Anthropic 研究测算机器人对岗位的暴露度：机器人可做 74% 的物理任务但仅 0.3% 具备成本竞争力

主要发现

• 我们提出了一个机器人暴露指数，该指数基于机器人目前执行工作任务的能力。
• 我们将机器人定义为能够感知和行动的自主物理机器，它们可以完成美国四分之三的体力任务，占工作时间的34%，但大多是在有限的环境中。接触机器人的工人更有可能是男性、受教育程度较低且薪酬较低。例如，驾驶和仓库工作高度暴露于当前可用的机器人；而护理和一般维修工作则不然，因为当今的机器人即使在高度受控的环境中也很少能完成这些工作。
• 总体而言，按工作时间计算，约80%的工作任务暴露于机器人或大型语言模型（LLM）。机器人可以完成LLM无法完成的工作。剩余未暴露的工作高度依赖人际交往，或需要当今机器人不具备的体力技能。
• 虽然机器人如今可以完成大多数体力工作任务，但它们比人力劳动昂贵得多。机器人仅在0.3%的工作任务上具有成本竞争力。如果机器人价格按照过去的趋势下降，这一比例达到10%需要40年。除了价格之外，能力、偏好和法规等因素也对机器人自动化构成了进一步障碍。
• 在过去50年中，更容易接触机器人的工作经历的工资和就业下降幅度大于其他工作。与此同时，工作暴露程度也在增加：每年，机器人能够完成约2%以前无法完成的体力工作。

引言

大型语言模型的进步提高了自动化大量工作的可能性。但许多工作是体力劳动。人工智能对经济的影响将部分取决于机器人技术。¹

预测机器人技术的进步速度很困难，但我们认为，列举当今的能力可以洞察未来几年的发展。我们开发了一种衡量工作对机器人暴露程度的方法，使用Claude评估当今机器人在执行工作任务方面的能力。当机器人能在更不受控制的环境中完成更多任务时，该工作暴露程度更高。当机器人能以低于人类工人的成本完成相同任务时，就具有成本竞争力。

我们发现，机器人已经可以完成美国74%的体力任务，占工作时间的34%。机器人和LLM共同暴露了五分之四的就业。

但我们也发现了采用的重大障碍：大多数机器人需要高度结构化的环境，并且仅在0.3%的工作中具有与人类竞争的成本优势。如果机器人价格按照过去的趋势下降，这一比例达到10%就需要40年。除了成本之外，主要障碍是能力，例如解开电线所需的灵巧性。人类偏好和法规进一步限制了机器人在相当一部分任务中的采用。

我们的核心前提是，当机器人今天已经可以完成某项工作时，该工作更有可能受到影响。一项跨越50年的回溯测试验证了这一方法：从1977年至今，更容易接触现有机器人的工作在随后几十年中经历了工资和就业下降。

如果以过去为鉴，出租车司机和仓库包装工将比护士和机械师更早看到变化。我们预计，体力工作将首先在机器人已有立足之地的地方实现自动化。

当今与未来的机器人

如今大多数机器人都在工厂等受控环境中运行。机器人通常需要被编程才能与物理世界交互，而人类则能适应自己的工作环境。这一直是机器人实现商业可行性的主要障碍。² 但 AI 帮助机器人解读并回应周围环境，使仓库机器人和自动驾驶车辆能够与人类协同工作。³

许多观察者预计 AI 将迅速提升机器人能力。⁴ 尽管机器人支出仍仅占美国设备总投资的约 1%，但商业调查显示，美国的机器人采用率可能在三年内几乎翻倍。⁵ 各公司正投入数十亿美元开发能力媲美人类体能的 AI 驱动机器人。⁶

很难准确预测这些努力将如何影响就业和生产率。按当前机器人任务覆盖范围对职业进行的排名，揭示了影响将最先出现在哪里。机器人应当先影响它们已经能胜任的工作，然后才是只有借助新技术、并在某些情况下需要配套法规才能完成的工作。当前能力也是具体且可衡量的，而对未来能力的预测则必须押注哪些技术会成功。⁷ 仍在开发中的机器人支持我们聚焦于当前暴露度：各公司正在汽车工厂和仓库中测试人形机器人，这些结构化环境中机器人已很常见。⁸

衡量暴露度

我们的分析使用 O*NET 的数据，这是一个包含约 900 种职业的数据库，并关联了约 19,000 项工作任务的描述。我们识别出一组我们认为没有机器人就无法自动化的体力工作任务。为此，我们让 Claude 依据一套衡量体力、认知和人际工作要求的评分标准，对任务描述进行打分。⁹ “挖沟渠”这项任务是体力性的；“教舞蹈学生”也是，尽管它还需要认知和人际技能。而打字员用手“使用加法机或计算器，在报表、请购单或账单上计算并核对总数”，这项任务则不算。附录 A 给出了细节，附录 F 列出了我们的提示词。

我们在任务层面对机器人暴露度的衡量提出这样一个问题：今天的机器人能否执行这项任务，如果能，是在什么情况下？我们将机器人定义为能够感知并行动的自主物理机器，这包括扫描车辆以调整喷水器的洗车机，但不包括完全由外科医生控制的远程操作手术机器。¹⁰ 我们根据机器人执行任务所需的工作环境类型来衡量暴露度的高低。¹¹

具体来说，我们按照机器人能在何处执行每项任务，将任务分为四个暴露度递增的层级：

• E0：机器人无法执行任务。
• E1：机器人可在专门建造的机器人工作环境中执行任务，如工厂装配线。
• E2：机器人可在结构化的人类工作设施中执行任务，如物流仓库。
• E3：机器人可在非结构化环境中执行任务，如城市道路。

图 1 也展示了这一评分标准。

我们认为环境控制是衡量近期自动化风险的一个良好指标。由于机器人难以适应不可预测的环境，大多数已部署的机器人都在工程化环境中工作，例如在装配线上为汽车喷漆。虽然过去人们认为这些困难减缓了物理自动化的进程，但由 AI 驱动的机器人可以更好地适应其环境。¹²

为了确定暴露程度，我们指示 Claude 搜索与每项任务相关的特定机器人，并评估其能力和运行环境，直接引用来源。我们询问机器人是否能在该任务典型工作环境中结构化程度更高或更低的版本中执行该任务。让机器人完成诸如装载洗碗机这样看似简单的任务需要许多复杂的物理技能，因此只有已展示的机器人能力才计入考量。¹³ 机器人还必须以与人类相似的水平完成任务，同时考虑可靠性、错误率和速度。

例如，在 2010 年代初期，自动驾驶汽车无法在真实道路上“驾驶出租车或私人车辆运送乘客”。但它们确实在为测试而建造的模拟城镇中行驶，这是迈向当今自动驾驶汽车的一步。¹⁴ 我们的评分标准在当时会将驾驶乘客评为暴露等级 E1（机器人可以在专门建造的机器人工作环境中执行任务），而今天则为 E3（在非结构化环境中）。¹⁵

这一评分标准需要做出许多主观判断。O*NET 的任务描述往往很简洁，省略了那些对人类来说可能很容易但对机器人来说却很难的细节。¹⁶ 为了更具体地描述工作，我们引出关于当今任务如何执行以及这些情况发生频率的详细示例。然后 Claude 使用网络搜索来评估机器人能力，为这些示例的暴露程度打分。引用的来源必须展示机器人部署、商业销售或演示，如果我们省略依赖演示的评分，结果也类似。任务暴露程度由多数规则决定：机器人能够完成至少一半任务示例（按时间加权）的最不结构化环境。¹⁷

考虑“挖沟”这项任务。我们首先要求 Claude 提供示例，描述工人如今如何执行这项任务以及频率如何。例如，Claude 估计有 25% 的时间，这项任务需要“在开阔地面上切割线性沟槽”。另外 20% 的时间，它需要“在现有埋地管道、流线、电缆和导管周围小心手工挖掘”。这些活动需要不同类型的物理能力，例如导航和力量（线性沟槽），与精细运动灵活性和感知能力（在管道周围手工挖掘）相比。

Claude 将第一个示例评为暴露等级 E3，引用了一个可改装液压挖掘机以自主挖沟的控制系统。¹⁸ 如今的机器人无法在埋地管道周围小心挖掘，因此后一个示例在 E0 等级下未暴露。对于“挖沟”而言，更多示例类似于小心挖掘而非线性挖沟，因此总体而言，在我们的评分中这项任务未暴露。

图 2 给出了任务评分示例，以及来自我们暴露提示中每个暴露等级的更详细描述，以及为支持每个评分而引用的机器人。我们还计算了每个等级中物理任务的占比：在这里以及全文中，每项任务都按执行它的工人数量以及他们花在其上的工作时间比例进行加权，后者由 Claude 估算。¹⁹

机器人无法完成的任务（E0）按估算工作时间计算约占体力任务的四分之一。这些任务需要机器人难以具备的身体技能，例如“使用涂抹器或刷子为头发漂白、染色或上色”所需的精细手部灵巧度，以及“搭建脚手架或梯子以在地面以上组装结构”所需的力量、平衡能力和移动能力。

机器人可以在专门建造的环境中完成一半的体力任务，但无法在更广泛的环境中完成（E1）。食品准备机器人在传送带上打包即食餐食，而搭载 AI 的新型号可以适应不同的食材、份量和餐盘，因此“按照患者饮食要求将食物装入餐盘”这一任务被评为 E1。

另有 22% 的体力任务可由机器人在结构化的人类工作场所中完成（E2）。例如，医院配送机器人可以“向患者、护士站或手术室配送药物或药品用品”。如今，机器人在非结构化环境中只能完成 2% 的体力任务（E3）。这些任务通常涉及驾驶，由自动驾驶汽车、拖拉机和其他车辆完成。²⁰

我们随附发布的数据包含 Claude 对所有已评分任务的推理过程和引用来源。例如，Claude 将“在平焊、立焊或仰焊位置焊接部件”这一任务评为 E1，并引用了一篇关于 AI 驱动机器人焊工的文章：“Path Robotics 表示，其两个焊接单元都能自主焊接钢制部件，并已部署在美国和加拿大的制造车间中。”²¹

我们为工作定义了一个机器人暴露指数，该指数按 0–3 分制对某项工作的各项任务暴露评分取平均值。该指数随暴露任务占比以及机器人在这些任务上的能力而上升，并按估算工作时间对任务进行加权。

理想的暴露指数应能完美预测未来几年的机器人自动化。由于我们无法预知未来，因此转而尝试利用有关工作和任务描述的历史数据来验证我们的衡量指标。我们对 1977 年以来若干年份的机器人暴露程度进行了评分，并将暴露程度与工资和就业变化联系起来。当某项工作对机器人的暴露程度更高时，其工资和就业在随后几十年中下降，即使在考虑行业趋势和其他潜在混杂因素之后也是如此。这些结果详见附录 B，让我们有一定信心认为我们的衡量指标能够预测未来的工作颠覆。

图 3 总结了美国经济中所有工作任务的机器人暴露程度，按某项职业在每个任务上花费的时间估计比例以及该职业的就业人数对任务进行加权。认知和人际工作按工作时间计算占任务的 54%。其余 46% 的任务是体力任务。在所有任务中，约 12% 目前任何机器人都无法完成（E0），23% 可由机器人在专门建造的环境中完成（E1），10% 可在结构化的人类工作场所中完成（E2），1% 可在非结构化环境中完成（E3）。计入任何暴露级别，74% 的体力工作，即全部工作的 34%，可在某些情况下由机器人完成。

机器人可以在经济的许多领域完成体力任务。自主移动仓库机器人行驶到装卸码头和拖车内部。它们使用吸盘抓取包裹并将其装载到移动传送带上。²² 由于这些机器人在结构化的人类工作场所中导航，许多仓储任务被评为 E2，例如“用手或使用卡车、拖拉机或其他设备，将货物、库存或其他材料移入或移出存储或生产区域、装卸码头、送货车辆、船舶或集装箱。”导航和搬运方面的进步使这类机器人在物流和运输领域变得普遍。²³

图 4 显示了按我们的机器人暴露指数排名最高的 10 个职业。在这 10 个职业中，9 个是车辆操作员。驾驶类工作所引用的机器人包括自动驾驶汽车、拖拉机、卡车和铺路机。出租车司机的暴露程度最高，指数为 2.2。²⁴ 由于驾驶是该工作的主要部分，其按工作时间计算的中位任务暴露级别为 E3，而诸如“吸尘和清洁汽车内饰，以及清洗和抛光汽车外部”等辅助任务可由机器人完成，但所处的环境比开放道路更为结构化。自动驾驶汽车尚未大规模颠覆驾驶类工作，但其能力表明这些工作比其他工作面临更高的自动化风险。²⁵

一个职业可以涵盖多种工作场景。在美国，机器人可能只在其中部分场景中运行，而国外的机器人有时能覆盖所有场景。例如，理货员和订单拣货员在零售店和仓库中执行“用新的或调拨的商品补货货架、货架、货箱、料箱和台面”这一任务。在美国的仓库中，配备触觉传感器的AI驱动机器人负责拣选和存放商品。²⁶在美国零售店中，移动机器人扫描空货架，但必须提醒人工来补货。²⁷日本便利店中的机器人则自行补充冰箱货物。²⁸由于理货机器人需要物品被送到它们附近，该任务被评为E1。我们估计理货员有70%的时间用于像这样被评为E1的任务，另有16%用于被评为E2的任务。总体而言，他们在机器人暴露指数上得分为1.0。

如果某个职业的许多任务具有中等暴露度，或者其某些最耗时的任务具有高暴露度，那么该职业就可能暴露于机器人。比较一下接缝工（负责完成干墙）和回收与再生利用工人（负责分拣回收物）：前者的暴露指数为1.6，后者为1.7。

接缝工的任务分为无暴露和高暴露两类。建造内墙的工人使用纸带和一种称为“泥”的膏状物来抹平干墙板上的接缝、连接处和螺丝。泥和纸带是脏乱的材料，因此机器人无法执行“将纸带压在接缝上，使纸带嵌入密封胶并密封接缝”这一任务。在第一层纸带干燥后，工人“再涂几层以填补孔洞并使表面光滑”。该任务被评为E3：一台自主干墙机器人使用AI扫描墙面，喷涂额外的密封层，并在干燥后打磨墙面。²⁹总体而言，按工作时间计算，接缝工有41%的任务暴露度为E3，而另有26%的任务机器人无法完成。

相比之下，对于回收与再生利用工人，按工作时间计算，机器人能以E2级别完成超过四分之三的任务。只有8%的工作时间无暴露。例如，配备计算机视觉和吸盘夹爪的机器人可以完成“将金属、玻璃、木材、纸张或塑料等材料分拣到相应的回收容器中”这一任务。³⁰这些机器人替代了从传送带上分拣回收物的工人，因此该任务及其他任务被评为E2。

机器人对接缝工的特定任务能力很强，而对回收工人来说暴露范围更广，但机器人的适应性较低。接缝工的脏乱任务可能成为自动化的瓶颈。也有可能最大的回收设施会率先使用机器人，而较小的设施则较晚采用。无论哪种情况，如果广泛部署，机器人似乎都可能改变这两类工作。

暴露岗位的特征

在建立了我们的暴露度量之后，我们现在来考虑暴露对劳动力市场可能意味着什么。

图 5 将机器人暴露指数最高五分位的工作者与未暴露工作者（约占另外 20% 的岗位）进行了比较。2020–2024 年美国社区调查的数据显示，高暴露工作者为女性的可能性低 20 个百分点，为西班牙裔的可能性高 16 个百分点。这些工作者拥有学士或更高学位的可能性也低 55 个百分点，每小时收入约少 30 美元，且面临的失业率是前者的两倍多。³¹

美国劳工统计局编制的职业要求统计数据还显示，机器人暴露岗位更可能需要搬运重物，以及在极端高温或靠近危险污染物的地方工作。这种暴露模式在许多方面与 LLM 的典型情况相反。³²

机器人与 LLM 暴露

我们接下来分析机器人是否将岗位受冲击的风险扩大到仅由 LLM 带来的风险之外。借鉴 Eloundou 等人（2024）以及 Massenkoff 和 McCrory（2026）的研究，我们比较两种职业暴露度量：

• LLM 暴露：LLM 可将所需时间减半的工作任务占比（在我们先前的工作中称为理论能力）。
• LLM 与机器人暴露：暴露于 LLM 或机器人的工作任务占比，对每项任务取 LLM 暴露与 E1+ 机器人暴露中的较高者。

机器人与 LLM 暴露评级衡量的是略有不同的概念，但两者都表明技术可以完成某项任务的大部分工作。虽然在实际岗位中整合 LLM 和机器人可能带来新挑战或催生新能力，但这两项度量基于当今的能力，为岗位暴露提供了初步洞察。

图 6 按大类职业绘制了仅暴露于 LLM（蓝色）以及暴露于 LLM 和机器人（橙色）的岗位暴露情况。与仅 LLM 相比，机器人使更多且不同类型的岗位暴露。虽然运输与搬运任务中仅有不到 15% 暴露于仅 LLM，但其中约 90% 的任务在机器人参与下暴露。同样，机器人将办公室与行政支持岗位（主要涉及计算机工作和轻度体力任务）的暴露提高到近 100%。总体而言，约一半的工作暴露于仅 LLM，但考虑机器人后这一比例升至 81%。

如今 LLM 和机器人无法完成哪些工作？以个人护理与服务类工作为例，其中约 40% 的任务是暴露的。面对面的社交互动和与人的身体接触——这两者都是 LLM 和机器人难以应对的——对这类工作很重要。这个例子表明，高度人际性的工作，或需要精细操作的工作，可能不太容易在近期被自动化。³³ 安装与维修、医疗支持、社区与社会服务等职业组也符合这一模式，暴露度相对较低。虽然 LLM 和机器人今天无法执行这些任务，但没有任何职业组总体暴露度低于 40%。而且随着技术进步，未来可能出现新型自动化。

未暴露的任务也有助于说明是什么让工作难以自动化。我们将文本相似的未暴露任务陈述分组，并让 Claude 描述这些任务（附录 C.1）。今天未暴露于 LLM 或机器人的任务往往需要动手和面对面，有时还受到监管。医疗任务常常兼具这些特征，例如“为患者给药并监测患者的反应或副作用”。

在附录 C.2 中，我们研究哪些障碍最可能阻碍体力任务的自动化。我们询问 Claude，如果这些障碍得不到解决，今天哪些障碍会阻止机器人完成每项体力任务中的很大一部分（无论是否暴露）。我们将这些障碍分为四类：能力、人类偏好、监管和成本。对于受能力限制的任务，我们让 Claude 从操作、规划与推理、移动与力量、感知中选出最重要的缺失技能。

我们发现，能力和成本是机器人应用的最大障碍。能力因素阻碍了约 70% 的体力任务的应用。操作能力尤为突出：除非机器人更擅长触摸和处理物体，否则一半的体力任务无法大规模自动化。规划与推理技能的不足——AI 似乎最有可能在这方面取得进步——限制了 8% 的体力任务的机器人自动化。对于几乎所有任务，成本也需要下降。

我们的估算表明，今天的监管不会允许机器人执行 14% 的体力任务。监管对医疗、保护性服务和教育领域的任务很重要，但对食品准备、清洁、生产、建筑、维修和物料搬运等任务则不那么重要，而这些约占体力工作的一半。人类偏好阻碍了机器人完成四分之一的任务。这可能是因为人们不信任机器人（去“给儿童穿衣和换尿布”），或者因为他们看重社交互动（去“迎接客人，引他们入座，并递上菜单和酒单”）。

如果机器人能力发展迅速，这些结果意味着什么？约 30% 的体力任务受到偏好或监管的限制。如果机器人变得有能力且具有成本效益，那么大多数体力工作都将暴露在风险之下，尽管其余部分可能成为拖累总体生产力提升速度的薄弱环节。能力进展的不均衡也可能减缓采用，如果机器人变得更聪明但仍无法像人类一样处理物体的话。话虽如此，更好的机器人也可能减少对自动化的抵制，例如通过缓解安全方面的担忧。我们接下来转向机器人成本，这是这些估计中最常见的障碍。

机器人成本与采用

尽管机器人今天可以完成大多数体力工作，但这样做对它们来说可能并不经济。用昂贵硬件制造的机器人不像软件或 LLM 那样容易被复制和分发。与评估能力但不评估成本的 LLM 暴露度指标相比，区分暴露于具有成本竞争力的机器人的工作尤为重要。³⁴

对于每一项暴露的任务，我们根据任务暴露度指标中引用的具体机器人，估算机器人执行该任务的年度成本。我们提示 Claude 使用网络搜索以及每项任务所引用的机器人列表来估算成本。自动化一项 E1 暴露的任务可能需要重新设计工作场所以适应机器人，而执行 E3 任务的机器人可能使用昂贵的传感器来在非结构化环境中导航。

估算涵盖了部署机器人执行每项任务的全部成本。³⁵ 为了使机器人和劳动力成本具有可比性，我们首先让 Claude 估算一名人类工人在一项任务上一年通常能产出多少。例如，负责“滚动、揉捏、切割或塑形面团”的面包师，估计每年能产出 180,000 个成型面团。然后 Claude 估算机器人生产相同产出需要多少成本。由于面包师在不同环境中执行这项任务，Claude 对零售、中型和工业面包店的机器人部署成本取平均值。固定成本按年化处理，以反映硬件寿命和资金的时间价值。

按这种方式计算，机器人执行一项任务的成本可以与工人薪酬进行比较。如果一项任务的机器人成本低于其劳动力成本，我们就说该任务暴露于具有成本竞争力的机器人；劳动力成本按 BLS 中该职业的总薪酬乘以花在该任务上的时间比例计算。³⁶ 标准经济模型预测这些任务会被自动化，不过考虑到采用摩擦和成本不确定性，这一指标更具提示性。³⁷ 详见附录 D.1 了解细节以及机器人任务成本的计算示例。

机器人通常执行多项工作任务，但我们估算的是一个职业中各项任务的成本。简单地将任务成本相加可能会重复计算机器人，或低估实际部署中的协调成本。我们按职业进行汇总，向 Claude 提供任务成本估算，并要求给出扣除冗余和新增项后的总成本。如果这一总机器人成本低于其总薪酬乘以机器人可执行任务所占时间比例，我们就说该职业暴露于具有成本竞争力的机器人。³⁸

图7展示了五种高度体力型职业，以及机器人执行其可自动化任务所需的成本估算。包装工和打包工是受成本竞争型机器人影响最大的职业。完成包装工和打包工的任务需要多种机器人，包括装填容器的机器人、在仓库中搬运物料的机器人、对货物进行视觉检测的机器人、竖立纸箱的机器人、打印并粘贴标签的机器人以及封箱的机器人。³⁹这些机器人的购置和安装成本估计超过200万美元，但可替代约14名工人的全年工作。固定成本按约10年使用寿命、8%的资本成本进行分摊。维护、兼职人工监督和能源等运营费用，使每年替代一名人工工人的成本约为45,000美元。由于包装工和打包工97%的时间都花在机器人可以完成的任务上，且其成本约为49,000美元，因此机器人完成这些工作的成本每年约低2,500美元。

美国目前雇用了约56万名包装工和打包工。企业面临采用摩擦，如监管或借贷限制，且这些成本估算仅为近似值。例如，Robotaxi的成本估计仅比出租车司机高约7,000美元，但面临监管障碍。⁴⁰尽管如此，对于成本竞争型机器人能够胜任的体力型工作，自动化风险可能更高。自2015年以来，包装工和打包工的就业人数下降了22%。BLS预计到2035年，该职业的岗位减少数量将在所有职业中排第11位，仅次于惩教人员和9个销售及办公室职业。⁴¹

对于图7中展示的其他职业，尽管机器人能完成大部分工作，但其成本远高于人类。在金属加工领域，配备摄像头和AI的机器人可以自主焊接。⁴²但人类焊工要完成许多为焊接做准备的任务，例如定位大型金属部件、爬梯子焊接难以触及的接头、检查质量以及打磨和精加工材料。要实现这些工作的自动化，所需机器人的总成本约为人类焊工的五倍。

洗碗工以及清洁工和保洁员的薪酬比焊工低25,000至30,000美元。但其机器人替代品的成本仍高出数倍。清洁工作难以标准化和流程化，且清洁机器人通常比人类慢，且仅限于特定任务。⁴³

图8通过绘制在假设的机器人成本统一下降情景下，对成本竞争型机器人的暴露程度，总结了机器人成本。例如，如果今天机器人成本降低20%，它们对280万工人所从事的体力型工作将具有成本竞争力。这些工人平均将42%的时间花在可被机器人替代的工作上，占经济中全部工作时间的0.8%。作为参考，行业和政府数据显示，自1990年代以来，机器人价格每年下降约3%。按此速度，大约需要七年才能实现20%的成本下降。附录E更详细地讨论了历史机器人价格数据。

虽然理论上机器人可以完成累计占全部工作时间 34% 的任务，但如今它们仅在 0.3% 的任务上具备成本竞争力。不过，这仍包括约 30 万名工人，机器人可以完成其 95% 的任务。⁴⁴ 如图 7 所示，如今机器人在包装工和打包工以及出租车司机岗位上已具备成本竞争力或接近具备，但对大多数其他工作而言还相差甚远。

要让机器人在今天 10% 的人类工作上具备成本竞争力，成本需要下降约 70%。按每年下降 3% 计算，这大约需要 40 年。话虽如此，像人形机器人这样能力更强的新机器人，或新的制造工艺，有可能降低自动化人类工作的成本。有报告表明，如果对人形机器人和其他机器人有强烈需求，全球产量可以迅速扩大。⁴⁵

人类与机器人之间的成本持平也不一定意味着高失业率或快速经济增长。工人仍在执行的任务可能成为瓶颈，而这些机器人成本估算往往包含人工监督、异常处理或维修工作。正如 Acemoglu 和 Restrepo（2019）以及 Jones 和 Tonetti（2026）所强调的，自动化只有在机器比人类便宜得多之后才会提升生产率，而不是在成本持平时。

我们在两个扩展中应用了这些成本估算。附录 D.2 计算了一种替代的职业暴露度量，它同时考虑了机器人能做什么以及它们能以多低的成本完成。与我们的主要、以能力为中心的度量相比，考虑成本后得出的工作排名非常相似（排名相关性为 0.95）。

附录 E 还给出了如果机器人变得更便宜、更高效时的机器人采用风格化情景。这些情景借鉴了自 1990 年代以来工业机器人的创新和采用历程，以及分析师对人形机器人的预测。我们固定任务和工资，并探讨历史成本下降和能力提升速度需要多久才能使机器人具备成本竞争力。

为了模拟机器人如何随时间承担新工作，我们用当今的机器人对 1977 年的工作任务进行暴露评级，并与 1977 年时的机器人暴露进行比较。1977 年，机器人无法完成 62% 的体力任务。如今，机器人可以完成这些相同任务中除 24% 之外的全部任务。随着时间推移，这意味着机器人每年能够完成此前无法完成的任务中的 2%（见附录 B.4）。

按每年成本下降 3% 计算，机器人在当今一半的体力工作中要到 2085 年才具备成本竞争力。但机器人技术的快速进步可能会加快这一时间线。在快速普及情景下，质量调整后成本下降速度最高可达四倍，且机器人掌握新任务的速度翻倍，那么到 2050 年机器人就能在一半的体力工作中具备成本竞争力。而若要自动化当今 90% 的体力工作，仍需 53 年。

需要说明的是，我们的情景对每项任务都施加相同的成本下降或能力提升幅度，因此各项任务会按照当前成本相对于工人工资的排序依次变得具有成本竞争力。这一排序反映了广泛的经济激励，但机器人公司也可能出于其他原因优先瞄准某些能力。例如，人形机器人演示有时会展示家务劳动，而机器人制造商可能会优先发展对自身运营有用的工厂工作，正如 AI 公司优先发展编程智能体一样。⁴⁶

总体而言，机器人需要在未来几十年保持创纪录的价格下降和质量提升速度，才能实现快速的体力自动化。即便如此，在这些成本预测下，对就业的影响仍不那么确定，因为这些预测排除了偏好和监管等力量，也排除了工资下降等反馈效应。

讨论

我们引入了一种衡量工作受机器人影响程度的新指标。当机器人能够执行某项工作的任务时，该工作即被视为受影响。像人类一样工作的机器人对受影响程度的贡献大于那些需要重建工作场所的机器人，比如工厂机器人。

我们发现，由于自动驾驶汽车的存在，驾驶类工作受影响程度最高。当今大多数机器人并不像无人驾驶汽车，只能在受控环境中运行。但机器人有可能以某种形式执行大多数体力工作，尽管在相同工作上它们的成本是人类的数倍。受机器人影响的工作与受 LLM 影响的工作截然不同。前者薪酬更低，体力要求更高。

我们希望这些指标有助于在 AI 不断扩展机器人能力的同时分析体力自动化。当今的影响模式或许能预示，能力更强的机器人未来将冲击哪些工作。定期更新这项工作还可以追踪机器人的能力进展。

我们的影响程度量表基于这样一个理念：如果当今的机器人已在某些场景中从事某项工作，那么该工作的风险就更大。我们发现，这种评估机器人的方式能够预测历史上的劳动力市场影响。但 AI 驱动的机器人可能会跨越我们的量表，完成它们今天无法完成的工作，例如学会爬梯子，或更灵巧地使用手臂和夹爪。

关于机器人和 AI 将如何重塑经济，仍有许多不确定性。AI 增长预测通常假设，机器人可以缓解的物理瓶颈将拖累认知工作自动化带来的收益。我们也没有考虑 AI 在不借助机器人的情况下如何影响体力工作，例如通过更好地预测工厂机器何时需要维护。近期而言，追踪 AI 影响的努力可以关注受影响职业，如司机和仓库包装工，以寻找早期受冲击的迹象。更长远来看，一个关键问题是工作本身将如何变化——随着机器人和 AI 的进步，工作或许会变得更加社交化和人际化。

附录

见此处。

作者

Russell Legate-Yang 和 Maxim Massenkoff。

致谢

James Akl、Tess Cotter、Sholto Douglas、Adam Farina、Megan Giacobetti、Ryan Heller、Johannes Hermle、Zoë Hitzig、Ben Jones、Chad Jones、Anton Korinek、Jan Leike、Eva Lyubich、Peter McCrory、Kerry Persen、Sarah Pollack、Santi Ruiz、Szymon Sacher、Monika Tuchowska、Zhengdong Wang、Heather Whitney、Nathan Wilmers、Kim Withee。

引用

请按如下方式引用：

@online{legateyang2026robots,
 author = {Legate-Yang, Russell and Massenkoff, Maxim},
 title = {What work can robots do?},
 date = {2026-09-30},
 year = {2026},
 url = {https://www.anthropic.com/research/what-work-can-robots-do}
}

参考文献

Abraham, Katharine G. and James R. Spletzer, "Are the New Jobs Good Jobs?," in Katharine G. Abraham, James R. Spletzer, and Michael J. Harper, eds., Labor in the New Economy, University of Chicago Press, 2010, 101-143.

Acemoglu, Daron and David Autor, "Skills, Tasks and Technologies: Implications for Employment and Earnings," in Handbook of Labor Economics, Vol. 4B, Elsevier, 2011, 1043-1171.

Acemoglu, Daron, David Autor, David Dorn, Gordon H. Hanson, and Brendan Price, "Import Competition and the Great US Employment Sag of the 2000s," Journal of Labor Economics, 2016, 34 (S1), S141-S198.

Acemoglu, Daron, Andrea Manera, and Pascual Restrepo, "Does the US Tax Code Favor Automation?," Brookings Papers on Economic Activity, Spring 2020, 231-300.

Acemoglu, Daron and Pascual Restrepo, "The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment," American Economic Review, 2018, 108 (6), 1488-1542.

Acemoglu, Daron and Pascual Restrepo, "Automation and New Tasks: How Technology Displaces and Reinstates Labor," Journal of Economic Perspectives, 2019, 33 (2), 3-30.

Acemoglu, Daron and Pascual Restrepo, "Robots and Jobs: Evidence from US Labor Markets," Journal of Political Economy, 2020, 128 (6), 2188-2244.

Acemoglu, Daron and Pascual Restrepo, "Tasks, Automation, and the Rise in U.S. Wage Inequality," Econometrica, 2022, 90 (5), 1973-2016.

Agaskar, Ameya, Sriram Siva, William Pickering, Kyle O'Brien, Charles Kekeh, Alexandre Ormiga Galvao Barbosa, Ang Li, Brianna Gallo Sarker, Alicia Chua, Mayur Nemade, Charun Thattai, Jiaming Di, Isaac Iyengar, Ramya Dharoor, Dino Kirouani, Jimmy Erskine, Tamir Hegazy, Scott Niekum, Usman A. Khan, Federico Pecora, and Joseph W. Durham, "DeepFleet: Multi-Agent Foundation Models for Mobile Robots," arXiv:2508.08574, 2025.

Aghion, Philippe, Benjamin F. Jones, and Charles I. Jones, "Artificial Intelligence and Economic Growth," in Ajay Agrawal, Joshua Gans, and Avi Goldfarb, eds., The Economics of Artificial Intelligence: An Agenda, University of Chicago Press, 2019, 237-282.

Airbus, "Airbus Concludes ATTOL with Fully Autonomous Flight Tests," press release, June 29, 2020. https://www.airbus.com/en/newsroom/press-releases/2020-06-airbus-concludes-attol-with-fully-autonomous-flight-tests

Airbus, "Airbus Tests New Technologies to Enhance Pilot Assistance," press release, January 12, 2023. https://www.airbus.com/en/newsroom/press-releases/2023-01-airbus-tests-new-technologies-to-enhance-pilot-assistance

Amazon, "Introducing Vulcan: Amazon's First Robot with a Sense of Touch," About Amazon, May 7, 2025. https://www.aboutamazon.com/news/operations/amazon-vulcan-robot-pick-stow-touch

Andreassen, Dag, "A Robotic Success from Jæren," The Norwegian Museum of Science and Technology, September 3, 2024. https://www.tekniskmuseum.no/en/stories/a-robotic-success-from-the-fighter

Andrychowicz, Marcin, Bowen Baker, Maciek Chociej, Rafal Józefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, Jonas Schneider, Szymon Sidor, Josh Tobin, Peter Welinder, Lilian Weng, and Wojciech Zaremba, "Learning Dexterous In-Hand Manipulation," International Journal of Robotics Research, 2020, 39 (1), 3-20.

Arantes, Ana P., Nadja Bressan, Ludymila R. Borges, and Chris A. McGibbon, "Evaluation of a Novel Real-Time Adaptive Assist-as-Needed Controller for Robot-Assisted Upper Extremity Rehabilitation Following Stroke," PLOS ONE, 2023, 18 (10), e0292627.

Atalay, Enghin, Phai Phongthiengtham, Sebastian Sotelo, and Daniel Tannenbaum, "The Evolution of Work in the United States," American Economic Journal: Applied Economics, 2020, 12 (2), 1-34.

Autor, David H., "Why Are There Still So Many Jobs? The History and Future of Workplace Automation," Journal of Economic Perspectives, 2015, 29 (3), 3-30.

Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller, "New Frontiers: The Origins and Content of New Work, 1940-2018," Quarterly Journal of Economics, 2024, 139 (3), 1399-1465.

Autor, David H. and David Dorn, "The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market," American Economic Review, 2013, 103 (5), 1553-1597.

Autor, David H., David Dorn, and Gordon H. Hanson, "The China Syndrome: Local Labor Market Effects of Import Competition in the United States," American Economic Review, 2013, 103 (6), 2121-2168.

Autor, David H., Lawrence F. Katz, and Alan B. Krueger, "Computing Inequality: Have Computers Changed the Labor Market?," Quarterly Journal of Economics, 1998, 113 (4), 1169-1213.

Autor, David H., Frank Levy, and Richard J. Murnane, "The Skill Content of Recent Technological Change: An Empirical Exploration," Quarterly Journal of Economics, 2003, 118 (4), 1279-1333.

Autor, David, David Mindell, and Elisabeth Reynolds, "The Work of the Future: Building Better Jobs in an Age of Intelligent Machines," MIT Task Force on the Work of the Future, 2020.

Autor, David H. and Neil Thompson, "Expertise," Journal of the European Economic Association, 2025, 23 (4), 1203-1271.

Badger Technologies, "Badger Technologies and Stop & Shop Stores Upgrade Marty the Robot at 300+ Locations to Improve Customer Shopping Experiences," press release, September 18, 2023. https://www.badger-technologies.com/news/press-releases/stop-and-shop-upgrade-marty-the-robot.html

Bank of America, "Humanoid Robots 101," BofA Global Research, April 29, 2025.

Bank of Japan, "Producer Price Index, 2020 Base: Robots [PRCG20_2201350029]," Time-Series Data Search, 2026. https://www.stat-search.boj.or.jp/

Baptista, Eduardo, Ju-min Park, and Laurie Chen, "Robots Poised for 'ChatGPT Moment,' Unitree CEO Says," Reuters, August 20, 2026. https://www.reuters.com/world/asia-pacific/robots-poised-chatgpt-moment-unitree-ceo-says-2026-08-20/

Billard, Aude and Danica Kragic, "Trends and Challenges in Robot Manipulation," Science, 2019, 364 (6446), eaat8414.

Binder, Damon, "AGI Will Set Off an Industrial Explosion," AI Frontiers, August 11, 2026. https://ai-frontiers.org/articles/agi-will-set-off-an-industrial-explosion

Bishop, Todd，"Amazon Tops 1 Million Robots: Here's What They Do，" GeekWire，2025年7月1日。https://www.geekwire.com/2025/amazons-robot-workforce-hits-1-million-heres-what-they-all-do/

BMW Group，"BMW Group Advances the Use of Physical AI in Production with Figure 03 Project in Spartanburg，" 新闻稿，2026年6月25日。https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-group-advances-the-use-of-physical-ai-in-production-with-figure-03-project-in-spartanburg?language=en

Boston Consulting Group，"The Robotics Revolution: The Next Great Leap in Manufacturing，" 报告，2015年9月。https://www.bcg.com/publications/2015/lean-manufacturing-innovation-robotics-revolution-next-great-leap-manufacturing

Boston Dynamics，"Stretch，" 产品页面，2026年。https://bostondynamics.com/products/stretch/

Brooks, Rodney，"Why Today's Humanoids Won't Learn Dexterity，" rodneybrooks.com，2025年9月26日。https://rodneybrooks.com/why-todays-humanoids-wont-learn-dexterity/

Built Robotics，"The World's First Fully Autonomous Trenching Solution，" Exosystem 产品页面，2026年。https://www.builtrobotics.com/solutions/trenching

Canvas，"Drywall Finishing Robot，" 产品页面，JLG Industries，2026年。https://www.jlg.com/en/canvas

Cazzaniga, Mauro, Florence Jaumotte, Longji Li, Giovanni Melina, Augustus J. Panton, Carlo Pizzinelli, Emma J. Rockall, and Marina Mendes Tavares，"Gen-AI: Artificial Intelligence and the Future of Work，" IMF Staff Discussion Note SDN/2024/001，2024年。

Chaparro-Rico, Betsy D. M., Daniele Cafolla, Paolo Tortola, and Giuseppe Galardi，"Assessing Stiffness, Joint Torque and ROM for Paretic and Non-Paretic Lower Limbs during the Subacute Phase of Stroke Using Lokomat Tools，" Applied Sciences，2020，10 (18)，6168。

Chen, Jiafeng and Jonathan Roth，"Logs with Zeros? Some Problems and Solutions，" Quarterly Journal of Economics，2024，139 (2)，891-936。

Cleveland Clinic，"da Vinci Robotic Surgery，" Health Library，最后更新于2025年4月29日。https://my.clevelandclinic.org/health/treatments/16908-da-vinci-surgery

Cockburn, Iain M., Rebecca Henderson, and Scott Stern，"The Impact of Artificial Intelligence on Innovation: An Exploratory Analysis，" 载于 Ajay Agrawal, Joshua Gans, and Avi Goldfarb 编，The Economics of Artificial Intelligence: An Agenda，University of Chicago Press，2019，115-146。

Code of Federal Regulations，"Composition of Flight Crew，" 14 CFR 121.385，检索于2026年9月21日。https://www.ecfr.gov/current/title-14/section-121.385

CORDIS，"Automated System for Packaging Fresh Meat with Reduced Waste/Giveaway, Processing Time, Human Involvement and Contamination (AUTOMEATIC)，" European Commission，资助协议 775913，定期报告摘要，2018年。https://cordis.europa.eu/project/id/775913

Counterpoint Research，"Global Humanoid Robot Shipments Soar Nearly 300% YoY in H1 2026, Driven by Commercial Deployments，" 2026年8月20日。https://counterpointresearch.com/en/insights/global-humanoid-robot-shipments-soar-nearly-300-percent-yoy-in-h1-2026

Davidson, Tom and Rose Hadshar，"The Industrial Explosion，" Forethought Research，2025年5月。https://www.forethought.org/research/the-industrial-explosion

Davidson, Tom, Basil Halperin, Thomas Houlden, and Anton Korinek, "When Does Automating Research Produce Explosive Growth?," EconTAI Working Paper, January 2026. https://www.econtai.org/research/AutomatingAIResearch.pdf

Deming, David J., "The Growing Importance of Social Skills in the Labor Market," Quarterly Journal of Economics, 2017, 132 (4), 1593-1640.

Denain, Jean-Stanislas and Yann Rivière, "How Fast Could Robot Production Scale Up?," Epoch AI, April 2026. https://epoch.ai/publications/how-fast-could-robot-production-scale-up

Drankhan, Holly, "Developments in Robotic Teat Sprayers Cut Costs and Labor," Progressive Dairy, September 30, 2016. https://www.agproud.com/articles/28362-developments-in-robotic-teat-sprayers-cut-costs-and-labor

Dürr Group, "See and Drive Perfectly with x-light and x-wheel," duerrmore magazine, November 13, 2018. https://www.durr-group.com/en/duerrmore/2019/autonomous-driving/x-light-and-x-wheel

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, "GPTs are GPTs: Labor Market Impact Potential of LLMs," Science, 2024, 384 (6702), 1306-1308.

Epiroc, "Autonomous Material Handling: A Crucial Step Toward the Mine of the Future," customer story, April 17, 2024. https://www.epiroc.com/en-me/customer-stories/2024/autonomous-material-handling-a-crucial-step-toward-the-mine-of-the-future

Ewing, Jack, "Robots That Walk and Talk Are Coming to Car Factories," The New York Times, August 11, 2026. https://www.nytimes.com/2026/08/11/business/humanoid-robots-car-factories.html

Fabrico, "Maintenance Cost as Percent of RAV: The Benchmark That Cuts Through Excuses," blog, June 26, 2026. https://www.fabrico.io/blog/maintenance-cost-as-percent-of-rav/

Federal Aviation Administration (FAA), "Safety Framework for Aircraft Automation," September 2025. https://www.faa.gov/aircraft/air_cert/step/safety_framework_aircraft_automation

Fine, Sidney A., "A Structure of Worker Functions," Personnel and Guidance Journal, 1955, 34 (2), 66-73.

Fine, Sidney A. and Wretha W. Wiley, An Introduction to Functional Job Analysis, Kalamazoo, MI: W.E. Upjohn Institute for Employment Research, 1971.

Firoozi, Roya, Johnathan Tucker, Stephen Tian, Anirudha Majumdar, Jiankai Sun, Weiyu Liu, Yuke Zhu, Shuran Song, Ashish Kapoor, Karol Hausman, Brian Ichter, Danny Driess, Jiajun Wu, Cewu Lu, and Mac Schwager, "Foundation Models in Robotics: Applications, Challenges, and the Future," International Journal of Robotics Research, 2025, 44 (5), 701-739.

Fori Automation, "Toe Tool Systems," product page, 2025. https://foriauto.com/our-products/end-of-line/alignment-systems/wheel-alignment-systems/toe-tool-systems

Frey, Carl Benedikt and Michael A. Osborne, "The Future of Employment: How Susceptible Are Jobs to Computerisation?," Technological Forecasting and Social Change, 2017, 114, 254-280.

Friedel, William C., Jr., Thomas S. Bettencourt, and Daniel L. Freeman, "Tomato Harvesting Employing Electronic Sorting," U.S. Patent 4,033,099, issued July 5, 1977. https://patents.google.com/patent/US4033099A/en

Friedel, William C., Jr. and Charles F. Dietz, "Tomato Harvesting Sorting System," U.S. Patent 4,118,311, issued October 3, 1978. https://patents.google.com/patent/US4118311A/en

Goldman, Sharon，"一年前，Nvidia 的黄仁勋说机器人的‘ChatGPT 时刻’即将到来。如今他说它‘近在眼前’。但真是如此吗？"，Fortune，2026年1月6日。https://fortune.com/2026/01/06/nvidia-jensen-huang-chatgpt-moment-for-robotics/

González Huarte, Javier、Maite Ortiz de Zarate 和 Aitor Ibarguren，"面向航空领域线束制造的基于 CAD 的机器人编程解决方案"，Robotics，2023，12 (5)，130。

Graetz, Georg 和 Guy Michaels，"工作中的机器人"，Review of Economics and Statistics，2018，100 (5)，753-768。

Gridwise，"Robotaxi 如何影响自动驾驶城市的司机收入？"，博客，2025年8月。https://gridwise.io/blog/how-are-robotaxis-affecting-driver-pay-in-av-cities

Hirsch, Barry T.、David A. Macpherson 和 William E. Even，"来自 CPS 的工会会员、覆盖率和收入" [数据集]，Unionstats.com，2026年。https://unionstats.com/

Honerkamp, Paxton，"在 Uber-Waymo 分道扬镳之际，一场关于自动驾驶汽车的关键劳工之战正在首都打响"，CNBC，2026年8月1日。https://www.cnbc.com/2026/08/01/uber-waymo-autonomous-vehicle-regulation.html

Huffman, Wallace E.，"美国果蔬采收中省力机械化的发展现状"，Choices，2012，27 (2)。https://www.choicesmagazine.org/choices-magazine/theme-articles/immigration-and-agriculture/the-status-of-labor-saving-mechanization-in-us-fruit-and-vegetable-harvesting

Humlum, Anders，"机器人采用与劳动力市场动态"，Rockwool Foundation Research Unit Study Paper 175，2021年。

Hwang, Jyh-Jing、Runsheng Xu、Hubert Lin、Wei-Chih Hung、Jingwei Ji、Kristy Choi、Di Huang、Tong He、Paul Covington、Benjamin Sapp、Yin Zhou、James Guo、Dragomir Anguelov 和 Mingxing Tan，"EMMA：面向自动驾驶的端到端多模态模型"，Transactions on Machine Learning Research，2025年。

Ibarz, Julian、Jie Tan、Chelsea Finn、Mrinal Kalakrishnan、Peter Pastor 和 Sergey Levine，"如何用深度强化学习训练你的机器人：我们学到的经验教训"，International Journal of Robotics Research，2021，40 (4-5)，698-721。

Ingenia，"水果采摘机器人"，第93期，2022年12月。https://www.ingenia.org.uk/articles/fruit-picking-robots/

国际标准化组织（ISO），"机器人技术——词汇"，ISO 8373:2021，日内瓦，2021年。

Jarrett, Cosette，"关于帮助递送你的亚马逊包裹的 AI 驱动机器人的12个有趣事实"，About Amazon，2023年12月11日。https://www.aboutamazon.com/news/operations/amazon-robotics-cool-facts

Jones, Charles I. 和 Christopher Tonetti，"过去的自动化与未来的人工智能：薄弱环节如何驯服增长爆炸"，工作论文，斯坦福大学，2026年。https://web.stanford.edu/~chadj/JonesTonetti_Automation.pdf

Jurkat, Anne、Rainer Klump 和 Florian Schneider，"追踪机器人的崛起：IFR 数据库及其应用综述"，MPRA Paper 111812，Munich Personal RePEc Archive，2021年10月。https://mpra.ub.uni-muenchen.de/111812/

Katz, Lawrence F. 和 Kevin M. Murphy，"1963-1987年相对工资的变化：供给与需求因素"，Quarterly Journal of Economics，1992，107 (1)，35-78。

Kawasaki Robotics，《站在 FRP 自动化前沿：KADO × IEC 开创的喷涂与切割工艺创新》，案例研究，2026。https://kawasakirobotics.com/case-studies/at-the-forefront-of-frp-automation-spray-up-and-cutting-process-innovations-pioneered-by-kado-x-iec/

Kemp, Charles C., Aaron Edsinger, and Eduardo Torres-Jara，《人类环境中机器人操作的挑战》，IEEE Robotics & Automation Magazine，2007，14 (1)，20-29。

Kim, Moo Jin, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P. Foster, Pannag R. Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn, "OpenVLA: An Open-Source Vision-Language-Action Model," in Proceedings of the 8th Conference on Robot Learning (CoRL 2024), Proceedings of Machine Learning Research, 2025, 270, 2679-2713.

Kober, Jens, J. Andrew Bagnell, and Jan Peters, "Reinforcement Learning in Robotics: A Survey," International Journal of Robotics Research, 2013, 32 (11), 1238-1274.

Kogan, Leonid, Dimitris Papanikolaou, Lawrence D. W. Schmidt, and Bryan Seegmiller, "Technology and Labor Displacement: Evidence from Linking Patents with Worker-Level Data," NBER Working Paper 31846, 2023.

Kroemer, Oliver, Scott Niekum, and George Konidaris, "A Review of Robot Learning for Manipulation: Challenges, Representations, and Algorithms," Journal of Machine Learning Research, 2021, 22 (30), 1-82.

Link, Albert N., Zachary T. Oliver, and Alan C. O'Connor, "Economic Analysis of Technology Infrastructure Needs for Advanced Manufacturing: Advanced Robotics and Automation," NIST GCR 16-005, National Institute of Standards and Technology, August 2016. https://doi.org/10.6028/NIST.GCR.16-005

Lise, Jeremy and Fabien Postel-Vinay, "Multidimensional Skills, Sorting, and Human Capital Accumulation," American Economic Review, 2020, 110 (8), 2328-2376.

Ludwig, Jens, Sendhil Mullainathan, and Ashesh Rambachan, "Large Language Models: An Applied Econometric Framework," NBER Working Paper 33344, 2025.

Lung, Natalie, "New York's Robotaxi Plan Pulled in Setback to Waymo Expansion," Bloomberg News, February 19, 2026. https://www.bloomberg.com/news/articles/2026-02-19/new-york-s-robotaxi-plan-pulled-in-setback-to-waymo-expansion

Madrigal, Alexis C., "Inside Waymo's Secret World for Training Self-Driving Cars," The Atlantic, August 23, 2017. https://www.theatlantic.com/technology/archive/2017/08/inside-waymos-secret-testing-and-simulation-facilities/537648/

Mason, Matthew T., "Toward Robotic Manipulation," Annual Review of Control, Robotics, and Autonomous Systems, 2018, 1, 1-28.

Massenkoff, Maxim and Peter McCrory, "Labor Market Impacts of AI: A New Measure and Early Evidence," Anthropic, March 5, 2026. https://www.anthropic.com/research/labor-market-impacts

Mayekawa, "Deboning Machines," product page, 2026a. https://mayekawa.com/products/deboning_machines/

Mayekawa, "Reducing Workloads with Reliable Deboning Technology and Stabilizing Production to Meet Market Demands for Pork," With Mayekawa, 2026b (undated; retrieved September 29, 2026). https://mayekawa.com/with/stories/hamdas/

McKinsey & Company, "The Future of Robotics: Intelligent, Adaptable, and on Your Team," The Next Normal, 2026. https://www.mckinsey.com/featured-insights/the-next-normal/robotics

Metzger, Jean-Claude, Olivier Lambercy, Antonella Califfi, Daria Dinacci, Claudio Petrillo, Paolo Rossi, Fabio M. Conti, and Roger Gassert, "Assessment-Driven Selection and Adaptation of Exercise Difficulty in Robot-Assisted Therapy: A Pilot Study with a Hand Rehabilitation Robot," Journal of NeuroEngineering and Rehabilitation, 2014, 11, 154.

Moravec, Hans, Mind Children: The Future of Robot and Human Intelligence, Cambridge, MA: Harvard University Press, 1988.

Morgan Stanley, "Humanoid Robot Market Expected to Reach $5 Trillion by 2050," May 14, 2025. https://www.morganstanley.com/insights/articles/humanoid-robot-market-5-trillion-by-2050

Morgan Stanley Research, "Humanoids: Investment Implications of Embodied AI," report, June 26, 2024.

Navas-Reascos, Gabriel E., David Romero, Johan Stahre, and Alberto Caballero-Ruiz, "Wire Harness Assembly Process Supported by Collaborative Robots: Literature Review and Call for R&D," Robotics, 2022, 11 (3), 65.

Nissyoku Shimbun, "前川製作所が豚もも肉脱骨ロボット「ハムダス」発売" [Mayekawa Launches HAMDAS Pork Ham Deboning Robot], Nihon Shokuryo Shimbun, September 14, 1998. https://news.nissyoku.co.jp/news/nss-8423-0032

Oitzman, Mike, "Pharm Robotics Advances Automated Dairy Cow Healthcare," The Robot Report, July 13, 2025. https://www.therobotreport.com/pharm-robotics-advances-automated-dairy-cow-healthcare/

Oxmaint, "Food Manufacturing Maintenance Budgeting: Planning, Benchmarks, and Cost Control," March 28, 2026. https://oxmaint.com/industries/food-manufacturing/food-manufacturing-maintenance-budgeting-benchmarks-cost-control

Pallottino, Lucia, "Robotics for Warehouses and Logistics: Technologies, Challenges, and Future Directions," Annual Review of Control, Robotics, and Autonomous Systems, 2026, 9, 377-401.

Polanyi, Michael, The Tacit Dimension, Garden City, NY: Doubleday, 1966.

Pumacay, Wilbert, Ishika Singh, Jiafei Duan, Ranjay Krishna, Jesse Thomason, and Dieter Fox, "THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation," arXiv:2402.08191, 2024.

Reuters, "Meet SwagBot, the AI-Powered Robot Cattle Herder Preventing Soil Degradation," Rappler, December 12, 2024. https://www.rappler.com/technology/innovations/swagbot-artificial-intelligence-powered-robot-cattle-herder/

Rivière, Yann and Jean-Stanislas Denain, "Where Autonomy Works: Evaluating Robot Capabilities in 2026," Epoch AI, 2026. https://epoch.ai/publications/where-autonomy-works-evaluating-robot-capabilities-in-2026

Robbins, Jacob, "Apptronik Raises $520M as VC Funding for Humanoid Robotics Explodes 300%," PitchBook News, February 11, 2026. https://pitchbook.com/news/articles/apptronik-raises-520m-as-vc-funding-for-humanoid-robotics-explodes-300

The Robot Report, "Path Robotics Raises $100M to Automate Welding," October 28, 2024. https://www.therobotreport.com/path-robotics-raises-100m-to-automate-welding/

The Robot Report, "Okibo Opens New Jersey HQ and Launches EG7 Robot in U.S. Market," April 11, 2025. https://www.therobotreport.com/okibo-opens-new-jersey-hq-and-launches-eg7-robot-in-u-s-market/

Rosen, C. A., "Machine Vision and Robotics: Industrial Requirements," in George G. Dodd and Lothar Rossol, eds., Computer Vision and Sensor-Based Robots, Springer, 1979, 3-22.

SAE International, "Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles," SAE Recommended Practice J3016_202104, April 2021.

Sandvik, "Sandvik Introduces AutoMine AutoLoad 2.0 for Improved Autonomous Bucket Loading," press release, June 14, 2023. https://www.mining.sandvik/en/news-and-media/news-archive/2023/06/sandvik-introduces-automine-autoload-2.0-for-improved-autonomous-bucket-loading

Schenck RoTec, "Wheel Assembly Lines: Individual & Complete Solutions," product page, 2025. https://www.schenck-rotec.com/en-us/product/wheel-assembly-lines/

Scoop, "‘Golden Opportunity’ for Scott Technology with Silver Fern Farms," Scott Technology announcement, September 14, 2022. https://www.scoop.co.nz/stories/BU2209/S00201/golden-opportunity-for-scott-technology-with-silver-fern-farms.htm

Simbe Robotics, "Tally," product page, 2026. https://www.simberobotics.com/store-intelligence/tally

Statista, "Average Retail Electricity Price for Industrial Consumers in the United States from 1970 to 2025," April 28, 2026. https://www.statista.com/statistics/190680/us-industrial-consumer-price-estimates-for-retail-electricity-since-1970/

Suresh, Sudharshan, Haozhi Qi, Tingfan Wu, Taosha Fan, Luis Pineda, Mike Lambeta, Jitendra Malik, Mrinal Kalakrishnan, Roberto Calandra, Michael Kaess, Joseph Ortiz, and Mustafa Mukadam, "NeuralFeels with Neural Fields: Visuotactile Perception for In-Hand Manipulation," Science Robotics, 2024, 9 (96), eadl0628.

Svanberg, Maja, Wensu Li, Martin Fleming, Brian Goehring, and Neil Thompson, "Beyond AI Exposure: Which Tasks Are Cost-Effective to Automate with Computer Vision?," MIT FutureTech working paper, January 2024. https://futuretech.mit.edu/publication/beyond-ai-exposure-which-tasks-are-cost-effective-to-automate-with-computer-vision

Tamkin, Alex, Miles McCain, Kunal Handa, Esin Durmus, Liane Lovitt, Ankur Rathi, Saffron Huang, Alfred Mountfield, Jerry Hong, Stuart Ritchie, Michael Stern, Brian Clarke, Landon Goldberg, Theodore R. Sumers, Jared Mueller, William McEachen, Wes Mitchell, Shan Carter, Jack Clark, Jared Kaplan, and Deep Ganguli, "Clio: Privacy-Preserving Insights into Real-World AI Use," arXiv:2412.13678, 2024.

Tamkin, Alex and Peter McCrory, "Estimating AI Productivity Gains from Claude Conversations," Anthropic, 2025. https://www.anthropic.com/research/estimating-productivity-gains

Tedrake, Russ, Robotic Manipulation: Perception, Planning, and Control, course notes for MIT 6.421, 2026. http://manipulation.mit.edu

Telexistence, "Telexistence to Begin Installing AI Robot 'TX SCARA' in 300 FamilyMart Stores," press release, August 10, 2022. https://tx-inc.com/en/blog/2022/08/10/11737/

Thomas, Charlotte, "Rise of the Robots," METSTRADE, October 23, 2023. https://www.metstrade.com/news/superyacht/rise-of-the-robots

Trammell, Philip and Anton Korinek, "Economic Growth Under Transformative AI," Annual Review of Economics, 2026, 18, 589-611.

Transformation Hub Wiring Harness, "Robotics Challenge 2026," 2026. https://www.leitungssatz-hub.de/en/robotik-challenge/robotik-challenge-2026/

UBS, "Is the World Ready for 1bn Robots?," UBS Global Research Q-Series, June 17, 2025.

University of Michigan, "U-M Opens Mcity Test Environment for Connected and Driverless Vehicles," Michigan Engineering News, July 20, 2015. https://news.engin.umich.edu/2015/07/u-m-opens-mcity-test-environment-for-connected-and-driverless-vehicles/

U.S. Bureau of Economic Analysis (BEA), "Price Indexes for Private Fixed Investment in Equipment by Type: Computers and Peripheral Equipment [B935RG3A086NBEA]," NIPA Table 5.5.4, retrieved from FRED, Federal Reserve Bank of St. Louis, 2026. https://fred.stlouisfed.org/series/B935RG3A086NBEA

U.S. Bureau of Labor Statistics (BLS), "A New Methodology for Estimating Hours Worked for U.S. Labor Productivity Measurement," Office of Productivity and Technology technical note, November 3, 2022. https://www.bls.gov/productivity/technical-notes/labor-productivity-hours-worked-method-using-ces-all-employee-hours-nov-2022.htm

U.S. Bureau of Labor Statistics (BLS), "Occupational Employment and Wage Statistics (OEWS): May 2025 National Occupational Employment and Wage Estimates," 2025. https://www.bls.gov/oes/tables.htm

U.S. Bureau of Labor Statistics (BLS), "Consumer Price Index for All Urban Consumers: New Vehicles, U.S. City Average, Not Seasonally Adjusted [CUUR0000SETA01]," 2026a. https://www.bls.gov/cpi/

U.S. Bureau of Labor Statistics (BLS), "Employment Level - Agriculture and Related Industries [LNS12034560] and Employment Level [CE16OV]," Current Population Survey, retrieved from FRED, Federal Reserve Bank of St. Louis, 2026b. https://fred.stlouisfed.org/series/LNS12034560

U.S. Bureau of Labor Statistics (BLS), "Occupational Employment and Wage Statistics: Frequently Asked Questions," 2026c, retrieved September 21, 2026. https://www.bls.gov/oes/oes_ques.htm

U.S. Census Bureau, "Annual Capital Expenditures Survey: Robotic Equipment Expenditures, 2019-2022," 2024. https://www.census.gov/programs-surveys/aces.html

U.S. Department of Transportation Office of Inspector General, "Enhanced FAA Oversight Could Reduce Hazards Associated with Increased Use of Flight Deck Automation," Audit Report AV-2016-013, January 7, 2016. https://www.oig.dot.gov/sites/default/files/FAA%20Flight%20Decek%20Automation_Final%20Report%5E1-7-16.pdf

Vitiello, V., K.-W. Kwok, and G.-Z. Yang, "Introduction to Robot-Assisted Minimally Invasive Surgery (MIS)," in Paula Gomes, ed., Medical Robotics: Minimally Invasive Surgery, Woodhead Publishing, 2012, 1-40.

Wallace, Neal, "Butcher Bot Pays Off in Dollars per Carcase," Farmers Weekly (New Zealand), September 30, 2022. https://www.farmersweekly.co.nz/technology/butcher-bot-pays-off-in-dollars-per-carcase/

Wallace, Neal, "One Giant LEAP for Robo Meatworks," Farmers Weekly (New Zealand), April 22, 2024. https://www.farmersweekly.co.nz/technology/one-giant-leap-for-robo-meatworks/

Wang, Hanchen, Tianfan Fu, Yuanqi Du, Wenhao Gao, Kexin Huang, Ziming Liu, Payal Chandak, Shengchao Liu, Peter Van Katwyk, Andreea Deac, Anima Anandkumar, Karianne Bergen, Carla P. Gomes, Shirley Ho, Pushmeet Kohli, Joan Lasenby, Jure Leskovec, Tie-Yan Liu, Arjun Manrai, Debora Marks, Bharath Ramsundar, Le Song, Jimeng Sun, Jian Tang, Petar Veličković, Max Welling, Linfeng Zhang, Connor W. Coley, Yoshua Bengio, and Marinka Zitnik, "Scientific Discovery in the Age of Artificial Intelligence," Nature, 2023, 620, 47-60.

Waste360, "Q&A: AMP's Matanya Horowitz on Bringing AI to the Recycling World and How the Tech Is Advancing," November 2023. https://www.waste360.com/waste-recycling/q-a-amp-s-matanya-horowitz-on-bringing-ai-to-the-recycling-and-how-the-tech-is-advancing

WATTAgNet, "The New Generation of Automated Deboning Machines," October 1, 2018. https://www.wattagnet.com/broilers-turkeys/processing-slaughter/article/15526256/the-new-generation-of-automated-deboning-machines-wattagnet

Webb, Michael, "The Impact of Artificial Intelligence on the Labor Market," working paper, Stanford University, 2020.

Wessling, Brianna, "With New Funding, Monumental Plans to Bring Its Construction Robots to the U.S.," The Robot Report, July 17, 2026. https://www.therobotreport.com/with-new-funding-monumental-plans-bring-construction-robots-u-s/

Wishart-Smith, Heather, "Built Robotics Unveils Autonomous Pile Driving Robot, Expediting Solar Rollout," Forbes, March 20, 2023. https://www.forbes.com/sites/heatherwishartsmith/2023/03/20/built-robotics-unveils-autonomous-pile-driving-robot-expediting-solar-rollout/

Yotzov, Ivan, Jose Maria Barrero, Nicholas Bloom, Philip Bunn, Steven J. Davis, Kevin M. Foster, Aaron Jalca, Brent H. Meyer, Paul Mizen, Michael A. Navarrete, Pawel Smietanka, Gregory Thwaites, and Ben Zhe Wang, "Firm Data on AI," NBER Working Paper 34836, 2026. https://doi.org/10.3386/w34836

Yu, Zhen, Jingyu Wang, Ting Zuo, James Alm, Xun Li, and Xue Li, "Robotaxis Reduce Taxi Drivers' Income," Humanities and Social Sciences Communications, 2026, 13, article 629.

Zeira, Joseph, "Workers, Machines, and Economic Growth," Quarterly Journal of Economics, 1998, 113 (4), 1091-1117.

Ziemerink, Jan, "How to Manage Automatic Calf Feeders Successfully," Förster-Technik North America presentation, Vita Plus, 2018. https://www.vitaplus.com/wp-content/uploads/2018/06/Successfull-Calfraising_with-Auto-feeders2018_2.pdf

Data sources

Althobaiti, Shahad, Ahmad Alabdulkareem, Judy Hanwen Shen, Iyad Rahwan, Morgan Frank, Esteban Moro, and Alex Rutherford. "Dictionary of Titles" [dataset]. Harvard Dataverse, 2022. https://doi.org/10.7910/DVN/DQW8IP

Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller. Replication Data for: "New Frontiers: The Origins and Content of New Work, 1940-2018." Harvard Dataverse, V1.1, 2024. https://doi.org/10.7910/DVN/7RYD2E, retrieved July 30-31, 2026.

Autor, David H. and David Dorn. 2013. "The Growth of Low-Skill Service Jobs and the Polarization of the US Labor Market." American Economic Review 103 (5): 1553-1597. Occupation crosswalks (occ1990dd) and offshorability measure: https://www.ddorn.net/data.htm.

Bank of Japan. Producer Price Index, 2020 Base: Robots [PRCG20_2201350029]. Time-Series Data Search, retrieved September 16, 2026. https://www.stat-search.boj.or.jp/

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2024. "GPTs are GPTs: Labor Market Impact Potential of LLMs." Science 384 (6702): 1306-1308. Task-level exposure labels: https://github.com/openai/GPTs-are-GPTs.

Hirsch, Barry T., David A. Macpherson, and William E. Even. Union Membership, Coverage, and Earnings from the CPS [dataset]. Unionstats.com, 2026. https://unionstats.com/. Described in Hirsch, Barry T. and David A. Macpherson. 2003. "Union Membership and Coverage Database from the Current Population Survey: Note." Industrial and Labor Relations Review 56 (2): 349-354.

National Academy of Sciences, Committee on Occupational Classification and Analysis. Dictionary of Occupational Titles (DOT): Part I - Current Population Survey, April 1971, Augmented With DOT Characteristics, and Part II - Fourth Edition Dictionary of DOT Scores for 1970 Census Categories. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2006-09-06. https://doi.org/10.3886/ICPSR07845.v2

National Center for O*NET Development. O*NET Database, versions 29.3 (2025), 13.0 (2008), and 4.0 (2002); the O*NET-SOC taxonomy crosswalks; and the DOT to O*NET-SOC 2019 Crosswalk. Washington, DC: U.S. Department of Labor, Employment and Training Administration. https://www.onetcenter.org/database.html; https://www.onetcenter.org/taxonomy.html; https://www.onetcenter.org/crosswalks/dot/

Ruggles, Steven, Sarah Flood, Matthew Sobek, Daniel Backman, Grace Cooper, Julia A. Rivera Drew, Stephanie Richards, Renae Rodgers, Jonathan Schroeder, and Kari C.W. Williams. IPUMS USA: Version 16.0 [dataset]. Minneapolis, MN: IPUMS, 2025. https://doi.org/10.18128/D010.V16.0

U.S. Bureau of Economic Analysis (BEA). Personal Consumption Expenditures: Chain-type Price Index [PCEPI], retrieved from FRED, Federal Reserve Bank of St. Louis. https://fred.stlouisfed.org/series/PCEPI, September 21, 2026.

U.S. Bureau of Economic Analysis (BEA). Table 5.5.4. Price Indexes for Private Fixed Investment in Equipment by Type: Computers and peripheral equipment [B935RG3A086NBEA], retrieved from FRED, Federal Reserve Bank of St. Louis. https://fred.stlouisfed.org/series/B935RG3A086NBEA, September 16, 2026.

U.S. Bureau of Labor Statistics (BLS). Consumer Price Index for All Urban Consumers, U.S. city average, not seasonally adjusted: New vehicles [CUUR0000SETA01]. Washington, DC: U.S. Department of Labor. Retrieved September 16, 2026. https://www.bls.gov/cpi/

U.S. Bureau of Labor Statistics (BLS). Employer Costs for Employee Compensation (ECEC), 2004-2023 and March 2026. Washington, DC: U.S. Department of Labor. https://www.bls.gov/ecec/

U.S. Bureau of Labor Statistics (BLS). Employment Projections, 2025-35: Table 1.6, Occupations with the Largest Job Declines. Washington, DC: U.S. Department of Labor, August 27, 2026. https://www.bls.gov/emp/tables/occupations-largest-job-declines.htm

U.S. Bureau of Labor Statistics (BLS). Occupational Employment and Wage Statistics (OEWS): 2002, May 2008, May 2015, and May 2025 National Occupational Employment and Wage Estimates. Washington, DC: U.S. Department of Labor. https://www.bls.gov/oes/tables.htm

U.S. Bureau of Labor Statistics (BLS). Occupational Requirements Survey (ORS), 2023 and 2025. Washington, DC: U.S. Department of Labor. https://www.bls.gov/ors/

U.S. Census Bureau. 2018 Census Occupation Code List with Crosswalk, and 2018 ACS PUMS and 2018 SIPP Public-Use Occupation Code List [datasets], 2018. https://www2.census.gov/programs-surveys/demo/guidance/industry-occupation/2018-occupation-code-list-and-crosswalk.xlsx; https://www2.census.gov/programs-surveys/demo/guidance/industry-occupation/2018-ACS-PUMS-and-2018-SIPP-Public-Use-Occupation-Code-List.xlsx

U.S. Census Bureau. Annual Capital Expenditures Survey: Capital Expenditures for Structures and Equipment, Tables 1a and 1b, 2019-2022 releases. https://www.census.gov/data/tables/2022/econ/aces/2022-aces-summary.html

U.S. Department of Labor, Employment and Training Administration. Dictionary of Occupational Titles, Fourth Edition. Washington, DC: U.S. Government Printing Office, 1977.

United States Department of Labor, United States Employment Service, and North Carolina Occupational Analysis Field Center. Dictionary of Occupational Titles (DOT): Revised Fourth Edition, 1991. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2026-02-19. https://doi.org/10.3886/ICPSR06100.v2. Missing text from the Department of Labor Office of Administrative Law Judges DOT pages and https://occupationalinfo.org.

Webb, Michael. 2020. "The Impact of Artificial Intelligence on the Labor Market." Working paper, Stanford University. Robot exposure percentiles and 2010 labor-supply weights: https://web.stanford.edu/~mww/, retrieved July 18, 2026; now at https://www.michaelwebb.co/.

Footnotes

1. See for example Aghion et al. (2019), Davidson et al. (2026), and Jones and Tonetti (2026) on the role of weak links and complementarities in explosive economic growth scenarios. AI is often thought to affect digital work more than physical work; see for example Cazzaniga et al. (2024).

2. See for example Autor (2015) on “environmental control” as a barrier to physical automation. See also Section 6 of Link et al. (2016). We discuss this point in more detail in “Measuring exposure.”

3. See Firoozi et al. (2025) on AI foundation models for robots. Rivière and Denain (2026) discuss robot capabilities afforded by AI. See Agaskar et al. (2025) and Hwang et al. (2025) on AI-powered robots in warehousing and driving. AI could also speed the pace of innovation in robotics; see Cockburn et al. (2019), Wang et al. (2023), and Davidson et al. (2026).

4. See Baptista et al. (2026) and Goldman (2026) on a potential “ChatGPT moment” in robotics. See also forecasts from Morgan Stanley Research (2024) and McKinsey & Company (2026).

5. Robot investment data from the 2022 Annual Capital Expenditures Survey (U.S. Census Bureau, 2024). Surveys of US executives in November 2025 from Yotzov et al. (2026) show that 13% of firms use robotics now while 22% expect to within three years of surveying, weighted by employment. Expected use is broadly similar in the United Kingdom, Germany, and Australia.

6. See Robbins (2026) on investment in humanoid robotics startups and Ewing (2026) on efforts by large automakers to use humanoids.

7. For example, some robots use cameras only while others use touch sensors. See Brooks (2025) on these approaches, as well as research on vision and tactile robots by Andrychowicz et al. (2020) and Suresh et al. (2024).

8. See Counterpoint Research (2026) for a humanoid market overview; most humanoids today are used for entertainment and data collection. See BMW Group (2026) for a humanoid deployment in a car factory.

9. Our taxonomy builds on a long tradition in labor economics and occupational analysis. The predecessor to O*NET, the Dictionary of Occupational Titles, categorized jobs by “Data/People/Things.” See Fine (1955), Fine and Wiley (1971), Autor et al. (2003), Acemoglu and Autor (2011), and Deming (2017).

10. Precise definitions of a robot vary. We describe robots in our exposure prompt (Appendix F.3), and rely on Claude to judge which machines meet this criterion. The International Organization for Standardization defines a robot as a “programmed actuated mechanism with a degree of autonomy to perform locomotion, manipulation or positioning” and autonomy as the “ability to perform intended tasks based on current state and sensing, without human intervention” (ISO 2021).

11. Our focus on what today’s robots can do is similar to Eloundou et al. (2024), who measure AI exposure by what tasks LLMs could speed up as of 2023. Other AI exposure measures are described in Massenkoff and McCrory (2026).

12. See Autor (2015) on “environmental control” as a bottleneck to automation, and Autor et al. (2020) on how integration slows robot adoption. A report on the economic impacts of robotics for the National Institute of Standards and Technology highlights robot capabilities in unstructured environments; see Section 6 of Link et al. (2016). Kober et al. (2013), Ibarz et al. (2021), and Kim et al. (2025) discuss how robot performance generalizes to new tasks and environments. See Firoozi et al. (2025) on robotic foundation models versus specialized robot control systems. For surveys on progress in robotic manipulation, see Kemp et al. (2007) and Kroemer et al. (2021). See Pumacay et al. (2024) for manipulation benchmarks. For an inventory of robot capabilities in complex environments, see Rivière and Denain (2026).

13. Mason (2018) and Tedrake (2026), for example, discuss the complex skills robots need to manipulate objects.

14. See for example Madrigal (2017) on a mock town at a former Air Force base and University of Michigan (2015) on Mcity, a similar facility.

15. Our ratings indeed score this task E3 today and E1 in a version of our measure that rates exposure as of 2008 (Appendix B). As in our rubric, performance in unstructured environments factors into standards for autonomous vehicle performance; see SAE International (2021).

16. Moravec (1988) notably makes this point. See also Polanyi (1966), applied to automation in Autor (2015).

17. Appendix A.4 presents variants of our exposure measure; these produce a similar ranking of jobs by exposure.

18. See the Built Robotics Exosystem (2026).

19. Results are similar when also weighting across occupations by usual weekly hours from the 2020–2024 American Community Survey.

20. Humans also work in controlled environments: tasks like “Deliver medications or pharmaceutical supplies to patients, nursing stations, or surgery” are done in hospitals and clinics. In Appendix A.4 we measure robot exposure by the gap between where robots can do a task and where humans do it. This produces nearly the same ranking of jobs by exposure.

21. The Robot Report (2024).

22. See the Boston Dynamics Stretch.

23. See Pallottino (2026).

24. Employment data in Figure 4 come from the BLS OEWS, so independent contractors for ride-sharing services are excluded from taxi driver employment.

25. Early evidence suggests pay declines for drivers exposed to autonomous ride-hailing. See Gridwise (2025) and Yu et al. (2026).

26. See Amazon’s Vulcan robot.

27. See the Badger Marty and the Simbe Tally.

28. See the Telexistence TX SCARA.

29. See the Canvas drywall robot.

30. See the AMP Cortex (Waste360, 2023).

31. Restricting to physical jobs (at least half physical by work time), exposed workers are still paid less, less educated, less often female, more often Hispanic, and more often unemployed, though most gaps are somewhat smaller.

32. See for example Massenkoff and McCrory (2026).

33. Deming (2017) and Frey and Osborne (2017).

34. Analyses of automation and the labor market in the past few decades generally emphasize computers and software over physical automation. See for example Autor et al. (2003) on computerization and Acemoglu and Restrepo (2020) on the limited total impact of robots on US jobs so far. Svanberg et al. (2024) estimate costs to automate computer vision tasks.

35. We instruct Claude to include costs of the robot itself, accessories, integration and installation, maintenance and service, software, energy and other operating inputs, oversight and operation, insurance, and end-of-life decommissioning, plus any other costs. Claude annualizes fixed costs with an estimated cost of capital.

36. We estimate total compensation by multiplying BLS occupation wages by the ratio between total compensation and wages plus paid leave by occupation group in the BLS Employer Costs for Employee Compensation survey. Total compensation includes paid leave, supplemental pay, insurance, retirement and savings, and legally required employer payments like FICA and unemployment insurance.

37. On automation in task models, see for example Zeira (1998), Acemoglu and Autor (2011), and Acemoglu and Restrepo (2018).

38. In these comparisons, robots may be cheaper than workers but only do a few of their tasks. For example, archivists perform two physical tasks; their only robot-exposed one is “Preserve records, documents, and objects, copying records to film, videotape, audiotape, disk, or computer formats as necessary.” We estimate that it costs $5,500 per year for robots to do this task. Archivists spend about 8% of their time on it, and their median total compensation is $85,000, so labor costs for this task are $6,800 per year. Robots are cost-competitive for this task, but automating one task is unlikely to displace archivists.

39. Cited robots and systems include the Douglas top-load case packer (packing), Boston Dynamics’ Stretch and Agility Robotics’ Digit (material handling), machine-vision inspection stations (inspection), the Combi robotic random case erector (carton assembly), print-and-apply labelers (labeling), and the 3M-Matic random case sealer (sealing).

40. See for example Honerkamp (2026) and Lung (2026).

41. August 2026 forecasts from the BLS Employment Projections program, 2025–35, Table 1.6. Packers and packagers rank 11th in job loss, behind correctional officers and jailers and 9 sales and office occupations (the SOC “high-level aggregation” name for groups 41 and 43). Employment data from the BLS OEWS for 2015 and 2025.

42. See for example the Path Robotics AW-3 autonomous welding cell.

43. Though cleaning robots might work longer hours, they are up to 10 times slower than humans. See Rivière and Denain (2026).

44. This figure includes tire builders and packers and packagers. In these occupations, at least 95% of tasks by work time are exposed to robots, and robot costs fall below total compensation for a share of workers.

45. See Davidson and Hadshar (2025) and Denain and Rivière (2026).

46. Compare LLM exposure in Eloundou et al. (2024) for computer programmers (25th most exposed by human raters, 6th by the LLM rater) and poets, lyricists and creative writers (11th by humans, 14th by the LLM). The two occupations were similarly exposed, but AI developers have since directed effort toward coding more than creative writing.