WEBVTT

1
00:00:01.300 --> 00:00:05.791
This is Anthropic, Read Aloud. Today: What work can robots do?.

2
00:00:07.091 --> 00:00:08.559
What work can robots do?

3
00:00:09.009 --> 00:00:21.470
What work can robots do? Anthropic's robot exposure index finds robots can do 3/4 of U S physical job tasks, but are cost-competitive for only 0.3 percent.

4
00:00:21.920 --> 00:00:26.348
From Anthropic Research, published September 30, 2026.

5
00:00:27.348 --> 00:00:28.320
Key findings.

6
00:00:28.770 --> 00:00:34.758
We present a robot exposure index based on how well robots can perform job tasks today.

7
00:00:34.998 --> 00:00:53.828
Robots, which we define as autonomous physical machines that sense and act, can perform three-quarters of physical tasks in the U S, making up 34 percent of working hours, but mostly in limited settings. Workers exposed to robots are more likely to be male, less educated, and lower paid.

8
00:00:53.948 --> 00:01:07.724
For example, driving and warehouse jobs are highly exposed to currently available robots; nursing and general repair jobs are not, since present-day robots can do little of their work even in highly controlled environments.

9
00:01:07.964 --> 00:01:25.728
Overall, about 80 percent of job tasks by working time are exposed to either robots or L L Ms. Robots do work where L L Ms cannot. The remaining unexposed work is highly interpersonal or requires physical skills that robots today don’t have.

10
00:01:25.968 --> 00:01:43.145
While robots can do most physical work tasks today, they are much more expensive than human labor. Robots are cost-competitive for just 0.3 percent of job tasks. If robot price declines follow past trends, it will take 40 years for that share to reach 10 percent.

11
00:01:43.265 --> 00:01:50.857
Beyond price, factors including capabilities, preferences, and regulations pose further barriers to robot automation.

12
00:01:51.097 --> 00:02:08.005
Over the past 50 years, jobs more exposed to robots experienced greater declines in wages and employment than others. At the same time, job exposure has grown: each year, robots have become able to do about 2 percent of the physical work they previously couldn’t.

13
00:02:08.805 --> 00:02:09.649
Introduction.

14
00:02:10.099 --> 00:02:22.887
Advances in large language models have raised the possibility of automating large swaths of work. But many jobs are physical. A I’s impact on the economy will in part depend on robotics.¹

15
00:02:23.337 --> 00:02:45.780
Predicting the pace of robot advances is difficult, but enumerating capabilities today, we argue, can give insight into the coming years. We develop a measure of job exposure to robots, using Claude to assess how well present-day robots can perform work tasks. A job is more exposed when robots can do more of its tasks in less controlled environments.

16
00:02:45.900 --> 00:02:51.175
A robot is cost-competitive when it can do the same task for cheaper than a human worker.

17
00:02:51.625 --> 00:03:04.895
We find that robots can already perform 74 percent of physical tasks in the U S, making up 34 percent of working hours. Robots and L L Ms together expose all but one-fifth of employment.

18
00:03:05.345 --> 00:03:28.619
But we also find significant barriers to adoption: most robots require highly structured environments, and are cost-competitive with people for just 0.3 percent of work. If robot price declines follow past trends, it will take 40 years for that share to reach just 10 percent. After cost, the main barrier is capability, such as the dexterity needed to untangle wires.

19
00:03:28.739 --> 00:03:34.819
Human preferences and regulations further limit robot adoption for a significant share of tasks.

20
00:03:35.269 --> 00:03:54.679
Our core premise is that jobs are more likely to be impacted when robots can already do their work today. A backtest across 50 years validates this approach: from 1977 to today, jobs that were more exposed to existing robots experienced wage and employment declines in later decades.

21
00:03:55.129 --> 00:04:07.593
If the past is any guide, taxi drivers and warehouse packers will see changes sooner than nurses and mechanics. We expect that physical work will first be automated where robots have a foothold today.

22
00:04:08.393 --> 00:04:10.009
Robots today and tomorrow.

23
00:04:10.459 --> 00:04:22.000
Most robots today operate in controlled environments like factories. Robots usually need to be programmed to interact with the physical world, whereas humans can adapt to their work environments.

24
00:04:22.120 --> 00:04:35.912
This has been a major hurdle for commercially viable robots.² But A I helps robots interpret and respond to their surroundings, allowing warehouse robots and autonomous vehicles to operate alongside humans.³

25
00:04:36.362 --> 00:04:59.657
Many observers expect A I to improve robot capabilities quickly.⁴ Although spending on robots remains about 1 percent of total U S equipment investment, business surveys suggest U S robot adoption could nearly double within three years.⁵ And firms are investing billions to develop A I-powered robots that match human physical abilities.⁶

26
00:05:00.107 --> 00:05:20.821
It’s difficult to predict exactly how these efforts will affect jobs and productivity. A ranking of occupations by current robot task coverage suggests where the impacts will appear first. Robots should affect jobs that they can already do before jobs they could do only with new technology and, in some cases, accommodating regulations.

27
00:05:20.941 --> 00:05:41.768
Current capabilities are also concrete and measurable, while forecasts of future capabilities must bet on which technologies will succeed.⁷ Robots still in development support our focus on current exposure: firms are testing humanoids in car factories and warehouses, structured environments where robots are already common.⁸

28
00:05:42.568 --> 00:05:43.802
Measuring exposure.

29
00:05:44.252 --> 00:05:59.631
Our analyses use data from ONET, a database of around 900 occupations linked with descriptions of around 19000 job tasks. We identify a set of physical job tasks that we think could not be automated without robots.

30
00:05:59.751 --> 00:06:16.301
To do so, we have Claude score task descriptions on a rubric measuring physical, cognitive, and interpersonal work requirements.⁹ The task “Dig trenches” is physical; so is “Teach dance students,” though this also requires cognitive and interpersonal skills.

31
00:06:16.421 --> 00:06:31.013
While typists use their hands to “Compute and verify totals on report forms, requisitions, or bills, using adding machine or calculator,” this task doesn’t count. Appendix A gives details, and Appendix F lists our prompts.

32
00:06:31.463 --> 00:06:39.310
Our task-level measure of robot exposure asks: can a robot today perform this task, and if so under what circumstances?

33
00:06:39.430 --> 00:06:58.743
We take robots to mean autonomous physical machines that sense and act, which includes car washes that scan cars to adjust sprayers and excludes teleoperated surgical machines fully controlled by a surgeon.¹⁰ We measure degrees of exposure by the kind of work environment a robot needs to perform a task.¹¹

34
00:06:59.193 --> 00:07:05.636
Specifically, we sort tasks into four tiers of increasing exposure by where a robot can perform each task:

35
00:07:06.086 --> 00:07:08.640
E0: Robot cannot perform task.

36
00:07:08.880 --> 00:07:15.597
E1: Robot can perform task in a purpose-built robotic work environment, like a factory assembly line.

37
00:07:15.837 --> 00:07:21.653
E2: Robot can perform task in a structured human work facility, like a logistics warehouse.

38
00:07:21.893 --> 00:07:26.921
E3: Robot can perform task in an unstructured environment, like a city road.

39
00:07:27.371 --> 00:07:29.765
Figure 1 also displays this rubric.

40
00:07:30.215 --> 00:07:48.199
Figure 1: Figure 1: Task robot exposure rubric Job tasks identified as physical are rated on this robot exposure rubric. For example, the job task of unloading boxes from a truck at a warehouse is rated E2 if there’s a warehouse robot that can do that task.

41
00:07:48.649 --> 00:08:04.862
We think environmental control is a good measure of near-term automation risk. Because it’s hard for robots to adapt to unpredictable environments, most deployed robots work in engineered environments, like those that spray paint cars on assembly lines.

42
00:08:04.982 --> 00:08:13.221
While these difficulties are thought to have slowed physical automation in the past, A I-powered robots could better adapt to their environments.¹²

43
00:08:13.671 --> 00:08:32.145
To determine exposure, we instruct Claude to search for specific robots relevant to each task and assess their capabilities and operating environments, quoting sources directly. We ask whether a robot could perform a task in versions of that task’s typical work environment that are more or less structured.

44
00:08:32.265 --> 00:08:49.870
Getting robots to do seemingly simple tasks like loading a dishwasher requires many complex physical skills, so only demonstrated robot capabilities count.¹³ Robots must also do a task similarly well to humans, factoring in reliability, error rates, and speed.

45
00:08:50.320 --> 00:09:00.771
For instance, self-driving cars couldn’t “Drive taxicabs or privately owned vehicles to transport passengers” on real roads in the early 2010s.

46
00:09:00.891 --> 00:09:19.664
But they did drive around mock towns built for testing, a step toward today’s autonomous vehicles.¹⁴ Our rubric would rate driving passengers at exposure level E1 at that time (robot can perform task in a purpose-built robotic work environment) and E3 today (in an unstructured environment).¹⁵

47
00:09:20.114 --> 00:09:42.984
This rubric requires many judgment calls. The ONET task statements are often terse, and omit details that may be easy for humans but hard for robots.¹⁶ To describe work more concretely, we elicit detailed examples of how tasks are performed today, and how often these occur. Claude then scores exposure for these examples using web search to gauge robot capabilities.

48
00:09:43.104 --> 00:10:02.538
Cited sources must show robot deployments, commercial sales, or demonstrations, and results are similar if we omit ratings relying on demonstrations. Task exposure is set by majority rule: the least structured environment in which robots can do at least half of a task’s examples, weighted by time.¹⁷

49
00:10:02.988 --> 00:10:11.483
Consider the task “Dig trenches.” We first ask Claude to provide examples describing how workers perform this task today and how often.

50
00:10:11.603 --> 00:10:31.610
For example, Claude estimates that 25 percent of the time, this task requires “cutting a linear trench in open ground.” Another 20 percent of the time, it requires “careful digging by hand around existing buried pipes, flowlines, cables and conduit.” These activities require different sorts of physical abilities,

51
00:10:31.730 --> 00:10:39.538
like navigation and strength (linear trenches) compared to fine motor dexterity and perception (digging by hand around pipes).

52
00:10:39.988 --> 00:11:03.732
Claude rates the first example at exposure E3, citing a control system that retrofits hydraulic excavators to autonomously dig trenches.¹⁸ Robots today can’t carefully dig around buried pipes, so the latter example is unexposed at E0. More examples for “Dig trenches” resemble careful digging than linear trenching, so overall this task is unexposed in our ratings.

53
00:11:04.182 --> 00:11:14.657
Figure 2 gives examples of task ratings, along with fuller descriptions of each exposure tier from our exposure prompt and a robot cited to support each rating.

54
00:11:14.777 --> 00:11:27.612
We also calculate the share of physical tasks in each tier: here and throughout, each task is weighted by the number of workers who do it, and by the fraction of working time they spend on it, which we estimate with Claude.¹⁹

55
00:11:28.062 --> 00:11:54.849
Figure 2: Figure 2: Example task exposure ratings Levels of robot exposure are as described in the prompt we use to rate exposure with Claude. Exposure is rated for 7594 job tasks classified as physical. Physical task shares in column 2 weight tasks by the fraction of time spent on each task in an occupation (estimated with Claude) and that occupation’s employment (BLS, 2025).

56
00:11:54.969 --> 00:12:15.632
For example, the RelayRx robot delivers drugs in hospitals. Hospitals are structured human workplaces, so the task “Deliver medications or pharmaceutical supplies to patients, nursing stations, or surgery” is rated E2. Tasks rated E2 make up 22 percent of physical tasks by working time.

57
00:12:16.082 --> 00:12:38.418
Tasks that robots can’t do (E0) make up about a quarter of physical tasks by estimated working time. These tasks require physical skills that are hard for robots, like fine motor dexterity for “Bleach, dye, or tint hair, using applicator or brush” and strength, balance, and mobility for “Erect scaffolding or ladders for assembling structures above ground level.”

58
00:12:38.868 --> 00:12:59.965
Robots can do half of physical tasks in purpose-built environments but not in wider settings (E1). Food prep robots pack ready-to-eat meals on conveyor belts, and new models with A I can adjust to different ingredients, portions, and trays, so the task “Assemble meal trays with foods in accordance with patients’ diets” is rated E1.

59
00:13:00.415 --> 00:13:21.653
Another 22 percent of physical tasks can be done by robots in structured human workplaces (E2). Hospital delivery robots, for example, “Deliver medications or pharmaceutical supplies to patients, nursing stations, or surgery.” Today’s robots do only 2 percent of physical tasks in unstructured environments (E3).

60
00:13:21.773 --> 00:13:26.915
These often involve driving, done by autonomous cars, tractors, and other vehicles.²⁰

61
00:13:27.365 --> 00:13:33.339
Our accompanying data release contains Claude’s reasoning and cited sources for all rated tasks.

62
00:13:33.459 --> 00:13:52.957
For example, Claude rates the task “Weld components in flat, vertical, or overhead positions” E1, and quotes an article about A I-powered robot welders: “Path Robotics said both of its welding cells can autonomously weld steel parts and are deployed in fabrication shops across the U S and Canada.”²¹

63
00:13:53.407 --> 00:14:10.045
We define a robot exposure index for jobs that averages a job’s task exposure ratings on a 0 to 3 scale. This index rises with the share of exposed tasks and robot capabilities on those tasks, weighting tasks by estimated working time.

64
00:14:10.495 --> 00:14:31.925
An ideal exposure index would perfectly predict robot automation in the coming years. Since we can’t look forward in time, we instead try to validate our measure using historical data on jobs and task descriptions. We rate robot exposure in several years since 1977, and link exposure to changes in wages and employment.

65
00:14:32.045 --> 00:14:49.783
When a job was more exposed to robots, its wages and employment fell in later decades, even after accounting for industry trends and other potential confounders. These results, detailed in Appendix B, give us some confidence that our measure can predict future job disruption.

66
00:14:50.583 --> 00:14:52.051
What work can robots do?

67
00:14:52.501 --> 00:15:14.647
Figure 3 summarizes robot exposure across all job tasks in the U S economy, weighting tasks by the estimated fraction of time spent per task for an occupation and that occupation’s employment. Cognitive and interpersonal work makes up 54 percent of tasks by working time. The remaining 46 percent of tasks are physical.

68
00:15:14.767 --> 00:15:40.239
As a share of all tasks, around 12 percent cannot be done by any robot today (E0), 23 percent can be done by robots in specially built environments (E1), 10 percent in structured human workplaces (E2), and 1 percent in unstructured environments (E3). Counting any exposure level, 74 percent of physical work, or 34 percent of all work, can be done by robots in some circumstances.

69
00:15:40.689 --> 00:16:04.959
Figure 3: Figure 3: Job tasks by robot exposure All job tasks are included. Tasks are weighted by the fraction of time an occupation spends on each task as well as that occupation’s employment. Robot exposure is rated for tasks identified as physical. For example, 23 percent of worker time is spent on tasks that robots can do in a purpose-built environment.

70
00:16:05.079 --> 00:16:28.810
Example tasks by category: teaching classes, designing systems (Cognitive and interpersonal); taking people into custody, providing first aid (E0); dispensing medications, assembling electronics (E1); cleaning floors, moving warehouse goods (E2); regulating field irrigation, inspecting property from the air, and driving, which robots perform in unstructured environments (E3).

71
00:16:28.930 --> 00:16:36.472
Employment data come from the BLS (2025), and we use Claude to estimate occupation time on tasks.

72
00:16:36.922 --> 00:16:46.266
Robots can do physical tasks in many parts of the economy. Autonomous mobile warehouse robots drive to loading docks and inside trailers.

73
00:16:46.386 --> 00:17:08.786
They use suction cups to grab packages and load them onto mobile conveyor belts.²² Since these robots navigate structured human workplaces, many warehousing tasks are rated E2, like “Move freight, stock, or other materials to and from storage or production areas, loading docks, delivery vehicles, ships, or containers, by hand or using trucks, tractors,

74
00:17:08.906 --> 00:17:17.735
or other equipment.” Advances in navigation and handling have made robots like these common in logistics and transportation.²³

75
00:17:18.185 --> 00:17:33.685
Figure 4 shows the 10 most exposed occupations by our robot exposure index. Of these 10 occupations, 9 are vehicle operators. Robots cited for driving jobs include autonomous cars, tractors, trucks, and pavers.

76
00:17:33.805 --> 00:17:54.178
Taxi drivers lead exposure with an index of 2.2.²⁴ Since driving is most of the job, their median task by working time is exposed at E3, while auxiliary tasks like “Vacuum and clean interiors and wash and polish exteriors of automobiles” are doable by robots but in more structured settings than open roads.

77
00:17:54.298 --> 00:18:03.766
Autonomous vehicles have yet to upend driving jobs at scale, but their capabilities suggest that these jobs are at higher risk of automation than others.²⁵

78
00:18:04.216 --> 00:18:28.749
Figure 4: Figure 4: Most exposed occupations Exposure ranges from 0 (no task can be done by a robot) to 3 (every task can be done by a robot in an unstructured environment). Job exposure averages task exposure, weighting by time spent on each task. A highly exposed task is listed for each occupation along with a cited robot that can perform this task.

79
00:18:28.869 --> 00:18:45.914
The 10 most exposed occupations with at least 20000 jobs are displayed. For example, averaging task exposure for shuttle drivers and chauffeurs gives an index of 2.0 out of 3. Claude cites Waymo autonomous vehicles as performing some of their tasks.

80
00:18:46.034 --> 00:19:00.835
Employment data from BLS (2025) sometimes combine several ONET occupation titles; we split their employment equally, as for recycling and reclamation workers. We estimate task time requirements with Claude.

81
00:19:01.285 --> 00:19:22.426
An occupation can cover many work settings. In the U S, robots may only operate in some of these settings, while robots abroad sometimes cover all of them. For example, stockers and order fillers do the task “Stock shelves, racks, cases, bins, and tables with new or transferred merchandise” in retail stores and in warehouses.

82
00:19:22.546 --> 00:19:45.144
In U S warehouses, A I-powered robots with touch sensors pick and stow merchandise.²⁶ In U S retail, mobile robots scan for empty shelves, but must alert humans to restock them.²⁷ Robots in Japanese convenience stores restock fridges themselves.²⁸ Since stocker robots require items to be brought near them, this task is rated E1.

83
00:19:45.264 --> 00:19:58.992
We estimate that stockers spend 70 percent of their time on tasks rated E1 like this one, and another 16 percent on tasks rated E2. In total, they score 1.0 on the robot exposure index.

84
00:19:59.442 --> 00:20:19.600
Jobs can be exposed to robots if many of their tasks are moderately exposed, or if some of their most time-intensive tasks are highly exposed. Compare tapers, who finish drywall and whose exposure index is 1.6, with recycling and reclamation workers, who sort recycling and whose exposure index is 1.7.

85
00:20:20.050 --> 00:20:32.892
Tasks for tapers are split between unexposed and highly exposed. Workers building interior walls use paper tape and a paste called mud to smooth over seams, joints, and screws in drywall sheets.

86
00:20:33.012 --> 00:20:56.735
Mud and tape are messy materials, so robots cannot perform the task “Press paper tape over joints to embed tape into sealing compound and to seal joints.” After the first tape coat dries, workers “Apply additional coats to fill in holes and make surfaces smooth.” This task is rated E3: an autonomous drywall robot uses A I to scan walls, spray additional coats of sealant,

87
00:20:56.855 --> 00:21:07.266
and sand walls once dry.²⁹ Overall, by work time, 41 percent of tapers’ tasks are exposed at E3, while robots cannot do another 26 percent.

88
00:21:07.716 --> 00:21:19.449
In contrast, for recycling and reclamation workers, robots can perform over three-quarters of tasks by work time at level E2. Only 8 percent of their work time is unexposed.

89
00:21:19.569 --> 00:21:38.992
For example, robots with computer vision and suction grippers can do the task “Sort materials, such as metals, glass, wood, paper or plastics, into appropriate containers for recycling.”³⁰ These robots stand in for workers picking recycling from conveyor belts, so this and other tasks are rated E2.

90
00:21:39.442 --> 00:22:03.344
Robots are highly capable at specific tasks for tapers, while for recycling workers exposure is broader, but robots are less adaptable. Messy tasks for tapers may bottleneck automation. It’s also possible that the largest recycling facilities use robots early, while smaller facilities adopt them later. Either way, robots seem likely to change both jobs if widely deployed.

91
00:22:04.144 --> 00:22:06.250
Characteristics of exposed jobs.

92
00:22:06.700 --> 00:22:12.374
Having established our exposure measure, we now consider what exposure might imply for the labor market.

93
00:22:12.824 --> 00:22:36.729
Figure 5 compares workers in the top quintile of the robot exposure index with unexposed workers, who make up another 20 percent or so of jobs. Data from the 2020 to 2024 American Community Survey show that highly exposed workers are 20 percentage points less likely to be female and 16 percentage points more likely to be Hispanic.

94
00:22:36.849 --> 00:22:48.841
These workers are also 55 percentage points less likely to hold a bachelor’s degree or higher, earn around 30 dollarsless per hour, and face an unemployment rate more than twice as high.³¹

95
00:22:49.291 --> 00:23:09.555
Occupational requirement statistics compiled by the Bureau of Labor Statistics also show that robot-exposed jobs are much more likely to require carrying weight and working in extreme heat or near hazardous contaminants. This exposure pattern is in many ways the opposite of what’s typically found for L L Ms.³²

96
00:23:10.005 --> 00:23:35.993
Figure 5: Figure 5: Occupation characteristics by robot exposure The most exposed workers are those in the top quintile of the robot exposure index; unexposed workers have zero exposure and make up about another 20 percent of jobs. Demographics, education, and labor market data come from the 2020 to 2024 American Community Survey, save for union membership and licensing.

97
00:23:36.113 --> 00:24:01.815
Union membership comes from statistics tabulated by Hirsch and others (2026) for the 2020 to 2024 Current Population Surveys. Licensing comes from the 2025 BLS Occupational Requirements Survey, as do physical demand and work environment variables (for some occupations, these come from the 2023 survey). Statistics are employment weighted.

98
00:24:02.615 --> 00:24:04.547
Robot and L L M exposure.

99
00:24:04.997 --> 00:24:21.722
We next analyze whether robots broaden job disruption risks beyond those from L L Ms alone. Drawing on Eloundou and others (2024) and Massenkoff and McCrory (2026), we compare two occupation exposure measures:

100
00:24:22.172 --> 00:24:31.275
L L M exposure: the share of job tasks for which an L L M could halve required time (called theoretical capability in our prior work ).

101
00:24:31.515 --> 00:24:44.881
L L M and robot exposure: the share of job tasks exposed to either L L Ms or robots, taking the higher of L L M exposure and E1+ robot exposure for each task.

102
00:24:45.331 --> 00:25:07.859
Robot and L L M exposure ratings measure somewhat different concepts, but both indicate that technology could do much of a given task. While integrating L L Ms and robots in actual jobs may present new challenges or enable new capabilities, these two measures provide initial insight into job exposure based on today’s capabilities.

103
00:25:08.309 --> 00:25:32.885
Figure 6 plots job exposure to L L Ms alone (in blue) and to L L Ms and robots (in orange), grouped by broad occupation category. Robots expose more and different kinds of jobs than L L Ms alone. While less than 15 percent of transportation and moving tasks are exposed to L L Ms alone, about 90 percent of these tasks are exposed with robots.

104
00:25:33.005 --> 00:25:51.863
Similarly, robots raise exposure in office and admin support jobs, which largely involve computer work and light physical tasks, to nearly 100 percent. Overall, around half of work is exposed to L L Ms alone, but that rises to 81 percent when considering robots.

105
00:25:52.313 --> 00:26:19.608
Figure 6: Figure 6: L L M and robot exposure by occupation group L L M exposure is the share of tasks rated exposed in Eloundou and others (2024) (positive 𝛽). Occupation exposure to L L Ms and robots adds tasks rated at E1 or higher robot exposure. Occupation groups are 2-digit SOC codes, and group averages weight by task time shares and employment.

106
00:26:19.728 --> 00:26:37.778
In protective service occupations, for example, 31 percent of job tasks are exposed to L L Ms, and 61 percent are exposed to L L Ms or robots. Employment data come from BLS (2025). We estimate task time shares with Claude.

107
00:26:38.228 --> 00:26:56.736
What work can’t L L Ms and robots do today? Consider personal care and service jobs, for which around 40 percent of tasks are exposed. In-person social interaction and physical contact with humans, which both L L Ms and robots struggle with, are important for these jobs.

108
00:26:56.856 --> 00:27:16.411
This example suggests that highly interpersonal work, or work requiring delicate manipulation, may be less susceptible to near-term automation.³³ Occupation groups like installation and repair, healthcare support, and community and social service fit this pattern too, with relatively low exposure.

109
00:27:16.531 --> 00:27:30.014
While L L Ms and robots can’t perform these tasks today, no occupation group overall is less than 40 percent exposed. And advances in technology may make possible new kinds of automation over time.

110
00:27:30.464 --> 00:27:49.991
Unexposed tasks also help describe what could make work hard to automate. We group unexposed task statements that share similar text, and ask Claude to describe these (Appendix C.1). Tasks not exposed to L L Ms or robots today tend to be hands-on and face-to-face, and are sometimes regulated.

111
00:27:50.111 --> 00:27:58.873
Healthcare tasks often combine these features, like “Administer medications to patients and monitor patients for reactions or side effects.”

112
00:27:59.323 --> 00:28:20.336
In Appendix C.2, we study which barriers may most prevent the automation of physical tasks. We ask Claude which barriers today would prevent robots from doing a significant share of each physical task, exposed or not, if unresolved. We group these into four categories: capabilities, human preferences, regulation, and cost.

113
00:28:20.456 --> 00:28:30.263
For tasks limited by capabilities, we have Claude pick the most important missing skill from manipulation, planning and reasoning, mobility and strength, and perception.

114
00:28:30.713 --> 00:28:49.913
We find that capabilities and costs are the biggest impediments to robot adoption. Capabilities prevent adoption for around 70 percent of physical tasks. Manipulation capabilities stand out: half of physical tasks wouldn’t be automated at scale unless robots become more adept at touching and handling objects.

115
00:28:50.033 --> 00:29:02.492
Shortcomings in planning and reasoning skills, which A I seems most likely to improve, limit robot automation for 8 percent of physical tasks. For nearly all tasks, costs would also need to fall.

116
00:29:02.942 --> 00:29:23.941
Our estimates indicate that today’s regulations wouldn’t allow robots to perform 14 percent of physical tasks. Regulation matters for tasks in healthcare, protective service, and education, but less so for those in food preparation, cleaning, production, construction, repair, and material moving, which make up about half of physical work.

117
00:29:24.061 --> 00:29:40.845
Human preferences hold back robots for a quarter of tasks. That may be because people wouldn’t trust a robot (to “Dress children and change diapers”) or because they value social interaction (to “Greet guests, escort them to their seats, and present them with menus and wine lists”).

118
00:29:41.295 --> 00:30:01.988
If robot capabilities develop quickly, what do these results imply? Around 30 percent of physical tasks are limited either by preferences or by regulation. That would leave most physical work exposed if robots became capable and cost-effective, though the rest may become weak links that drag down aggregate productivity speedups.

119
00:30:02.108 --> 00:30:22.225
Uneven progress in capabilities could also slow adoption, if robots became smarter but still couldn’t handle objects like humans. That said, better robots could also reduce resistance to automation, for example by assuaging safety concerns. We turn next to robot costs, the most common barrier in these estimates.

120
00:30:23.025 --> 00:30:24.661
Robot costs and adoption.

121
00:30:25.111 --> 00:30:48.806
Though robots can perform most physical work today, it may not be economical for them to do so. Robots made with pricey hardware aren’t as easily copied and distributed as software or L L Ms. Compared to L L M exposure measures that assess capabilities but not costs, it’s especially important to distinguish jobs exposed to cost-competitive robots.³⁴

122
00:30:49.256 --> 00:31:04.936
For every exposed task, we estimate the annual cost for robots to perform this task based on the specific robots cited in our task exposure measures. We prompt Claude to estimate costs using web search and the list of robots cited for each task.

123
00:31:05.056 --> 00:31:17.676
Automating an E1-exposed task could require redesigning workplaces to suit robots, while robots that do E3 tasks may use expensive sensors to navigate unstructured environments.

124
00:31:18.126 --> 00:31:40.031
Estimates cover the full cost of deploying a robot to perform each task.³⁵ To make robot and labor costs comparable, we first ask Claude to estimate how much a human worker typically produces on a task in a year. Bakers who “Roll, knead, cut, or shape dough,” for example, produce an estimated 180000 shaped pieces per year.

125
00:31:40.151 --> 00:31:59.690
Claude then estimates what it would cost a robot to produce the same output. Because bakers do this task in different environments, Claude averages robot deployment costs across retail, mid-scale, and industrial bakeries. Fixed costs are annualized to reflect hardware lifespans and the time value of money.

126
00:32:00.140 --> 00:32:04.901
Calculated this way, costs for robots to do a task can be compared to worker pay.

127
00:32:05.021 --> 00:32:24.345
We say a task is exposed to cost-competitive robots if its robot cost is less than its labor cost, calculated as the occupation’s total compensation from the BLS scaled by the fraction of time spent on this task.³⁶ Standard economic models predict that these tasks are automated,

128
00:32:24.465 --> 00:32:35.448
though given adoption frictions and cost uncertainties this measure is more suggestive.³⁷ See Appendix D.1 for details and a worked example of robot task costs.

129
00:32:35.898 --> 00:32:58.314
Robots often perform multiple job tasks, but we estimate costs for individual tasks in an occupation. Simply adding up task costs may double count robots or undercount coordination costs in real deployments. We aggregate by occupation, providing Claude with task cost estimates and asking for a total cost, net of redundancies and new additions.

130
00:32:58.434 --> 00:33:11.511
We say that an occupation is exposed to cost-competitive robots if this total robot cost is less than its total compensation scaled by the fraction of time spent on tasks that robots can do.³⁸

131
00:33:11.961 --> 00:33:25.044
Figure 7 shows five highly physical jobs and cost estimates for robots to perform their exposed tasks. Packers and packagers are the largest occupation exposed to cost-competitive robots.

132
00:33:25.164 --> 00:33:48.161
Several robots are needed to do packers and packagers’ tasks, including those that pack containers, move materials around warehouses, visually inspect goods, erect cardboard boxes, print and apply labels, and seal packages.³⁹ These robots are estimated to cost over 2 dollarsmillion to purchase and install, but replace the yearly work of around 14 workers.

133
00:33:48.281 --> 00:34:05.279
Fixed costs are spread over a roughly 10-year service life at an 8 percent cost of capital. Operating expenses, such as maintenance, part-time human supervision, and energy, bring annual costs to around 45000 dollarsto replace one human worker per year.

134
00:34:05.399 --> 00:34:19.541
Since packers and packagers spend 97 percent of their time on tasks that robots can do and cost around 49000, dollarsrobots cost about 2500 dollarsless per year to do that work.

135
00:34:19.991 --> 00:34:46.011
Figure 7: Figure 7: Robot costs for physical occupations Employment and total compensation use data from the BLS (OEWS 2025, ECEC 2026). Exposed tasks are the share of that occupation’s tasks at exposure E1 to E3, weighted by estimated time spent on each task. Claude estimates annual costs for robots to perform exposed tasks for each occupation.

136
00:34:46.131 --> 00:34:56.754
Costs include annualized fixed costs and variable costs. The final column lists one robot included in each cost estimate and its estimated purchase price.

137
00:34:57.204 --> 00:35:09.189
The U S currently employs around 560000 packers and packagers. Firms face adoption frictions, like regulation or borrowing limits, and these cost estimates are approximate.

138
00:35:09.309 --> 00:35:30.668
For example, robotaxis are estimated to cost only around 7000 dollarsmore than taxi drivers, but face regulatory hurdles.⁴⁰ That said, automation risks are likely higher for physical jobs that cost-competitive robots can do. Employment for packers and packagers has fallen 22 percent since 2015.

139
00:35:30.788 --> 00:35:43.673
And the BLS projects this occupation will shed the 11th-most jobs of any occupation by 2035, behind only correctional officers and 9 sales and office occupations.⁴¹

140
00:35:44.123 --> 00:36:07.547
For the other jobs shown in Figure 7, robots cost much more than humans even though they can do much of the work. In metalworking, robots with cameras and A I can weld autonomously.⁴² But human welders perform many tasks that enable welding, like positioning large metal parts, climbing ladders to weld hard-to-reach joints, checking quality, and grinding and finishing materials.

141
00:36:07.667 --> 00:36:13.640
Automating this work requires robots that together cost around five times more than human welders.

142
00:36:14.090 --> 00:36:34.503
Dishwashers, along with janitors and cleaners, are paid 25000 dollarsto 30000 dollarsless than welders. But their robot analogs are still several times more expensive. Cleaning is hard to standardize and streamline, and robot cleaners are often slower than humans and limited to specific tasks.⁴³

143
00:36:34.953 --> 00:36:52.852
Figure 8 summarizes robot costs by plotting exposure to cost-competitive robots at hypothetical, uniform declines in robot costs. For example, if robots cost 20 percent less today, they would be cost-competitive for the physical work done by 2.8 million workers.

144
00:36:52.972 --> 00:37:16.111
These workers spend on average 42 percent of their time on work exposed to robots, 0.8 percent of all working time in the economy. For reference, industry and government data suggest that robot prices have declined roughly 3 percent per year since the 1990s. At this rate, it would take around seven years to reach a 20 percent cost decline.

145
00:37:16.231 --> 00:37:20.705
Appendix E discusses historical robot price data in more detail.

146
00:37:21.155 --> 00:37:30.246
Though robots can theoretically do tasks that add up to 34 percent of all work time, they are cost-competitive for just 0.3 percent today.

147
00:37:30.366 --> 00:37:47.869
Still, that includes roughly 300000 workers for whom robots can do 95 percent of their tasks.⁴⁴ As Figure 7 suggested, robots today are cost-competitive or close to it for packers and packagers and taxi drivers, but far from it for most other jobs.

148
00:37:48.319 --> 00:38:09.139
For robots to be cost-competitive for 10 percent of human work today, costs would need to decline about 70 percent. At a 3 percent decline per year, that would take around 40 years. That said, it’s possible that new, more capable robots like humanoids or new manufacturing processes will drive down costs to automate human work.

149
00:38:09.259 --> 00:38:16.442
Reports suggest that global production could scale quickly if there were intense demand for humanoids and other robots.⁴⁵

150
00:38:16.892 --> 00:38:39.838
Figure 8: Figure 8: Work exposed under hypothetical robot cost declines Calculated by comparing robot and labor costs. Labor costs are the fraction of time spent on exposed tasks multiplied by occupation total compensation. Given robot cost declines, the curve plots the share of all job tasks for which robots cost less than labor.

151
00:38:39.958 --> 00:39:02.625
Tasks are weighted by estimated time demands and occupation employment. Robot and labor costs are compared at six points in the worker pay distribution; see Appendix D.2 for details. Selected occupations are marked where their robot costs fall below their median labor cost. For example, robots do 94 percent of Postal Service mail carrier tasks.

152
00:39:02.745 --> 00:39:34.159
To match what workers produce in a year would cost 166000 dollarswith robots. Median total compensation for Postal Service mail carriers is 82000, dollarsso robots would be cost-competitive if cost declined 54 percent (166000 dollarsx 46 percent< 82000 dollarsx 94 percent). Total compensation is calculated with data from the BLS (OEWS 2025, ECEC 2026).

153
00:39:34.609 --> 00:39:52.033
Cost parity between humans and robots also needn’t imply high unemployment or rapid economic growth. Tasks that workers still perform may become bottlenecks, and these robot cost estimates often include human supervisory, exception-handling, or repair work.

154
00:39:52.153 --> 00:40:06.654
As emphasized by Acemoglu and Restrepo (2019) and Jones and Tonetti (2026), automation boosts productivity only after machines are much cheaper than humans, rather than at cost parity.

155
00:40:07.104 --> 00:40:29.355
We apply these cost estimates in two extensions. Appendix D.2 calculates an alternative occupation exposure measure that factors in both what robots can do and how cheaply they can do it. Accounting for cost produces a very similar ranking of jobs compared to our main, capabilities-focused measure (0.95 rank correlation).

156
00:40:29.805 --> 00:40:47.853
Appendix E also presents stylized scenarios for robot adoption if robots become cheaper and more productive. The scenarios draw on episodes of innovation and adoption for industrial robots since the 1990s, as well as analyst forecasts for humanoids.

157
00:40:47.973 --> 00:40:57.061
We hold fixed tasks and wages and ask how soon historical rates of cost declines and capability gains would make robots cost-competitive.

158
00:40:57.511 --> 00:41:19.671
To model how robots take on new work over time, we rate exposure for 1977 job tasks using today’s robots, and compare with robot exposure as of 1977. In 1977, robots could not do 62 percent of physical tasks. Today’s robots can do all but 24 percent of those same tasks.

159
00:41:19.791 --> 00:41:27.683
Over time, that means each year robots became able to do 2 percent of the tasks they previously could not (see Appendix B.4).

160
00:41:28.133 --> 00:41:40.783
Adding in 3 percent cost declines per year, robots aren’t cost-competitive for half of physical work today until 2085. But rapid advances in robotics may speed up these timelines.

161
00:41:40.903 --> 00:42:00.125
In a fast adoption scenario, where quality-adjusted costs fall up to four times faster and robots become able to do new tasks twice as fast, robots become cost-competitive for half of physical work by 2050. Automating 90 percent of physical work today still takes 53 years.

162
00:42:00.575 --> 00:42:21.055
One caveat is that our scenarios apply the same cost decline or capability increase to every task, so tasks become cost-competitive at today’s ranking of costs relative to worker pay. That ranking captures broad economic incentives, but robotics companies may also target certain capabilities first for other reasons.

163
00:42:21.175 --> 00:42:34.721
Humanoid demonstrations sometimes showcase household chores, for example, while robot makers may prioritize factory work useful to their own operations, much as A I companies prioritized coding agents.⁴⁶

164
00:42:35.171 --> 00:42:56.182
Overall, robots would need to sustain record rates of price declines and quality improvements over the coming decades to enable rapid physical automation. Even so, job impacts are less certain with these cost projections, since they set aside forces like preferences and regulation as well as feedback effects like falling wages.

165
00:42:56.982 --> 00:42:57.769
Discussion.

166
00:42:58.219 --> 00:43:13.598
We introduce a new measure of job exposure to robots. A job is exposed when robots can perform its tasks. Robots that work like humans count more toward exposure than those that require rebuilding the workplace, like factory robots.

167
00:43:14.048 --> 00:43:37.424
We find that driving jobs are most exposed, due to autonomous vehicles. Most robots today aren’t like driverless cars, and only operate in controlled environments. But robots could perform most physical work in some capacity, though they are several times more expensive than humans for the same jobs. Jobs exposed to robots are quite different from those exposed to L L Ms.

168
00:43:37.544 --> 00:43:40.057
They pay less, and are more physically demanding.

169
00:43:40.507 --> 00:43:56.913
We hope these measures help analyze physical automation as A I expands what robots can do. Exposure patterns today may inform which jobs more capable robots will disrupt tomorrow. And periodic updates to this work can track robot capabilities.

170
00:43:57.363 --> 00:44:20.105
Our exposure scale is predicated on the idea that jobs are more at risk when today’s robots already do them in some settings. We find that this way of assessing robots predicts historical labor market impacts. But A I-powered robots could leapfrog our scale and do work they cannot today, for example by learning to climb ladders or use their arms and grippers more deftly.

171
00:44:20.555 --> 00:44:43.288
Much remains uncertain about how robots and A I could reshape the economy. Growth forecasts for A I often assume that physical bottlenecks, which robots could loosen, will drag down gains from automating cognitive work. We also have not considered how A I could affect physical work without robots, for example by better predicting when factory machines need maintenance.

172
00:44:43.408 --> 00:45:01.775
In the nearer term, efforts to track A I’s impacts could watch exposed occupations, like drivers and warehouse packers, for early signs of disruption. Further out, a key question is how work itself will change, perhaps becoming more social and interpersonal as robots and A I advance.

173
00:45:02.575 --> 00:45:03.369
Appendix.

174
00:45:03.819 --> 00:45:04.780
Available here.

175
00:45:05.580 --> 00:45:06.305
Authors.

176
00:45:06.755 --> 00:45:09.414
Russell Legate-Yang and Maxim Massenkoff.

177
00:45:10.214 --> 00:45:11.191
Acknowledgments.

178
00:45:11.641 --> 00:45:32.206
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.

179
00:45:33.006 --> 00:45:33.869
Citation.

180
00:45:34.319 --> 00:45:35.470
Please cite this as:

181
00:45:35.920 --> 00:45:40.383
A code block of 7 lines appears here; see the original post for the code.

182
00:45:41.183 --> 00:45:42.013
References.

183
00:45:42.463 --> 00:46:00.057
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 to 143.

184
00:46:00.507 --> 00:46:17.483
Acemoglu, Daron and David Autor, "Skills, Tasks and Technologies: Implications for Employment and Earnings," in Handbook of Labor Economics, Vol. 4 billion, Elsevier, 2011, 1043 to 1171.

185
00:46:17.933 --> 00:46:36.396
Acemoglu, Daron, David Autor, David Dorn, Gordon H. Hanson, and Brendan Price, "Import Competition and the Great U S Employment Sag of the 2000s," Journal of Labor Economics, 2016, 34 (S1), S141-S198.

186
00:46:36.846 --> 00:46:50.456
Acemoglu, Daron, Andrea Manera, and Pascual Restrepo, "Does the U S Tax Code Favor Automation?," Brookings Papers on Economic Activity, Spring 2020, 231 to 300.

187
00:46:50.906 --> 00:47:09.456
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 to 1542.

188
00:47:09.906 --> 00:47:23.645
Acemoglu, Daron and Pascual Restrepo, "Automation and New Tasks: How Technology Displaces and Reinstates Labor," Journal of Economic Perspectives, 2019, 33 (2), 3 to 30.

189
00:47:24.095 --> 00:47:40.383
Acemoglu, Daron and Pascual Restrepo, "Robots and Jobs: Evidence from U S Labor Markets," Journal of Political Economy, 2020, 128 (6), 2188 to 2244.

190
00:47:40.833 --> 00:47:54.493
Acemoglu, Daron and Pascual Restrepo, "Tasks, Automation, and the Rise in U S Wage Inequality," Econometrica, 2022, 90 (5), 1973 to 2016.

191
00:47:54.943 --> 00:48:16.914
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.

192
00:48:17.034 --> 00:48:28.037
Durham, "DeepFleet: Multi-Agent Foundation Models for Mobile Robots," arXiv:2508.08574, 2025.

193
00:48:28.487 --> 00:48:49.483
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 to 282.

194
00:48:49.933 --> 00:48:58.583
Airbus, "Airbus Concludes ATTOL with Fully Autonomous Flight Tests," press release, June 29, 2020. airbus dot com

195
00:48:59.033 --> 00:49:07.958
Airbus, "Airbus Tests New Technologies to Enhance Pilot Assistance," press release, January 12, 2023. airbus dot com

196
00:49:08.408 --> 00:49:18.602
Amazon, "Introducing Vulcan: Amazon's First Robot with a Sense of Touch," About Amazon, May 7, 2025. aboutamazon dot com

197
00:49:19.052 --> 00:49:29.651
Andreassen, Dag, "A Robotic Success from Jæren," The Norwegian Museum of Science and Technology, September 3, 2024. tekniskmuseum dot no

198
00:49:30.101 --> 00:49:53.293
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 to 20.

199
00:49:53.743 --> 00:50:14.160
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.

200
00:50:14.610 --> 00:50:29.436
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 to 34.

201
00:50:29.886 --> 00:50:42.211
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 to 30.

202
00:50:42.661 --> 00:51:02.864
Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller, "New Frontiers: The Origins and Content of New Work, 1940 to 2018," Quarterly Journal of Economics, 2024, 139 (3), 1399 to 1465.

203
00:51:03.314 --> 00:51:21.182
Autor, David H. and David Dorn, "The Growth of Low-Skill Service Jobs and the Polarization of the U S Labor Market," American Economic Review, 2013, 103 (5), 1553 to 1597.

204
00:51:21.632 --> 00:51:40.347
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 to 2168.

205
00:51:40.797 --> 00:51:58.539
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 to 1213.

206
00:51:58.989 --> 00:52:17.190
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 to 1333.

207
00:52:17.640 --> 00:52:30.139
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.

208
00:52:30.589 --> 00:52:44.626
Autor, David H. and Neil Thompson, "Expertise," Journal of the European Economic Association, 2025, 23 (4), 1203 to 1271.

209
00:52:45.076 --> 00:53:01.693
Badger Technologies, "Badger Technologies and Stop and Shop Stores Upgrade Marty the Robot at 300+ Locations to Improve Customer Shopping Experiences," press release, September 18, 2023. badger-technologies dot com

210
00:53:02.143 --> 00:53:09.908
Bank of America, "Humanoid Robots 101," BofA Global Research, April 29, 2025.

211
00:53:10.358 --> 00:53:27.123
Bank of Japan, "Producer Price Index, 2020 Base: Robots [PRCG20 2201350029]," Time-Series Data Search, 2026. stat-search dot boj dot or dot jp

212
00:53:27.573 --> 00:53:39.998
Baptista, Eduardo, Ju-min Park, and Laurie Chen, "Robots Poised for 'ChatGPT Moment,' Unitree C E O Says," Reuters, August 20, 2026. reuters dot com

213
00:53:40.448 --> 00:53:54.641
Billard, Aude and Danica Kragic, "Trends and Challenges in Robot Manipulation," Science, 2019, 364 (6446), eaat8414.

214
00:53:55.091 --> 00:54:04.824
Binder, Damon, "A G I Will Set Off an Industrial Explosion," A I Frontiers, August 11, 2026. ai-frontiers dot org

215
00:54:05.274 --> 00:54:14.120
Bishop, Todd, "Amazon Tops 1 Million Robots: Here's What They Do," GeekWire, July 1, 2025. geekwire dot com

216
00:54:14.570 --> 00:54:29.234
BMW Group, "BMW Group Advances the Use of Physical A I in Production with Figure 03 Project in Spartanburg," press release, June 25, 2026. press dot bmwgroup dot com

217
00:54:29.684 --> 00:54:39.806
Boston Consulting Group, "The Robotics Revolution: The Next Great Leap in Manufacturing," report, September 2015. bcg dot com

218
00:54:40.256 --> 00:54:46.506
Boston Dynamics, "Stretch," product page, 2026. bostondynamics dot com

219
00:54:46.956 --> 00:54:56.656
Brooks, Rodney, "Why Today's Humanoids Won't Learn Dexterity," rodneybrooks.com, September 26, 2025. rodneybrooks dot com

220
00:54:57.106 --> 00:55:06.642
Built Robotics, "The World's First Fully Autonomous Trenching Solution," Exosystem product page, 2026. builtrobotics dot com

221
00:55:07.092 --> 00:55:14.998
Canvas, "Drywall Finishing Robot," product page, JLG Industries, 2026. jlg dot com

222
00:55:15.448 --> 00:55:36.876
Cazzaniga, Mauro, Florence Jaumotte, Longji Li, Giovanni Melina, Augustus J. Panton, Carlo Pizzinelli, Emma J. Rockall, and Marina Mendes Tavares, "Gen-A I: Artificial Intelligence and the Future of Work," IMF Staff Discussion Note SDN/2024/001, 2024.

223
00:55:37.326 --> 00:55:57.535
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.

224
00:55:57.985 --> 00:56:12.577
Chen, Jiafeng and Jonathan Roth, "Logs with Zeros? Some Problems and Solutions," Quarterly Journal of Economics, 2024, 139 (2), 891 to 936.

225
00:56:13.027 --> 00:56:23.338
Cleveland Clinic, "da Vinci Robotic Surgery," Health Library, last updated April 29, 2025. my dot clevelandclinic dot org

226
00:56:23.788 --> 00:56:46.636
Cockburn, Iain M., Rebecca Henderson, and Scott Stern, "The Impact of Artificial Intelligence on Innovation: An Exploratory Analysis," in Ajay Agrawal, Joshua Gans, and Avi Goldfarb, eds., The Economics of Artificial Intelligence: An Agenda, University of Chicago Press, 2019, 115 to 146.

227
00:56:47.086 --> 00:56:59.356
Code of Federal Regulations, "Composition of Flight Crew," 14 CFR 121.385, retrieved September 21, 2026. ecfr dot gov

228
00:56:59.806 --> 00:57:20.273
CORDIS, "Automated System for Packaging Fresh Meat with Reduced Waste/Giveaway, Processing Time, Human Involvement and Contamination (AUTOMEATIC)," European Commission, grant agreement 775913, periodic reporting summary, 2018. cordis dot europa dot eu

229
00:57:20.723 --> 00:57:36.014
Counterpoint Research, "Global Humanoid Robot Shipments Soar Nearly 300 percent YoY in H1 2026, Driven by Commercial Deployments," August 20, 2026. counterpointresearch dot com

230
00:57:36.464 --> 00:57:44.720
Davidson, Tom and Rose Hadshar, "The Industrial Explosion," Forethought Research, May 2025. forethought dot org

231
00:57:45.170 --> 00:57:58.620
Davidson, Tom, Basil Halperin, Thomas Houlden, and Anton Korinek, "When Does Automating Research Produce Explosive Growth?," EconTAI Working Paper, January 2026. econtai dot org

232
00:57:59.070 --> 00:58:13.987
Deming, David J., "The Growing Importance of Social Skills in the Labor Market," Quarterly Journal of Economics, 2017, 132 (4), 1593 to 1640.

233
00:58:14.437 --> 00:58:24.109
Denain, Jean-Stanislas and Yann Rivière, "How Fast Could Robot Production Scale Up?," Epoch A I, April 2026. epoch dot ai

234
00:58:24.559 --> 00:58:34.465
Drankhan, Holly, "Developments in Robotic Teat Sprayers Cut Costs and Labor," Progressive Dairy, September 30, 2016. agproud dot com

235
00:58:34.915 --> 00:58:43.937
Dürr Group, "See and Drive Perfectly with x-light and x-wheel," duerrmore magazine, November 13, 2018. durr-group dot com

236
00:58:44.387 --> 00:59:03.130
Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, "GPTs are GPTs: Labor Market Impact Potential of L L Ms," Science, 2024, 384 (6702), 1306 to 1308.

237
00:59:03.580 --> 00:59:13.930
Epiroc, "Autonomous Material Handling: A Crucial Step Toward the Mine of the Future," customer story, April 17, 2024. epiroc dot com

238
00:59:14.380 --> 00:59:24.250
Ewing, Jack, "Robots That Walk and Talk Are Coming to Car Factories," The New York Times, August 11, 2026. nytimes dot com

239
00:59:24.700 --> 00:59:35.788
Fabrico, "Maintenance Cost as Percent of RAV: The Benchmark That Cuts Through Excuses," blog, June 26, 2026. fabrico dot io

240
00:59:36.238 --> 00:59:44.871
Federal Aviation Administration (FAA), "Safety Framework for Aircraft Automation," September 2025. faa dot gov

241
00:59:45.321 --> 00:59:55.102
Fine, Sidney A., "A Structure of Worker Functions," Personnel and Guidance Journal, 1955, 34 (2), 66 to 73.

242
00:59:55.552 --> 01:00:07.708
Fine, Sidney A. and Wretha W. Wiley, An Introduction to Functional Job Analysis, Kalamazoo, MI: W.E. Upjohn Institute for Employment Research, 1971.

243
01:00:08.158 --> 01:00:33.039
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 to 739.

244
01:00:33.489 --> 01:00:39.936
Fori Automation, "Toe Tool Systems," product page, 2025. foriauto dot com

245
01:00:40.386 --> 01:00:56.842
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 to 280.

246
01:00:57.292 --> 01:01:14.476
Friedel, William C., Jr., Thomas S. Bettencourt, and Daniel L. Freeman, "Tomato Harvesting Employing Electronic Sorting," U S Patent 4033099, issued July 5, 1977. patents dot google dot com

247
01:01:14.926 --> 01:01:30.444
Friedel, William C., Jr. and Charles F. Dietz, "Tomato Harvesting Sorting System," U S Patent 4118311, issued October 3, 1978. patents dot google dot com

248
01:01:30.894 --> 01:01:45.362
Goldman, Sharon, "A Year Ago, Nvidia's Jensen Huang Said the 'ChatGPT Moment' for Robotics Was Around the Corner. Now He Says It's 'Nearly Here.' But Is It?," Fortune, January 6, 2026. fortune dot com

249
01:01:45.812 --> 01:02:01.779
González Huarte, Javier, Maite Ortiz de Zarate, and Aitor Ibarguren, "CAD-Based Robot Programming Solution for Wire Harness Manufacturing in Aeronautic Sector," Robotics, 2023, 12 (5), 130.

250
01:02:02.229 --> 01:02:14.098
Graetz, Georg and Guy Michaels, "Robots at Work," Review of Economics and Statistics, 2018, 100 (5), 753 to 768.

251
01:02:14.548 --> 01:02:22.974
Gridwise, "How Are Robotaxis Affecting Driver Pay in AV Cities?," blog, August 2025. gridwise dot io

252
01:02:23.424 --> 01:02:36.396
Hirsch, Barry T., David A. Macpherson, and William E. Even, "Union Membership, Coverage, and Earnings from the CPS" [dataset], Unionstats.com, 2026. unionstats dot com

253
01:02:36.846 --> 01:02:49.923
Honerkamp, Paxton, "Within Uber-Waymo Split, a Key Labor Battle over AVs Is Being Waged in the Nation's Capital," CNBC, August 1, 2026. cnbc dot com

254
01:02:50.373 --> 01:03:01.792
Huffman, Wallace E., "The Status of Labor-Saving Mechanization in U S Fruit and Vegetable Harvesting," Choices, 2012, 27 (2). choicesmagazine dot org

255
01:03:02.242 --> 01:03:12.386
Humlum, Anders, "Robot Adoption and Labor Market Dynamics," Rockwool Foundation Research Unit Study Paper 175, 2021.

256
01:03:12.836 --> 01:03:33.167
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, and Mingxing Tan, "EMMA: End-to-End Multimodal Model for Autonomous Driving," Transactions on Machine Learning Research, 2025.

257
01:03:33.617 --> 01:03:53.666
Ibarz, Julian, Jie Tan, Chelsea Finn, Mrinal Kalakrishnan, Peter Pastor, and Sergey Levine, "How to Train Your Robot with Deep Reinforcement Learning: Lessons We Have Learned," International Journal of Robotics Research, 2021, 40 (4 to 5), 698 to 721.

258
01:03:54.116 --> 01:04:01.667
Ingenia, "Fruit-Picking Robots," Issue 93, December 2022. ingenia dot org dot uk

259
01:04:02.117 --> 01:04:14.055
International Organization for Standardization (ISO), "Robotics, Vocabulary," ISO 8373:2021, Geneva, 2021.

260
01:04:14.505 --> 01:04:26.188
Jarrett, Cosette, "12 Cool Facts About the A I-Powered Robots That Help Deliver Your Amazon Packages," About Amazon, December 11, 2023. aboutamazon dot com

261
01:04:26.638 --> 01:04:40.675
Jones, Charles I. and Christopher Tonetti, "Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion," working paper, Stanford University, 2026. web dot stanford dot edu

262
01:04:41.125 --> 01:05:00.366
Jurkat, Anne, Rainer Klump, and Florian Schneider, "Tracking the Rise of Robots: A Survey of the IFR Database and its Applications," MPRA Paper 111812, Munich Personal RePEc Archive, October 2021. mpra dot ub dot uni-muenchen dot de

263
01:05:00.816 --> 01:05:17.776
Katz, Lawrence F. and Kevin M. Murphy, "Changes in Relative Wages, 1963 to 1987: Supply and Demand Factors," Quarterly Journal of Economics, 1992, 107 (1), 35 to 78.

264
01:05:18.226 --> 01:05:32.098
Kawasaki Robotics, "At the Forefront of FRP Automation: Spray-Up and Cutting Process Innovations Pioneered by KADO times IEC," case study, 2026. kawasakirobotics dot com

265
01:05:32.548 --> 01:05:47.261
Kemp, Charles C., Aaron Edsinger, and Eduardo Torres-Jara, "Challenges for Robot Manipulation in Human Environments," IEEE Robotics and Automation Magazine, 2007, 14 (1), 20 to 29.

266
01:05:47.711 --> 01:05:57.483
Kim, Moo Jin, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan P. Foster, Pannag R.

267
01:05:57.603 --> 01:06:22.891
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 to 2713.

268
01:06:23.341 --> 01:06:40.408
Kober, Jens, J. Andrew Bagnell, and Jan Peters, "Reinforcement Learning in Robotics: A Survey," International Journal of Robotics Research, 2013, 32 (11), 1238 to 1274.

269
01:06:40.858 --> 01:06:58.119
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.

270
01:06:58.569 --> 01:07:14.665
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 to 82.

271
01:07:15.115 --> 01:07:35.430
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 to 005, National Institute of Standards and Technology, August 2016. doi dot org

272
01:07:35.880 --> 01:07:52.499
Lise, Jeremy and Fabien Postel-Vinay, "Multidimensional Skills, Sorting, and Human Capital Accumulation," American Economic Review, 2020, 110 (8), 2328 to 2376.

273
01:07:52.949 --> 01:08:06.656
Ludwig, Jens, Sendhil Mullainathan, and Ashesh Rambachan, "Large Language Models: An Applied Econometric Framework," NBER Working Paper 33344, 2025.

274
01:08:07.106 --> 01:08:17.581
Lung, Natalie, "New York's Robotaxi Plan Pulled in Setback to Waymo Expansion," Bloomberg News, February 19, 2026. bloomberg dot com

275
01:08:18.031 --> 01:08:28.944
Madrigal, Alexis C., "Inside Waymo's Secret World for Training Self-Driving Cars," The Atlantic, August 23, 2017. theatlantic dot com

276
01:08:29.394 --> 01:08:40.346
Mason, Matthew T., "Toward Robotic Manipulation," Annual Review of Control, Robotics, and Autonomous Systems, 2018, 1, 1 to 28.

277
01:08:40.796 --> 01:08:52.338
Massenkoff, Maxim and Peter McCrory, "Labor Market Impacts of A I: A New Measure and Early Evidence," Anthropic, March 5, 2026. anthropic dot com

278
01:08:52.788 --> 01:08:58.888
Mayekawa, "Deboning Machines," product page, 2026a. mayekawa dot com

279
01:08:59.338 --> 01:09:16.338
Mayekawa, "Reducing Workloads with Reliable Deboning Technology and Stabilizing Production to Meet Market Demands for Pork," With Mayekawa, 2026b (undated; retrieved September 29, 2026). mayekawa dot com

280
01:09:16.788 --> 01:09:26.641
McKinsey and Company, "The Future of Robotics: Intelligent, Adaptable, and on Your Team," The Next Normal, 2026. mckinsey dot com

281
01:09:27.091 --> 01:09:51.990
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.

282
01:09:52.440 --> 01:10:02.013
Moravec, Hans, Mind Children: The Future of Robot and Human Intelligence, Cambridge, MA: Harvard University Press, 1988.

283
01:10:02.463 --> 01:10:13.457
Morgan Stanley, "Humanoid Robot Market Expected to Reach 5 dollarsTrillion by 2050," May 14, 2025. morganstanley dot com

284
01:10:13.907 --> 01:10:22.574
Morgan Stanley Research, "Humanoids: Investment Implications of Embodied A I," report, June 26, 2024.

285
01:10:23.024 --> 01:10:39.936
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 and D," Robotics, 2022, 11 (3), 65.

286
01:10:40.386 --> 01:11:07.422
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. news dot nissyoku dot co dot jp

287
01:11:07.872 --> 01:11:18.278
Oitzman, Mike, "Pharm Robotics Advances Automated Dairy Cow Healthcare," The Robot Report, July 13, 2025. therobotreport dot com

288
01:11:18.728 --> 01:11:28.503
Oxmaint, "Food Manufacturing Maintenance Budgeting: Planning, Benchmarks, and Cost Control," March 28, 2026. oxmaint dot com

289
01:11:28.953 --> 01:11:45.091
Pallottino, Lucia, "Robotics for Warehouses and Logistics: Technologies, Challenges, and Future Directions," Annual Review of Control, Robotics, and Autonomous Systems, 2026, 9, 377 to 401.

290
01:11:45.541 --> 01:11:51.663
Polanyi, Michael, The Tacit Dimension, Garden City, NY: Doubleday, 1966.

291
01:11:52.113 --> 01:12:11.226
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.

292
01:12:11.676 --> 01:12:21.894
Reuters, "Meet SwagBot, the A I-Powered Robot Cattle Herder Preventing Soil Degradation," Rappler, December 12, 2024. rappler dot com

293
01:12:22.344 --> 01:12:34.467
Rivière, Yann and Jean-Stanislas Denain, "Where Autonomy Works: Evaluating Robot Capabilities in 2026," Epoch A I, 2026. epoch dot ai

294
01:12:34.917 --> 01:12:49.468
Robbins, Jacob, "Apptronik Raises 520 million dollars as VC Funding for Humanoid Robotics Explodes 300 percent," PitchBook News, February 11, 2026. pitchbook dot com

295
01:12:49.918 --> 01:13:00.023
The Robot Report, "Path Robotics Raises 100 million dollars to Automate Welding," October 28, 2024. therobotreport dot com

296
01:13:00.473 --> 01:13:11.900
The Robot Report, "Okibo Opens New Jersey HQ and Launches EG7 Robot in U S Market," April 11, 2025. therobotreport dot com

297
01:13:12.350 --> 01:13:26.627
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 to 22.

298
01:13:27.077 --> 01:13:43.772
SAE International, "Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles," SAE Recommended Practice J3016 202104, April 2021.

299
01:13:44.222 --> 01:13:56.187
Sandvik, "Sandvik Introduces AutoMine AutoLoad 2.0 for Improved Autonomous Bucket Loading," press release, June 14, 2023. mining dot sandvik

300
01:13:56.637 --> 01:14:05.460
Schenck RoTec, "Wheel Assembly Lines: Individual and Complete Solutions," product page, 2025. schenck-rotec dot com

301
01:14:05.910 --> 01:14:17.176
Scoop, "‘Golden Opportunity’ for Scott Technology with Silver Fern Farms," Scott Technology announcement, September 14, 2022. scoop dot co dot nz

302
01:14:17.626 --> 01:14:23.430
Simbe Robotics, "Tally," product page, 2026. simberobotics dot com

303
01:14:23.880 --> 01:14:36.909
Statista, "Average Retail Electricity Price for Industrial Consumers in the United States from 1970 to 2025," April 28, 2026. statista dot com

304
01:14:37.359 --> 01:15:00.175
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.

305
01:15:00.625 --> 01:15:18.331
Svanberg, Maja, Wensu Li, Martin Fleming, Brian Goehring, and Neil Thompson, "Beyond A I Exposure: Which Tasks Are Cost-Effective to Automate with Computer Vision?," MIT FutureTech working paper, January 2024. futuretech dot mit dot edu

306
01:15:18.781 --> 01:15:32.206
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.

307
01:15:32.326 --> 01:15:50.064
Sumers, Jared Mueller, William McEachen, Wes Mitchell, Shan Carter, Jack Clark, Jared Kaplan, and Deep Ganguli, "Clio: Privacy-Preserving Insights into Real-World A I Use," arXiv:2412.13678, 2024.

308
01:15:50.514 --> 01:16:00.675
Tamkin, Alex and Peter McCrory, "Estimating A I Productivity Gains from Claude Conversations," Anthropic, 2025. anthropic dot com

309
01:16:01.125 --> 01:16:13.372
Tedrake, Russ, Robotic Manipulation: Perception, Planning, and Control, course notes for MIT 6.421, 2026. manipulation dot mit dot edu

310
01:16:13.822 --> 01:16:26.315
Telexistence, "Telexistence to Begin Installing A I Robot 'TX SCARA' in 300 FamilyMart Stores," press release, August 10, 2022. tx-inc dot com

311
01:16:26.765 --> 01:16:34.121
Thomas, Charlotte, "Rise of the Robots," METSTRADE, October 23, 2023. metstrade dot com

312
01:16:34.571 --> 01:16:46.961
Trammell, Philip and Anton Korinek, "Economic Growth Under Transformative A I," Annual Review of Economics, 2026, 18, 589 to 611.

313
01:16:47.411 --> 01:16:55.815
Transformation Hub Wiring Harness, "Robotics Challenge 2026," 2026. leitungssatz-hub dot de

314
01:16:56.265 --> 01:17:04.515
UBS, "Is the World Ready for 1bn Robots?," UBS Global Research Q-Series, June 17, 2025.

315
01:17:04.965 --> 01:17:18.764
University of Michigan, "U-M Opens Mcity Test Environment for Connected and Driverless Vehicles," Michigan Engineering News, July 20, 2015. news dot engin dot umich dot edu

316
01:17:19.214 --> 01:17:41.399
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. fred dot stlouisfed dot org

317
01:17:41.849 --> 01:17:58.297
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. bls dot gov

318
01:17:58.747 --> 01:18:13.760
U S Bureau of Labor Statistics (BLS), "Occupational Employment and Wage Statistics (OEWS): May 2025 National Occupational Employment and Wage Estimates," 2025. bls dot gov

319
01:18:14.210 --> 01:18:30.658
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. bls dot gov

320
01:18:31.108 --> 01:18:52.451
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. fred dot stlouisfed dot org

321
01:18:52.901 --> 01:19:07.685
U S Bureau of Labor Statistics (BLS), "Occupational Employment and Wage Statistics: Frequently Asked Questions," 2026c, retrieved September 21, 2026. bls dot gov

322
01:19:08.135 --> 01:19:20.082
U S Census Bureau, "Annual Capital Expenditures Survey: Robotic Equipment Expenditures, 2019 to 2022," 2024. census dot gov

323
01:19:20.532 --> 01:19:39.725
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 to 013, January 7, 2016. oig dot dot dot gov

324
01:19:40.175 --> 01:19:56.263
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 to 40.

325
01:19:56.713 --> 01:20:07.322
Wallace, Neal, "Butcher Bot Pays Off in Dollars per Carcase," Farmers Weekly (New Zealand), September 30, 2022. farmersweekly dot co dot nz

326
01:20:07.772 --> 01:20:18.266
Wallace, Neal, "One Giant LEAP for Robo Meatworks," Farmers Weekly (New Zealand), April 22, 2024. farmersweekly dot co dot nz

327
01:20:18.716 --> 01:20:30.703
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.

328
01:20:30.823 --> 01:20:55.044
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 to 60.

329
01:20:55.494 --> 01:21:08.891
Waste360, "Q and A: AMP's Matanya Horowitz on Bringing A I to the Recycling World and How the Tech Is Advancing," November 2023. waste360 dot com

330
01:21:09.341 --> 01:21:17.341
WATTAgNet, "The New Generation of Automated Deboning Machines," October 1, 2018. wattagnet dot com

331
01:21:17.791 --> 01:21:26.039
Webb, Michael, "The Impact of Artificial Intelligence on the Labor Market," working paper, Stanford University, 2020.

332
01:21:26.489 --> 01:21:38.764
Wessling, Brianna, "With New Funding, Monumental Plans to Bring Its Construction Robots to the U S," The Robot Report, July 17, 2026. therobotreport dot com

333
01:21:39.214 --> 01:21:50.457
Wishart-Smith, Heather, "Built Robotics Unveils Autonomous Pile Driving Robot, Expediting Solar Rollout," Forbes, March 20, 2023. forbes dot com

334
01:21:50.907 --> 01:22:11.954
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 A I," NBER Working Paper 34836, 2026. doi dot org

335
01:22:12.404 --> 01:22:27.671
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.

336
01:22:28.121 --> 01:22:41.063
Zeira, Joseph, "Workers, Machines, and Economic Growth," Quarterly Journal of Economics, 1998, 113 (4), 1091 to 1117.

337
01:22:41.513 --> 01:22:52.433
Ziemerink, Jan, "How to Manage Automatic Calf Feeders Successfully," Förster-Technik North America presentation, Vita Plus, 2018. vitaplus dot com

338
01:22:53.233 --> 01:22:54.278
Data sources.

339
01:22:54.728 --> 01:23:08.963
Althobaiti, Shahad, Ahmad Alabdulkareem, Judy Hanwen Shen, Iyad Rahwan, Morgan Frank, Esteban Moro, and Alex Rutherford. "Dictionary of Titles" [dataset]. Harvard Dataverse, 2022. doi dot org

340
01:23:09.413 --> 01:23:29.978
Autor, David, Caroline Chin, Anna Salomons, and Bryan Seegmiller. Replication Data for: "New Frontiers: The Origins and Content of New Work, 1940 to 2018." Harvard Dataverse, V1.1, 2024. doi dot org, retrieved July 30 to 31, 2026.

341
01:23:30.428 --> 01:23:53.694
Autor, David H. and David Dorn. 2013. "The Growth of Low-Skill Service Jobs and the Polarization of the U S Labor Market." American Economic Review 103 (5): 1553 to 1597. Occupation crosswalks (occ1990dd) and offshorability measure: ddorn dot net.

342
01:23:54.144 --> 01:24:13.237
Bank of Japan. Producer Price Index, 2020 Base: Robots [PRCG20 2201350029]. Time-Series Data Search, retrieved September 16, 2026. stat-search dot boj dot or dot jp

343
01:24:13.687 --> 01:24:35.436
Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2024. "GPTs are GPTs: Labor Market Impact Potential of L L Ms." Science 384 (6702): 1306 to 1308. Task-level exposure labels: github dot com.

344
01:24:35.886 --> 01:25:06.301
Hirsch, Barry T., David A. Macpherson, and William E. Even. Union Membership, Coverage, and Earnings from the CPS [dataset]. Unionstats.com, 2026. unionstats dot 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 to 354.

345
01:25:06.751 --> 01:25:28.874
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.

346
01:25:28.994 --> 01:25:39.406
Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2006 to 09-06. doi dot org

347
01:25:39.856 --> 01:26:05.237
National Center for ONET Development. ONET Database, versions 29.3 (2025), 13.0 (2008), and 4.0 (2002); the ONET-SOC taxonomy crosswalks; and the DOT to ONET-SOC 2019 Crosswalk. Washington, D C: U S Department of Labor, Employment and Training Administration. onetcenter dot org; onetcenter dot org; onetcenter dot org

348
01:26:05.687 --> 01:26:25.341
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. doi dot org

349
01:26:25.791 --> 01:26:42.431
U S Bureau of Economic Analysis (BEA). Personal Consumption Expenditures: Chain-type Price Index [PCEPI], retrieved from FRED, Federal Reserve Bank of St. Louis. fred dot stlouisfed dot org, September 21, 2026.

350
01:26:42.881 --> 01:27:06.625
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. fred dot stlouisfed dot org, September 16, 2026.

351
01:27:07.075 --> 01:27:28.073
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, D C: U S Department of Labor. Retrieved September 16, 2026. bls dot gov

352
01:27:28.523 --> 01:27:44.692
U S Bureau of Labor Statistics (BLS). Employer Costs for Employee Compensation (ECEC), 2004 to 2023 and March 2026. Washington, D C: U S Department of Labor. bls dot gov

353
01:27:45.142 --> 01:28:04.459
U S Bureau of Labor Statistics (BLS). Employment Projections, 2025 to 35: Table 1.6, Occupations with the Largest Job Declines. Washington, D C: U S Department of Labor, August 27, 2026. bls dot gov

354
01:28:04.909 --> 01:28:25.269
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, D C: U S Department of Labor. bls dot gov

355
01:28:25.719 --> 01:28:39.663
U S Bureau of Labor Statistics (BLS). Occupational Requirements Survey (ORS), 2023 and 2025. Washington, D C: U S Department of Labor. bls dot gov

356
01:28:40.113 --> 01:29:00.254
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. www2 dot census dot gov; www2 dot census dot gov

357
01:29:00.704 --> 01:29:14.848
U S Census Bureau. Annual Capital Expenditures Survey: Capital Expenditures for Structures and Equipment, Tables 1a and 1b, 2019 to 2022 releases. census dot gov

358
01:29:15.298 --> 01:29:28.722
U S Department of Labor, Employment and Training Administration. Dictionary of Occupational Titles, Fourth Edition. Washington, D C: U S Government Printing Office, 1977.

359
01:29:29.172 --> 01:29:52.653
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 to 02-19. doi dot org.

360
01:29:52.773 --> 01:30:00.915
Missing text from the Department of Labor Office of Administrative Law Judges DOT pages and occupationalinfo dot org.

361
01:30:01.365 --> 01:30:21.878
Webb, Michael. 2020. "The Impact of Artificial Intelligence on the Labor Market." Working paper, Stanford University. Robot exposure percentiles and 2010 labor-supply weights: web dot stanford dot edu, retrieved July 18, 2026; now at michaelwebb dot co.

362
01:30:22.678 --> 01:30:23.426
Footnotes.

363
01:30:23.876 --> 01:30:47.169
1. See for example Aghion and others (2019), Davidson and others (2026), and Jones and Tonetti (2026) on the role of weak links and complementarities in explosive economic growth scenarios. A I is often thought to affect digital work more than physical work; see for example Cazzaniga and others (2024).

364
01:30:47.409 --> 01:31:03.222
2. See for example Autor (2015) on “environmental control” as a barrier to physical automation. See also Section 6 of Link and others (2016). We discuss this point in more detail in “Measuring exposure.”.

365
01:31:03.462 --> 01:31:25.045
3. See Firoozi and others (2025) on A I foundation models for robots. Rivière and Denain (2026) discuss robot capabilities afforded by A I. See Agaskar and others (2025) and Hwang and others (2025) on A I-powered robots in warehousing and driving.

366
01:31:25.165 --> 01:31:37.666
A I could also speed the pace of innovation in robotics; see Cockburn and others (2019), Wang and others (2023), and Davidson and others (2026).

367
01:31:37.906 --> 01:31:55.263
4. See Baptista and others (2026) and Goldman (2026) on a potential “ChatGPT moment” in robotics. See also forecasts from Morgan Stanley Research (2024) and McKinsey and Company (2026).

368
01:31:55.503 --> 01:32:20.250
5. Robot investment data from the 2022 Annual Capital Expenditures Survey (U S Census Bureau, 2024). Surveys of U S executives in November 2025 from Yotzov and others (2026) show that 13 percent of firms use robotics now while 22 percent expect to within three years of surveying, weighted by employment.

369
01:32:20.370 --> 01:32:25.232
Expected use is broadly similar in the United Kingdom, Germany, and Australia.

370
01:32:25.472 --> 01:32:37.449
6. See Robbins (2026) on investment in humanoid robotics startups and Ewing (2026) on efforts by large automakers to use humanoids.

371
01:32:37.689 --> 01:32:55.949
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 and others (2020) and Suresh and others (2024).

372
01:32:56.189 --> 01:33:12.521
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.

373
01:33:12.761 --> 01:33:37.930
9. Our taxonomy builds on a long tradition in labor economics and occupational analysis. The predecessor to ONET, the Dictionary of Occupational Titles, categorized jobs by “Data/People/Things.” See Fine (1955), Fine and Wiley (1971), Autor and others (2003), Acemoglu and Autor (2011), and Deming (2017).

374
01:33:38.170 --> 01:33:49.733
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.

375
01:33:49.853 --> 01:34:11.666
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).

376
01:34:11.906 --> 01:34:31.126
11. Our focus on what today’s robots can do is similar to Eloundou and others (2024), who measure A I exposure by what tasks L L Ms could speed up as of 2023. Other A I exposure measures are described in Massenkoff and McCrory (2026).

377
01:34:31.366 --> 01:34:56.554
12. See Autor (2015) on “environmental control” as a bottleneck to automation, and Autor and others (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 and others (2016).

378
01:34:56.674 --> 01:35:27.140
Kober and others (2013), Ibarz and others (2021), and Kim and others (2025) discuss how robot performance generalizes to new tasks and environments. See Firoozi and others (2025) on robotic foundation models versus specialized robot control systems. For surveys on progress in robotic manipulation, see Kemp and others (2007) and Kroemer and others (2021).

379
01:35:27.260 --> 01:35:39.847
See Pumacay and others (2024) for manipulation benchmarks. For an inventory of robot capabilities in complex environments, see Rivière and Denain (2026).

380
01:35:40.087 --> 01:35:50.044
13. Mason (2018) and Tedrake (2026), for example, discuss the complex skills robots need to manipulate objects.

381
01:35:50.284 --> 01:36:02.425
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.

382
01:36:02.665 --> 01:36:22.774
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).

383
01:36:23.014 --> 01:36:34.107
16. Moravec (1988) notably makes this point. See also Polanyi (1966), applied to automation in Autor (2015).

384
01:36:34.347 --> 01:36:42.704
17. Appendix A.4 presents variants of our exposure measure; these produce a similar ranking of jobs by exposure.

385
01:36:42.944 --> 01:36:47.251
18. See the Built Robotics Exosystem (2026).

386
01:36:47.491 --> 01:36:58.491
19. Results are similar when also weighting across occupations by usual weekly hours from the 2020 to 2024 American Community Survey.

387
01:36:58.731 --> 01:37:22.429
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.

388
01:37:22.669 --> 01:37:26.029
21. The Robot Report (2024).

389
01:37:26.269 --> 01:37:28.826
22. See the Boston Dynamics Stretch.

390
01:37:29.066 --> 01:37:32.274
23. See Pallottino (2026).

391
01:37:32.514 --> 01:37:43.287
24. Employment data in Figure 4 come from the BLS OEWS, so independent contractors for ride-sharing services are excluded from taxi driver employment.

392
01:37:43.527 --> 01:37:55.464
25. Early evidence suggests pay declines for drivers exposed to autonomous ride-hailing. See Gridwise (2025) and Yu and others (2026).

393
01:37:55.704 --> 01:37:58.449
26. See Amazon’s Vulcan robot.

394
01:37:58.689 --> 01:38:01.783
27. See the Badger Marty and the Simbe Tally.

395
01:38:02.023 --> 01:38:04.750
28. See the Telexistence TX SCARA.

396
01:38:04.990 --> 01:38:07.730
29. See the Canvas drywall robot.

397
01:38:07.970 --> 01:38:13.230
30. See the AMP Cortex (Waste360, 2023).

398
01:38:13.470 --> 01:38:28.592
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.

399
01:38:28.832 --> 01:38:33.799
32. See for example Massenkoff and McCrory (2026).

400
01:38:34.039 --> 01:38:39.400
33. Deming (2017) and Frey and Osborne (2017).

401
01:38:39.640 --> 01:39:05.325
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 and others (2003) on computerization and Acemoglu and Restrepo (2020) on the limited total impact of robots on U S jobs so far. Svanberg and others (2024) estimate costs to automate computer vision tasks.

402
01:39:05.565 --> 01:39:26.786
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.

403
01:39:27.026 --> 01:39:43.245
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.

404
01:39:43.365 --> 01:39:54.645
Total compensation includes paid leave, supplemental pay, insurance, retirement and savings, and legally required employer payments like FICA and unemployment insurance.

405
01:39:54.885 --> 01:40:07.311
37. On automation in task models, see for example Zeira (1998), Acemoglu and Autor (2011), and Acemoglu and Restrepo (2018).

406
01:40:07.551 --> 01:40:13.924
38. In these comparisons, robots may be cheaper than workers but only do a few of their tasks.

407
01:40:14.044 --> 01:40:34.374
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 5500 dollarsper year for robots to do this task.

408
01:40:34.494 --> 01:40:53.304
Archivists spend about 8 percent of their time on it, and their median total compensation is 85000, dollarsso labor costs for this task are 6800 dollarsper year. Robots are cost-competitive for this task, but automating one task is unlikely to displace archivists.

409
01:40:53.544 --> 01:41:16.580
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 3 million-Matic random case sealer (sealing).

410
01:41:16.820 --> 01:41:22.848
40. See for example Honerkamp (2026) and Lung (2026).

411
01:41:23.088 --> 01:41:52.668
41. August 2026 forecasts from the BLS Employment Projections program, 2025 to 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.

412
01:41:52.908 --> 01:41:58.345
42. See for example the Path Robotics AW-3 autonomous welding cell.

413
01:41:58.585 --> 01:42:07.950
43. Though cleaning robots might work longer hours, they are up to 10 times slower than humans. See Rivière and Denain (2026).

414
01:42:08.190 --> 01:42:23.614
44. This figure includes tire builders and packers and packagers. In these occupations, at least 95 percent of tasks by work time are exposed to robots, and robot costs fall below total compensation for a share of workers.

415
01:42:23.854 --> 01:42:30.765
45. See Davidson and Hadshar (2025) and Denain and Rivière (2026).

416
01:42:31.005 --> 01:42:56.713
46. Compare L L M exposure in Eloundou and others (2024) for computer programmers (25th most exposed by human raters, 6th by the L L M rater) and poets, lyricists and creative writers (11th by humans, 14th by the L L M). The two occupations were similarly exposed, but A I developers have since directed effort toward coding more than creative writing.

417
01:42:57.713 --> 01:43:13.898
That was What work can robots do?, from Anthropic Research, published September 30, 2026. Read the original at anthropic dot com. This unofficial audio edition was generated automatically for the podcast Anthropic, Read Aloud.