.
三个决策Three decisions
三个决策都是推断。All three decisions are inferences.
决策一：先做哪个市场Decision 1: which market to start with
先做运营方管得住的有人公共场所，以长隆这类运营方为灯塔；工厂并行，当第二个证明场；家庭不放进主叙事。Start with people-filled public venues that an operator controls, using Chimelong-type operators as the lighthouse. Run factories in parallel as a second proving ground. Keep the home out of the main story.
置信度：中Confidence: Medium（事实：公司口径的部署案例、机构占比）(the facts: company-stated deployments, agency shares)置信度：低Confidence: Low（决策本身）(the decision itself)
依据：智元创新日披露的部署案例（公司口径，2026-10-29）、Counterpoint、买家与运营方的二手发言；顺序是推断。Basis: deployment cases AgiBot gave at its Innovation Day (company statements, 2026-10-29), Counterpoint, and secondhand buyer and operator remarks. The order is an inference.
细节：事实、推断与依据Details: facts, inference and basis
圆点只表示先后顺序，不表示规模。长隆超过 300 台，近 500 台为单一来源；两者都是公司口径，只有智元一家的说法。灯塔的判据是经运营考核后留下来的岗位，不是展演热度，也不是台数。The dots show order only, not size. Chimelong is over 300 units, and the close-to-500 count is a single source; both are company statements from AgiBot alone. The test for a lighthouse site is posts that stay after operator review, not demo buzz or unit counts.
- 2026-09-24 智元称横琴长隆飞船乐园常态化上岗机器人超过 300 台，是公司口径。（IT之家，智元第 20,000 台下线，2026-09-24）On 2026-09-24 AgiBot said the Hengqin Chimelong spaceship park has more than 300 robots in routine service. This is a company statement. (ITHome, AgiBot 20,000th unit, 2026-09-24)
- 2026-10-29 创新日智元又称长隆具身智能主题乐园近 500 台常态化部署，计划年底增至 1,000 台、覆盖 100 多个岗位。近 500 台为单一来源的公司口径，未见长隆方或第三方确认。两处口径不完全相同，是否同一园区未核，这里不比较增幅。（投资界，智元创新日，2026-10-29；潮新闻，智元创新日，2026-10-29）At its 2026-10-29 Innovation Day AgiBot said the Chimelong embodied-AI theme park runs close to 500 robots as routine, with plans for 1,000 by year end across more than 100 posts. The close-to-500 count is a single-source company statement, and neither Chimelong nor a third party has confirmed it. The two statements are not worded identically, and whether they describe the same park was not checked, so no growth is compared here. (Pedaily, AgiBot Innovation Day, 2026-10-29; Tidenews, AgiBot Innovation Day, 2026-10-29)
- 公司披露：现场机器人与人员配置约 3:1，目标 10:1、理想 20:1；每台机器人接受运营考核，体验人数下降的节目会被淘汰。（搜狐转载，智元披露，2026-09-29）The company disclosed a ratio of about 3 robots to 1 staff member on site, with a target of 10:1 and an ideal of 20:1. Each robot is reviewed by operations, and shows with falling visitor numbers are dropped. (Sohu, AgiBot disclosure, 2026-09-29)
- Counterpoint 估算 2026 上半年各场景占出货：服务与导览约 19%，智能制造约 13%，仓储物流约 5%。TrendForce 指出市场距离大规模标准化重复采购仍有差距。（Counterpoint，2026-08-20；TrendForce，2026-08-19）Counterpoint estimates H1 2026 shipments by use at about 19% for services and guided tours, about 13% for smart manufacturing and about 5% for warehousing and logistics. TrendForce notes the market is still some way from large-scale standardised repeat purchasing. (Counterpoint, 2026-08-20; TrendForce, 2026-08-19)
- 买家与运营方一侧：酒店经理自己算出两三年回本，商场负责人因性价比没有采购。这两条都是非人形机器人，商场一条还是 2020 年的发言，只能说明买家怎么算账，不能当作人形的采购证据。On the buyer and operator side, a hotel manager worked out payback in two to three years, and a mall manager declined to buy on value grounds. Both concern non-humanoid robots, and the mall remark dates from 2020. They show how buyers do the sums, not that anyone bought a humanoid.
- 家庭：公众评论只有有条件的购买意愿，未见已购买者；照护场所运营方的发言也未见。证据没有指向家庭。Home: public comments show only conditional willingness to buy, with no actual buyers seen, and no remarks from care operators were found. The evidence does not point to the home.
推断Inference推断Inference先做运营方能自己核算岗位的有人公共场所，工厂并行。长隆之外，商场和酒店买家没有人形机器人的发言，所以商场与酒店跟进只是假设，要靠试点证明。Start where the operator can judge each post itself, and run factories in parallel. Beyond Chimelong, no mall or hotel buyer has been heard on humanoids, so following with malls and hotels is a hypothesis for pilots to test.
依据编号（点击跳到附录 A5）Basis IDs (click to jump to appendix A5)F-33F-97F-03F-11T-02N-B02N-B03N-B04N-B05N-B07N-B11N-B12N-B13N-B14N-F01N-F16V1-2V1-3V1-5
来源：智元披露，公司口径；日期与链接见附录证据。按 20 台机器人示意，不是现场人数。Source: AgiBot’s own statement; dates and links in the evidence appendix. Drawn for 20 robots, not a headcount.
决策二：品牌对谁讲、讲什么Decision 2: who the brand speaks to, and what it says
对场所运营方讲四件事：已经在岗、算得清账、坏了有人修、记录可核。Speak to venue operators about four things: already in post, payback you can compute, someone fixes it when it breaks, a record that can be checked.
置信度：低Confidence: Low（买家排序来自二手小样本；决策本身是推断）(the buyer ranking is a small secondhand sample; the decision itself is an inference)
依据：本页验证部分的顾虑排序（买家与运营方 22 条，二手发言）；主张与顾虑的对应是推断。Basis: the concern ranking in the validation section of this page (22 buyer and operator items, secondhand). Pairing claims with concerns is an inference.
细节：事实、推断与依据Details: facts, inference and basis
条数来自二手发言的编码，不是调查比例。可靠性、速度各有 3 条，没有单独成为主张。Counts come from coding secondhand remarks and are not survey shares. Reliability and speed have 3 items each and do not become claims of their own.
- 展演仍是中国社媒的主要曝光方式；评论里有真实性怀疑和安全联想，证明点正从动作转向运行记录，但这个总判断证据偏薄。（Allianz GI，2026-09-02；The Robot Report，2026-10-26）Demos remain the main exposure on Chinese social media; comments voice doubts about authenticity and fears of injury. Proof is moving from stunts to operating records, but the overall claim is thinly supported. (Allianz GI, 2026-09-02; The Robot Report, 2026-10-26)
- 买家与运营方更常谈回本、新鲜感与使用率、维保；谈安全与责任的只有 3 条，不是排在前面的顾虑。Buyers and operators talk most about payback, novelty and utilisation, and maintenance. Only 3 items concern safety and liability, so it is not a leading concern.
- 买家与运营方没有人谈品牌区分，也没有人把行为层承诺当作采购条件。这一空位仍在，但只能当加分项。No buyer or operator discussed brand differences, and none treated behaviour-layer promises as a buying condition. The gap is still there, but it can only be a bonus.
- 智元公开标价只见部分型号（如远征 A2 青春版 198,000 元），旗舰款未标价；宇树 G1 官网标价 8.5 万元起，并写明保修。条款透明这一项，智元落后于宇树。（宇树 G1 官网，2026-10-28）AgiBot publicly lists prices for only some models (such as the Expedition A2 youth edition, CNY 198,000), not the flagships. Unitree lists the G1 from CNY 85,000 and states the warranty. On transparent terms AgiBot trails Unitree. (Unitree G1 page, 2026-10-28)
- 智元的公司口径部署案例集中在商场、酒店、地铁、导诊和工厂，客户名称与合同条款多未披露。（投资界，智元创新日，2026-10-29）AgiBot’s company-stated deployments sit mostly in malls, hotels, metro stations, hospital guidance and factories, with customer names and contract terms largely undisclosed. (Pedaily, AgiBot Innovation Day, 2026-10-29)
推断Inference推断Inference把在岗运行记录做成品牌的主要证据，并附买家可以复算的回本账。安全与责任降为支撑：写清标准、测试范围和责任条款，但不当作第一主张。不跟宇树比声量，不跟优必选比场景跨度。Make in-post operating records the brand’s main proof, with a payback sum a buyer can recompute. Treat safety and liability as support: state the standard, the test scope and the liability terms, but do not lead with them. Do not compete with Unitree on noise or with UBTech on breadth of settings.
依据编号（点击跳到附录 A5）Basis IDs (click to jump to appendix A5)T-04F-29F-32F-99N-B02N-B03N-B06N-B10N-B14N-B16N-F14N-F16N-B01N-B07N-B11N-B18N-B20N-B22V2-4V2-5V2-6V2-7
决策三：靠什么赚钱Decision 3: how to earn
试点期用租赁或运营服务换取买家的回本账，出现重复采购后再卖整机或签长期服务；按运行小时收费只是设想。Use rental or an operating service for pilots to earn the buyer’s payback sums, and sell whole units or sign long service contracts only after repeat purchases appear; per-hour charging is only an idea.
置信度：低Confidence: Low（没有买家说过愿意怎么付钱；收入没有拆分；决策本身是推断）(no buyer has said how they would pay; revenue is not split; the decision itself is an inference)
依据：智元对收费方式的表述（公司口径，2026-10-29）、租赁价格、买家顾虑排序；顺序是推断。Basis: AgiBot’s remarks on how it might charge (company statements, 2026-10-29), rental prices, and the ranking of buyer concerns. The order is an inference.
细节：事实、推断与依据Details: facts, inference and basis
虚线框是设想或弱线索：按运行小时收费还没有价格和客户；保险与厂商存续各只有一条二手发言，只能当线索去核对。Dashed boxes are an idea or weak leads. Per-hour charging has no price and no customer yet. Insurance and supplier survival each rest on a single secondhand item and are only leads to check.
- 智元高管在创新日采访中，把按工时消耗扣费的套餐称为设想，并说当前刚刚起步；另一位高管称可按客户需求卖本体，也可提供平台或服务。没有价格，没有客户（公司口径）。（潮新闻，智元创新日，2026-10-29；新浪财经，智元创新日，2026-10-29）In Innovation Day interviews, AgiBot executives called a package billed by hours consumed an idea and said it was only just starting; another executive said it could sell units or offer a platform or service as the customer needs. There is no price and no customer (company statements). (Tidenews, AgiBot Innovation Day, 2026-10-29; Sina Finance, AgiBot Innovation Day, 2026-10-29)
- 租赁标价在下降：2025 年初日租最高超 2 万元，2026 年 10 月部分商家 300 元起。租赁价格内卷是租赁商自述的风险，对智元品牌溢价的影响未见。（新浪财经，2026-10-09）Rental prices are falling: daily rates peaked above CNY 20,000 in early 2025 and start from CNY 300 at some merchants in October 2026. Price competition is a risk rental firms report themselves; its effect on AgiBot’s brand premium has not been seen. (Sina Finance, 2026-10-09)
- 买家与运营方没有人明说先租试试，也没有人谈到按小时或按产出付费。有试用和租用的事实（工厂试用、活动租用），但多数不是人形，或是租赁商而非终端买家。No buyer or operator said they would rent to try first, and none mentioned paying by the hour or by output. Trials and rentals did happen (factory trials, event hire), but mostly with non-humanoids, or with rental firms rather than end buyers.
- 回本和新鲜感是最常见的两类顾虑：回本在 22 条里有 10 条，在 2020 年之后的 18 条里有 8 条，只看场景方的 14 条里有 6 条。Payback and novelty are the two most common concerns: payback has 10 of 22 items, 8 of the 18 items from 2020 onward, and 6 of the 14 items from venue-side speakers only.
- 智元的营收没有拆分：公司口述 2026 上半年收入近 20 亿元，未审计，无法判断整机、租赁和服务各占多少。AgiBot’s revenue is not split: the company said revenue for H1 2026 was close to CNY 2 billion, unaudited, so the shares of whole units, rental and services cannot be told.
推断Inference推断Inference收费顺序：试点期租赁或运营服务，转常态后整机或长期服务合同。试点合同带回本账、维保响应条款和转常态的留存判据。按运行小时收费只是设想，核对前不对外承诺。Order of charging: rental or operating service during pilots, then whole units or long service contracts in steady use. The pilot contract carries the payback sums, service response terms and the test for moving to steady use. Per-hour charging is only an idea; do not promise it before it is checked.
依据编号（点击跳到附录 A5）Basis IDs (click to jump to appendix A5)F-98F-96F-94F-16F-11T-01T-02N-B02N-B03N-B07N-B12N-B13N-B14N-B15N-B19N-B17V3-1V3-2V3-3V3-4V3-5V3-6V3-7
用买家的话验证Checking against buyers’ remarks
基于公开讨论的二手验证，买家样本有限Secondhand validation from public discussion; the buyer sample is limited.
- 原来的假设The hypothesis
- 运营方最担心安全和责任，而不是价格Operators fear safety and liability more than price
- 公开发言里In public remarks
- 回本 10 条、新鲜感 9 条，安全与责任 3 条（共 22 条）Payback 10, novelty 9, safety and liability 3 (of 22)
- 决策因此改为So the decision became
- 在岗记录和回本账当主要证据，安全与责任放在支撑位置In-post records and a payback sum as the main proof, with safety and liability in support
20 个验证假设：支持 2 · 部分支持 13 · 未见 4 · 相反 120 hypotheses: supported 2 · partly 13 · not found 4 · contradicted 1
回本和新鲜感最常被提到Payback and novelty come up most
推断Inference推断Inference买家先问的是算不算得过账，不是敢不敢；所以先让他们算得清账、看得到在岗记录。Buyers first ask whether the sums work, not whether they dare; so first make the payback clear and the in-post record visible.
依据：本研究的公开讨论抽样（78 条，其中买家与运营方 22 条）；二手发言，买家样本有限。Basis: this study’s sample of public discussion (78 items, 22 from buyers and operators). Secondhand remarks; the buyer sample is limited.
细节：样本、敏感性检查、假设结果与弱线索Details: sample, sensitivity check, hypothesis results and weak leads
图中前四项是决策二的依据。条数来自对公开发言的编码，没有复核，不是调查比例。The top four feed Decision 2. Counts come from coding public remarks, with no review, and are not survey shares.
- 样本：78 条发言，来自 29 个页面。其中买家与运营方 22 条（买家 10、运营 12），从业者与公众各自单列。Sample: 78 items from 29 pages. Of these, 22 come from buyers and operators (10 buyers and 12 operators); practitioners and the public are counted separately.
- 22 条里有 16 条是媒体或会议报道转述的二手说法，6 条是 X 或 HN 上的自述（无法核验，多为英文）。至少 10 条谈的是非人形机器人，8 条来自租赁商，4 条是 2020 年的发言。Of the 22, 16 are secondhand accounts relayed by media or conference reports and 6 are self-descriptions on X or HN (unverifiable, mostly in English). At least 10 concern non-humanoid robots, 8 come from rental firms and 4 date from 2020.
- 敏感性检查：回本在全部 22 条里 10 条，新鲜感 9 条；回本在 2020 年之后的 18 条里 8 条，只看场景方的 14 条里 6 条。样本小，回本只领先一两条，所以只说这两类最常见，不排第一第二。Sensitivity check: payback has 10 of all 22 items and novelty 9; payback has 8 of the 18 from 2020 onward and 6 of the 14 from venue-side speakers only. The sample is small and payback leads by only one or two items, so these two are called the most common, not ranked.
- 可靠性在从业者的 21 条里有 11 条，买家与运营方里只有 3 条。可能是厂商更爱强调它，也可能只是买家样本小，这里不下结论。Reliability appears in 11 of 21 practitioner items but in only 3 buyer and operator items. Vendors may stress it more, or the buyer sample may simply be small. No conclusion is drawn here.
对照假设的结果How the hypotheses came out
- 20 个验证假设：支持 2，部分支持 13，未见 4，相反 1。Of 20 validation hypotheses, 2 are supported, 13 partly supported, 4 not seen and 1 contradicted.
- 相反：假设运营方最担心安全和责任而不是价格。样本里反过来：回本和新鲜感排前面，安全与责任只有 3 条。所以安全与责任不再当第一主张。Contradicted: the guess that operators fear safety and liability more than price. The sample shows the reverse: payback and novelty rank first, and safety and liability has only 3 items. So safety and liability are no longer the lead claim.
- 支持：回本是买家的核心算法；维保、故障和停机是真实痛点。但没有买家说愿意为服务合同付费。Supported: payback is the buyer’s core calculation, and maintenance, faults and downtime are real pain points. But no buyer said they would pay for a service contract.
- 未见：长隆的采购方式与价格，买家如何区分品牌，行为层承诺是否被需要，按小时或按产出付费。未见不等于没有，只是本次没读到。Not seen: how Chimelong buys and at what price, how buyers tell brands apart, whether behaviour-layer promises are wanted, and per-hour or per-output pricing. Not seen does not mean absent; it means not found this time.
两条弱线索，只作核对用Two weak leads, to be checked only
- 弱线索 · 保险：一位匿名的 HN 用户自述向保险公司咨询时拿不到保险，并说高校要求投保一般责任险。这是美国的单条匿名发言，不能写成保险公司拒保，也不能推到国内。Weak lead, insurance: one anonymous HN user said they could not get cover when asking insurers, and that a university required general liability insurance. It is a single anonymous US remark; do not write that insurers refuse cover, and do not carry it over to China.
- 弱线索 · 厂商存续：一位具名运营者复盘 2020 年的非人形沙拉机器人生意，称若继续经营，将面临厂商被收购后买不到机器人、没有技术支持。单条、非人形、年代久远。Weak lead, supplier survival: a named operator, looking back on a 2020 non-humanoid salad-robot business, said that had they carried on they would have faced a supplier that was acquired, with no robots to buy and no support. One item, non-humanoid, and old.
依据编号（点击跳到附录 A5）Basis IDs (click to jump to appendix A5)N-B19N-B17V3-6V3-7
推断Inference推断Inference安全与责任写清楚，但放在支撑位置。Spell out safety and liability, but as support.
依据编号（点击跳到附录 A5）Basis IDs (click to jump to appendix A5)V1-1V1-2V2-4V2-5V3-3V3-4V3-5N-B01N-B02N-B03N-B04N-B05N-B07N-B10N-B16
给智元的建议Recommendations for AgiBot
以下四条都是推断，不是研究结论。All four lines below are inferences, not research conclusions.
先把一个有人公共场所做成灯塔，工厂并行当第二个证明场。Turn one people-filled public venue into the lighthouse, with factories as a second proving ground.
灯塔的判据是经运营考核后留下来的岗位，不是展演热度或台数。The test is posts that stay after operator review, not demo buzz or unit counts.
把在岗运行记录当成品牌的主要证据，再附一份买家可以复算的回本账。Make in-post operating records the brand’s main proof, and add a payback sum a buyer can recompute.
记录写明任务、小时数、是否遥操作、失败与恢复，数字有出处。回本账含替代了多少人力、现场仍需多少人、停机与维护成本。Records state the task, hours, whether teleoperated, failures and recovery, each number sourced. The payback sum covers how much labour is replaced, how many people are still needed on site, and downtime and maintenance cost.
售后先行：公开响应时间、本地维修和备件。Put after-sales first: publish response times, local repair and spare parts.
补上更多型号的标价、保修和故障处理条款，缩小与宇树在条款透明上的差距。Add prices, warranty and fault handling for more models to close the transparency gap with Unitree.
收费顺序：试点期租赁或运营服务，转常态后整机或长期服务；按运行小时收费只是设想，核对前不对外承诺。Charging order: rental or operating service for pilots, then whole units or long service contracts; per-hour charging is only an idea, not to be promised before it is checked.
试点合同带回本账、维保条款和转常态的留存判据。The pilot contract carries the payback sums, service terms and the test for moving to steady use.
加分项：先示意，再动作A bonus: signal, then move
买家没有把行为层承诺当作采购条件，所以它只当加分项。动作前的示意能否帮助旁观者理解机器人意图、判断安全边界，不是本次桌面研究的发现，没有做过一手访谈与实地测试。Buyers did not treat behavior-layer commitments as a purchase condition, so this stays a bonus. Whether a signal before movement helps bystanders understand robot intent and judge safety boundaries is not a finding of this desk research, and it has not been through first-hand interviews or field testing.
细节：每条建议的依据Details: the basis for each
局限Limits
- 买家声音少，且多为二手。买家与运营方 22 条里，16 条是媒体或会议报道转述，6 条是 X 或 HN 自述，无法核验；至少 10 条谈的是非人形机器人。2026 年人形机器人的买家发言，只有工厂方和租赁商。Buyer voices are few and mostly secondhand. Of the 22 buyer and operator items, 16 are relayed by media or conference reports and 6 are unverifiable self-descriptions on X or HN; at least 10 concern non-humanoids. The only 2026 humanoid buyer remarks come from a factory side and from rental firms.
- 智元的部署证据几乎都是公司口径，客户没有确认。长隆超过 300 台（9 月 24 日）与近 500 台（创新日）都是公司口径，近 500 台为单一来源；两处口径不完全相同，是否同一园区未核，不能互相印证。AgiBot’s deployment evidence is almost all company statements, with no customer confirmation. Chimelong’s over 300 units (24 September) and close to 500 (Innovation Day) are both company statements, and the close-to-500 count is a single source. The two statements are not worded identically and whether they describe the same park was not checked, so they cannot confirm each other.
- 长隆等案例的商业条款（买断、租赁、服务、价格、续约）未知，营收没有拆分，所以三个决策里的收费顺序是推断。租赁价格与营收数字都是预测或公司口述。The commercial terms of Chimelong and similar cases (purchase, rental, service, price, renewal) are unknown and revenue is not split, so the charging order in Decision 3 is an inference. Rental figures and revenue are forecasts or company statements.
- 样本没有覆盖知乎、B 站、抖音、脉脉、Reddit、领英；小红书只有上文的评论抽样。公众评论是按点赞排序的首屏小样本，不代表全体用户。The sample does not cover Zhihu, Bilibili, Douyin, Maimai, Reddit or LinkedIn; Xiaohongshu appears only through the comment sample above. Public comments are a small first-screen sample sorted by likes and do not represent all users.
- 评分只是判断，不是盲评，收集证据和评分没有分开。Ratings are judgments, not blind, and evidence gathering and scoring were not kept separate.
- 行为层得分只代表本次核对的公开页面，未见不等于没有。Behaviour-layer scores cover only the public pages checked this time. Not seen does not mean absent.
- 标准与法规：国际标准 ISO 25785-1 仍是委员会草案；欧盟 (EU) 2026/1744 转引自英国政府备忘录，欧盟原文未见。尚无已实施的人形专用标准，范围未核。Standards and regulation: ISO 25785-1 is still a committee draft, and (EU) 2026/1744 is quoted from a UK government memorandum, with the EU original not seen. No implemented humanoid-specific standard was found, and its scope was not checked.
- 保险与厂商存续两条线索各只有一条发言，其中一条是美国的匿名发言，另一条是 2020 年的非人形，国内情况未核。The insurance and supplier-survival leads each rest on one remark: one anonymous and American, the other from 2020 and non-humanoid. The situation in China was not checked.
- 决策和建议是推断，不是研究结论：没有客户访谈，也没有一手的买家数据。顾虑编码没有经过复核。The decisions and recommendations are inferences, not research conclusions: there are no client interviews and no first-hand buyer data. The concern coding was not reviewed.
证据、评分和方法Evidence, ratings and methods
查看全部证据View all evidence
A1 证据A1 Evidence
| 来源Source | 日期Date | 链接Link |
|---|---|---|
| Counterpoint | 2026-08-20 | 查看来源View source |
| SAG | 2026-08-10 | 查看来源View source |
| IDC，经澎湃IDC, via The Paper | 2026-10-01 | 查看来源View source |
| 宇树 X 官方账号Unitree on X | 2026-10-28 读取Read 2026-10-28 | 查看来源View source |
| 宇树小红书评论Unitree comment sample | 2026-10-28 读取Read 2026-10-28 | — |
| 宇树小红书评论Unitree comment sample | 2026-10-28 读取Read 2026-10-28 | — |
| 智元小红书评论AgiBot comment sample | 2026-10-28 读取Read 2026-10-28 | — |
| 优必选 U1 预售帖UBTech U1 pre-order | 2026-10-28 读取Read 2026-10-28 | — |
| 宇树 G1 官网Unitree G1 | 2026-10-28 读取Read 2026-10-28 | 查看来源View source |
| 智元 X2 官网AgiBot X2 | 2026-10-28 读取Read 2026-10-28 | 查看来源View source |
| 傅利叶 GR-3，TechWalkerFourier GR-3, TechWalker | 2025-08-06 | 查看来源View source |
| 傅利叶 CES 新闻稿Fourier CES release | 2026-01-07 | 查看来源View source |
| IFR《Global Sales of Professional Service Robots Surge 24%》，World Robotics 2026 | 2026-09（页面未写具体日）2026-09 (no day on the page) | 查看来源View source |
| IDC 2025 年数据，经央视IDC 2025 data via CCTV | 2026-01-23 至 2026-01-252026-01-23 to 2026-01-25 | 查看来源View source |
| 高盛文章Goldman Sachs article | 2024-02-27 | 查看来源View source |
| 高盛报告 PDFGoldman Sachs report PDF | 2024-01-08 | 查看来源View source |
| 租赁价格，新浪财经Rental prices, Sina Finance | 2026-10-09 | 查看来源View source |
| The Robot Report，Agility 与 GXOThe Robot Report, Agility and GXO | 2026-10-26 | 查看来源View source |
| 优必选 U1，IT之家UBTech U1, ITHome | 2026-06-30 | 查看来源View source |
| Allianz GI《When robots start earning their keep》 | 2026-09-02 | 查看来源View source |
| ISO/CD 25785-1 页面ISO/CD 25785-1 page | 2026-05-08（登记 CD）2026-05-08 (CD registered) | 查看来源View source |
| 同花顺财经，首批国家标准要点Tonghuashun Finance, first national standards | 2026-10-23 | 查看来源View source |
| 工信部科技司，标准体系建设指南MIIT standards-system guide | 2026-08-24 | 查看来源View source |
| 工信部征求稿 PDFMIIT draft PDF | 2026-08-24 | 查看来源View source |
| 新华社，标准体系建设指南Xinhua, standards-system guide | 2026-08-24 | 查看来源View source |
| 中国经济网，首批 6 项国家标准China Economic Net, first 6 national standards | 2026-10-16 | 查看来源View source |
| 欧盟委员会 Machinery 页面European Commission Machinery page | 2026-10-30 读取Read 2026-10-30 | 查看来源View source |
| 英国政府说明备忘录，(EU) 2026/1744 的转引来源，欧盟原文未见UK government explanatory memorandum, the quoted source for (EU) 2026/1744; EU original not seen | 2026-08-03 | 查看来源View source |
| 投资界，智元创新日报道Pedaily, AgiBot Innovation Day report | 2026-10-29 | 查看来源View source |
| 潮新闻，智元创新日采访Tidenews, AgiBot Innovation Day interview | 2026-10-29 | 查看来源View source |
| 新浪财经，智元创新日Sina Finance, AgiBot Innovation Day | 2026-10-29 | 查看来源View source |
| IT之家，智元第 20,000 台下线ITHome, AgiBot 20,000th unit | 2026-09-24 | 查看来源View source |
| 搜狐转载，智元披露长隆人员配置Sohu, AgiBot disclosure on Chimelong staffing | 2026-09-29 | 查看来源View source |
| TrendForce，中国人形机器人产值预测TrendForce, China humanoid output forecast | 2026-08-19 | 查看来源View source |
以下证据来自 2026 年 10 月的桌面研究，没有对原始页面重新核对。出货区间：SAG 和 Counterpoint 的估算，日期为 2026-08-10 和 2026-08-20，见两家机构的页面。科研、展演和数据采集占出货量：IDC，经澎湃新闻 2026-10-01 转述。X 展示量：官方账号页面，2026-10-28 读取。评论：小红书高赞笔记的首屏，2026-10-28 读取。条款与免责：宇树官网 G1 页，2026-10-28 读取。The evidence below comes from desk research in October 2026, and the original pages were not re-checked. Shipment ranges: SAG and Counterpoint estimates dated 10 and 20 August 2026, from the two agencies’ pages. Research, shows and data collection as a share of shipments: IDC, as reported by The Paper on 1 October 2026. X impressions: the official account page, read on 28 October 2026. Comments: the first screen of top-liked Xiaohongshu posts, read on 28 October 2026. Terms and disclaimer: Unitree's G1 page, read on 28 October 2026.
A2 评分A2 Ratings
评分图Ratings chart
来源：对公开页面的评分；依据与方法见下文“方法与局限”。Source: ratings of public pages; evidence and methods in “Methods and limits” below.
| 对象Product | 文字层Text | 行为层Behavior |
|---|---|---|
| 宇树Unitree | 3 | 1 · 保持安全距离1 · Distance disclaimer |
| 优必选 U1UBTech U1 | 2 | N/E |
| 智元AgiBot | 1 | N/E |
| 傅利叶Fourier | 1 | N/E |
| 优必选 Walker S2UBTech Walker S2 | 1 | N/E |
文字层：宇树 3，优必选 U1 2，智元 1，傅利叶 1，优必选 Walker S2 1。行为层：宇树 1，其余 N/E。N/E 表示证据不足，不是 0 分。两个层的分数不换算成坐标。Text layer: Unitree 3, UBTech U1 2, AgiBot 1, Fourier 1, UBTech Walker S2 1. Behavior layer: Unitree 1, all others N/E. N/E means there was not enough evidence, and it is not a zero. The two layers are not converted into coordinates.
评分经一轮锚点澄清，未经访谈校准。第二轮不是盲评。The anchors were clarified once, the scores have not been calibrated with interviews, and the second round was not blind.
A3 方法与局限A3 Methods and limits
评论取按点赞排序的首屏，每个帖子约 10 到 15 条，置信度低，不给比例。出货区间不是置信区间。时间长度不同的数据只看方向。Comments come from the first screen sorted by likes, about 10 to 15 per post. Confidence is low and no proportions are given. Shipment ranges are not confidence intervals. Data covering periods of different lengths shows direction only.
没有放进正文的线索：55.5% 按项目数统计，不能用来印证 69%。智元自报的"超过 8,600 台"没有并入区间。其余低置信度推测也没有放入。Leads outside the main text: 55.5% counts projects and cannot corroborate 69%. AgiBot's own report of "over 8,600" units is excluded from the range. Other low-confidence inferences are also left out.
来源：人形机器人场景应用联盟统计的媒体转述（转载，联盟原报告未打开）；IDC 经澎湃；智元公司自述经第一财经、澎湃转述。Sources: media reports of the humanoid robot application alliance statistics (reprint; original report not opened); IDC via The Paper; AgiBot company statement via Yicai and The Paper.
据第三方估算，2026 上半年智元出货约 8,400–9,700 台，宇树约 5,900–7,000+ 台，按出货量智元排在宇树之前。宇树官方号 10 月 21 日在 X 上发的 H2-X 帖，展示量约 1,612 万（页面读取值）。在小红书上按点赞排序的前 8 篇宇树笔记里，有 7 篇来自官方账号。另一方面，科研教育、性能展示和政府数据采集这类科研、展演和数据采集占出货量的比例，从 83.8%（2025 全年）降到 69%（2026 上半年）。这个比例在下降，但仍然是大头。By third-party estimates, AgiBot shipped about 8,400–9,700 units in H1 2026 and Unitree about 5,900–7,000+, putting AgiBot ahead on shipments. Unitree's 21 October post on X about H2-X drew about 16.12 million impressions (as read from the page). Of the 8 most-liked Unitree posts on Xiaohongshu in the search, 7 came from the official account. Meanwhile the share of shipments going to research, education, performance and government data-collection uses fell from 83.8% (all of 2025) to 69% (H1 2026). The share is falling, but it is still the largest.
第三方估算；区间的两端来自两家不同机构（SAG 与 Counterpoint），不是置信区间。展示量是页面读取值。Third-party estimates; the two ends of each range come from two different agencies (SAG and Counterpoint) and are not confidence intervals. Impressions are as read from the page.
评论只取按点赞排序的首屏，样本小，置信度低，反映的是话题，不是比例。隐私和联网的顾虑来自优必选 U1 预售帖。Only the first screen of comments sorted by likes was taken. The sample is small and confidence is low, so it shows topics, not proportions. The privacy and networking concerns came from the UBTech U1 pre-order post.
本研究将"承诺"分成两层。一层是写在条款里的，比如价格、保修和数据政策。另一层是现场看得见的，比如灯、动作、声音，以及出故障时的行为。评分经一轮锚点澄清，未经访谈校准。行为层只有宇树有分（1 分），依据是官网上"保持安全距离"的免责文案，其余四个是 N/E。This research separates "commitment" into two layers. One is written in the terms, such as price, warranty and data policy. The other is visible in the moment, such as light, motion, sound and what the robot does when something fails. The anchors were clarified once, and the scores have not been calibrated with interviews. On the behavior layer only Unitree scores (1), based on the "keep a safe distance" disclaimer on its website. The other four are N/E.
行为层只有宇树有分（1），内容仅是官网“保持安全距离”的免责文案；其余四个是 N/E。On the behavior layer only Unitree scores, a 1, for a "keep a safe distance" disclaimer on its website; the other four are N/E.
N/E = 证据不足，不等于 0；评分只是判断，不是结论。第二轮不是盲评。N/E means insufficient evidence, not zero. Judgments, not conclusions. The second round was not blind.
这是桌面研究，用的是公开资料、五个对象的公开页面，以及社交平台上高点赞评论的抽样。五个对象是宇树、智元、傅利叶，和优必选的 Walker S2 与 U1。本研究未开展一手访谈。This is desk research. It uses public sources, the public pages of five products, and a sample of highly liked comments on social platforms. The five are Unitree, AgiBot, Fourier, and UBTech's Walker S2 and U1. This research includes no first-hand interviews.
五个对象：宇树、智元、傅利叶、优必选 Walker S2 与 U1。X 帖展示量 1,612 万；小红书前 8 篇中 7 篇为官方。出货区间两端来自 SAG 与 Counterpoint。智元 X2 灯效与交互屏已有，缺统一语义。Five products: Unitree, AgiBot, Fourier, UBTech Walker S2 and U1. 16.12 million X impressions; seven of eight Xiaohongshu posts official. SAG and Counterpoint supply shipment endpoints. AgiBot X2 has lighting and a screen, but lacks unified semantics.
A4 证据编号A4 Evidence IDs
决策、验证和建议部分的依据编号：F 与 H 是发现，T 是趋势，N 开头的是公开发言记录（N-B 为买家与运营方，N-F 为从业者，N-G 为公众），V 是验证假设。编号在页面上是链接，点击跳到下面的 A5 对应行。The basis IDs in the decision, validation and recommendation sections: F and H are findings, T are trends, those starting with N are records of public remarks (N-B buyers and operators, N-F practitioners, N-G the public), and V are validation hypotheses. On the page each ID is a link to its row in A5 below.
A5 依据编号对照A5 Evidence ID index
页面上每个依据编号都链接到这里的一行。N 开头的是买家、从业者和公众发言记录（只列日期、身份、顾虑主题和来源链接，不摘录原话），V 是验证假设（没有外部链接，依据见同一行里的结果）。买家与运营方发言列出被引用的 19 条，另外 3 条没有被引用，图中计数包含它们。内容是研究记录的中文概括。Each basis ID on the page links to one row here. N entries are the records of buyer, practitioner and public remarks (date, role, concern themes and source link only, no quotations); V entries are the validation hypotheses (no external link; the result is in the row). Of the buyer and operator remarks, the 19 that are cited are listed; the other 3 are not cited, and the counts in the figure include them. Descriptions are summaries of the research records, in Chinese.
| 编号ID | 内容Entry | 来源Source |
|---|---|---|
| F-03 发现Finding | Counterpoint 口径：2026上半年全球人形机器人应用中，“娱乐表演”与“数据生产/科研”合计占比在H1略有下降（原文未写明对比基期），但仍高于60%；服务与导览约19%，智能制造约13%，仓储物流约5%。 Chinese-language record; see the source link. | 来源 1Source 1 |
| F-11 发现Finding | TrendForce：2026年中国人形机器人产值约150亿元，2027年增速不低于60%；同时指出市场“距离大规模标准化重复采购仍有差距”，需验证复购与投资回报。 Chinese-language record; see the source link. | 来源 1Source 1 |
| F-16 发现Finding | 租赁价格（媒体转述，已打开原文，2026-10-09）：租赁价格区间大幅分化、整体下行。（1）历史：2025年初春晚后，宇树G1在二手/租赁店铺日租标价多在8,000–15,000元，热门档期最高超2万元； Chinese-language record; see the source link. | 来源 1Source 1 |
| F-29 发现Finding | 在上述笔记的高互动评论中，还出现对真实性/自主性的怀疑：评论原话不在此摘录。 Chinese-language record. | — |
| F-32 发现Finding | 智元官方商店页检索摘要：灵犀X2青春焕新版155,000元、远征A2青春版198,000元；旗舰款与精灵G2未公开标价。 Chinese-language record; see the source link. | 来源 1Source 1 |
| F-33 发现Finding | 智元宣布累计下线第20,000台机器人（2026-09-24，远征A3 Ultra），该台交付横琴长隆飞船乐园，园区常态化上岗机器人超300台。 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| F-94 发现Finding | 2026智元创新日（2026-10-29，上海世博中心）发布内容（公司口径，AGIBOT英文新闻稿）：物理AI价值度量单位（有效智能工时＝在岗机器人数×自主作业时长×产出效能比）；第三代人形硬件架构AHA3；皮下式关节液冷系统与IP67关节；灵巧手OmniHand 3 Lite/3/3 Ultra与夹爪OmniPicker 3； Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| F-95 发现Finding | 精灵G2 Max（公司口径，创新日发布）：定位安规级重载工业机器人，21自由度，单臂负载20公斤、双臂合计40公斤、峰值50公斤，7自由度力控双臂，作业高度约0.1–2.5米，双电池热插拔；公司称已在食品、日化快消、电商企业落地工业搬运。未公布价格。 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| F-96 发现Finding | 智元经营数字（公司口径，联合创始人、总裁兼CTO彭志辉在创新日现场披露，多家媒体同日报道）：2026年上半年出货9,700台；上半年整体收入接近20亿元；9月下线第20,000台机器人；占全球人形机器人43%市场份额；6个月内完成量产1万台到2万台。 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| F-97 发现Finding | 智元在创新日列出的部署案例（公司口径）：长隆具身智能主题乐园近500台机器人常态化部署，计划年底增至1,000台、覆盖100多个岗位；广州地铁120台精灵G2承担安检、咨询、巡逻；精灵G2在龙旗工厂平板量产质检工段全覆盖；芯片封测上下料、CNC上下料； Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| F-98 发现Finding | 智元关于工时度量单位与收费方式的表述（采访，公司口径）：总裁助理/合伙人王闯称第一版该度量单位只做了粗略定义，设想客户购买“机器人生产力套餐”、按消耗的该度量单位扣除，“这是我们的设想，当前我们还是刚刚起步”；彭志辉称可按客户需求卖本体，也可提供平台或服务。 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 |
| F-99 发现Finding | 智元商城2025-08-18首发预售价（三家媒体当日报道）：远征A2青春版定价198,000元（限时优惠168,000元）；灵犀X2青春版定价98,000元；精灵G1售价45万元；远征A2旗舰版、灵犀X2探索版/旗舰版未标价，需联系购买。 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| T-01 趋势Trend | 出货放量、价格下探，但“出货”的口径差异很大 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| T-02 趋势Trend | 落地场景：从展演/科研向工厂、仓储、服务慢慢转移，但“规模化重复采购”尚未出现，家庭最远 Chinese-language record; see the source link. | 来源 1Source 1 |
| T-03 趋势Trend | 安全标准与法规：2027年前后进入“有规则可依”阶段，但服务/家庭场景的人形专用标准仍空白 Chinese-language record; see the source link. | 来源 1Source 1 · 来源 2Source 2 · 来源 3Source 3 |
| T-04 趋势Trend | 品牌叙事：证明点从“炫技视频”转向“运行小时数、完成量、复购”，但中国社媒上头部内容仍以展演为主 Chinese-language record; see the source link. | 来源 1Source 1 |
| N-B01 买家与运营方发言Buyer or operator item | 2025-04-29 · 运营·场景方 · 媒体转述 · 顾虑主题：总成本与回本、功能范围窄 2025-04-29 · Operator, venue side · media report · concerns: cost and payback, narrow function | 来源 1Source 1 |
| N-B02 买家与运营方发言Buyer or operator item | 2025-04-26 · 运营·场景方 · 媒体转述 · 顾虑主题：总成本与回本 2025-04-26 · Operator, venue side · media report · concerns: cost and payback | 来源 1Source 1 |
| N-B03 买家与运营方发言Buyer or operator item | 2020-01-16 · 买家·场景方 · 媒体转述 · 顾虑主题：总成本与回本、新鲜感/使用率 2020-01-16 · Buyer, venue side · media report · concerns: cost and payback, novelty and utilisation | 来源 1Source 1 |
| N-B04 买家与运营方发言Buyer or operator item | 2020-01-16 · 运营·场景方 · 媒体转述 · 顾虑主题：速度/效率、新鲜感/使用率 2020-01-16 · Operator, venue side · media report · concerns: speed, novelty and utilisation | 来源 1Source 1 |
| N-B05 买家与运营方发言Buyer or operator item | 2020-03-17 · 运营·场景方 · 媒体转述 · 顾虑主题：新鲜感/使用率、总成本与回本 2020-03-17 · Operator, venue side · media report · concerns: novelty and utilisation, cost and payback | 来源 1Source 1 |
| N-B06 买家与运营方发言Buyer or operator item | 2026-04-21 · 运营·场景方 · 媒体转述 · 顾虑主题：维保/备件/售后、可靠性/故障、自主性不足/依赖人工 2026-04-21 · Operator, venue side · media report · concerns: maintenance and after-sales, reliability and faults, autonomy, human back-up | 来源 1Source 1 |
| N-B07 买家与运营方发言Buyer or operator item | 2025-04-24 · 买家·场景方（租用方） · 媒体转述 · 顾虑主题：功能范围窄、新鲜感/使用率 2025-04-24 · Buyer, venue side (renter) · media report · concerns: narrow function, novelty and utilisation | 来源 1Source 1 |
| N-B10 买家与运营方发言Buyer or operator item | 2026-07-10 · 买家·场景方 · 会议展会报道 · 顾虑主题：维保/备件/售后、可靠性/故障、安全与责任、隐私与网络安全 2026-07-10 · Buyer, venue side · conference report · concerns: maintenance and after-sales, reliability and faults, safety and liability, privacy and security | 来源 1Source 1 |
| N-B11 买家与运营方发言Buyer or operator item | 2025-04-30 · 买家·租赁商 · 媒体转述 · 顾虑主题：功能范围窄、新鲜感/使用率 2025-04-30 · Buyer, rental firm · media report · concerns: narrow function, novelty and utilisation | 来源 1Source 1 |
| N-B12 买家与运营方发言Buyer or operator item | 2025-04-30 · 买家·租赁商 · 媒体转述 · 顾虑主题：新鲜感/使用率、总成本与回本 2025-04-30 · Buyer, rental firm · media report · concerns: novelty and utilisation, cost and payback | 来源 1Source 1 |
| N-B13 买家与运营方发言Buyer or operator item | 2025-04-30 · 买家·租赁商 · 媒体转述 · 顾虑主题：新鲜感/使用率、总成本与回本 2025-04-30 · Buyer, rental firm · media report · concerns: novelty and utilisation, cost and payback | 来源 1Source 1 |
| N-B14 买家与运营方发言Buyer or operator item | 2025-04-24 · 运营·租赁商 · 媒体转述 · 顾虑主题：维保/备件/售后、新鲜感/使用率、总成本与回本 2025-04-24 · Operator, rental firm · media report · concerns: maintenance and after-sales, novelty and utilisation, cost and payback | 来源 1Source 1 |
| N-B15 买家与运营方发言Buyer or operator item | 2025-04-24 · 运营·租赁商 · 媒体转述 · 顾虑主题：总成本与回本 2025-04-24 · Operator, rental firm · media report · concerns: cost and payback | 来源 1Source 1 |
| N-B16 买家与运营方发言Buyer or operator item | 2026-06-05 · 运营·租赁商 · 会议展会报道 · 顾虑主题：维保/备件/售后、安全与责任 2026-06-05 · Operator, rental firm · conference report · concerns: maintenance and after-sales, safety and liability | 来源 1Source 1 |
| N-B17 买家与运营方发言Buyer or operator item | 2024-04-19 · 运营·场景方 · X · 顾虑主题：集成与现场改造、维保/备件/售后、可靠性/故障 2024-04-19 · Operator, venue side · X post · concerns: integration, maintenance and after-sales, reliability and faults | 来源 1Source 1 |
| N-B18 买家与运营方发言Buyer or operator item | 2026-03-19 · 买家·租赁商 · X · 顾虑主题：功能范围窄 2026-03-19 · Buyer, rental firm · X post · concerns: narrow function | 来源 1Source 1 |
| N-B19 买家与运营方发言Buyer or operator item | 2026-09-16 · 买家·租赁商 · HN · 顾虑主题：安全与责任 2026-09-16 · Buyer, rental firm · Hacker News · concerns: safety and liability | 来源 1Source 1 |
| N-B20 买家与运营方发言Buyer or operator item | 2025-03-08 · 运营·场景方 · HN · 顾虑主题：功能范围窄 2025-03-08 · Operator, venue side · Hacker News · concerns: narrow function | 来源 1Source 1 |
| N-B22 买家与运营方发言Buyer or operator item | 2026-03-05 · 运营·场景方 · HN · 顾虑主题：新鲜感/使用率、功能范围窄 2026-03-05 · Operator, venue side · Hacker News · concerns: novelty and utilisation, narrow function | 来源 1Source 1 |
| N-F01 从业者发言Practitioner item | 2026-10-16 · 从业者 · 媒体转述 · 顾虑主题：新鲜感/使用率 2026-10-16 · Practitioner · media report · concerns: novelty and utilisation | 来源 1Source 1 |
| N-F14 从业者发言Practitioner item | 2026-10-23 · 从业者 · 媒体转述 · 顾虑主题：总成本与回本、自主性不足/依赖人工 2026-10-23 · Practitioner · media report · concerns: cost and payback, autonomy, human back-up | 来源 1Source 1 |
| N-F16 从业者发言Practitioner item | 2026-09-29 · 从业者 · 媒体转述 · 顾虑主题：总成本与回本、新鲜感/使用率 2026-09-29 · Practitioner · media report · concerns: cost and payback, novelty and utilisation | 来源 1Source 1 |
| V1-1 验证假设Validation hypothesis | 有人公共场所的运营方最担心安全和责任，而不是价格 Chinese-language record; see the source link. | 无外部链接No external link |
| V1-2 验证假设Validation hypothesis | 长隆类场所真实使用方式（买断/租赁/服务） Chinese-language record; see the source link. | 无外部链接No external link |
| V1-3 验证假设Validation hypothesis | 酒店前台、门店导购、导诊岗位有运营方谈过实际使用效果 Chinese-language record; see the source link. | 无外部链接No external link |
| V1-5 验证假设Validation hypothesis | 家庭与照护：围观者的期待是否伴随真实购买意愿 Chinese-language record; see the source link. | 无外部链接No external link |
| V2-4 验证假设Validation hypothesis | 买家如何看智元、宇树、优必选；品牌区分是否是考虑因素 Chinese-language record; see the source link. | 无外部链接No external link |
| V2-5 验证假设Validation hypothesis | 行为层承诺（减速、让行、停机提示、故障行为）是买家想要的信息 Chinese-language record; see the source link. | 无外部链接No external link |
| V2-6 验证假设Validation hypothesis | 条款透明（标价、保修、故障处理）买家是否在意 Chinese-language record; see the source link. | 无外部链接No external link |
| V2-7 验证假设Validation hypothesis | 买家/运营希望上岗前写清“能做什么、不做什么” Chinese-language record; see the source link. | 无外部链接No external link |
| V3-1 验证假设Validation hypothesis | 买家更愿意先租/试点，再决定买不买 Chinese-language record; see the source link. | 无外部链接No external link |
| V3-2 验证假设Validation hypothesis | 租赁价格下行使买家把机器人当可替换租赁品，而非品牌资产 Chinese-language record; see the source link. | 无外部链接No external link |
| V3-3 验证假设Validation hypothesis | 买家如何理解“按小时/按产出收费” Chinese-language record; see the source link. | 无外部链接No external link |
| V3-4 验证假设Validation hypothesis | 回本逻辑：替代人工、吸引客流还是别的 Chinese-language record; see the source link. | 无外部链接No external link |
| V3-5 验证假设Validation hypothesis | 维护、故障、停机谁负责，是否是买家愿为服务合同付钱的理由 Chinese-language record; see the source link. | 无外部链接No external link |
| V3-6 验证假设Validation hypothesis | 试点合同需写明公众责任由谁承担、是否有保险 Chinese-language record; see the source link. | 无外部链接No external link |
| V3-7 验证假设Validation hypothesis | 厂商被收购/停止支持的风险影响买家是否买断 Chinese-language record; see the source link. | 无外部链接No external link |
A6 其余建议A6 More recommendations
正文列出四条核心建议，其余四条放在这里。都是推断。The page lists four core recommendations; the other four are here. All are inferences.
- 推断Inference推断Inference
上岗前，每个岗位写清能做什么、不做什么、哪些要人在旁。功能范围窄是买家与运营方的第三大顾虑。Before a post goes live, write down what the robot does, what it does not do, and what needs a person beside it. Narrow function is the third most common concern among buyers and operators.
- 推断Inference推断Inference
行为层承诺（减速、让行、怎么停）作为加分项，不当主打。精灵 G2 Max 的安规级只是公司的定位用语，不要把它当成已有认证或行为描述。Treat behaviour-layer promises (slowing, yielding, how to stop it) as a bonus, not the headline. The safety-grade wording for the Genie G2 Max is the company’s positioning language; do not present it as a certification or a description of behaviour.
- 推断Inference推断Inference
把保险与公众责任、厂商存续与支持，当作两条待核对的线索：试点前先问清能否投保、责任由谁承担，转常态合同写明备件和支持的最短保障期。Treat insurance and public liability, and supplier survival and support, as two leads to check: before a pilot find out whether cover is available and who bears liability, and in steady-use contracts state a minimum period of spare parts and support.
- 推断Inference推断Inference
引用数字时写明机构、口径、日期，并把公司口径单独标出。规则时间表写进路线图，不要暗示产品已有尚不存在的标准认证。When quoting numbers, state the agency, scope and date, and mark company statements separately. Put the regulatory timetable into the roadmap, and do not imply a product holds a certification that does not exist yet.