Executive Summary

Key Figures

The first two numbers set the direction the index points. The second two set how far that direction spreads across the occupation table and the map.

Source: How exposed is work to robotics? Occupational, industry, and geographic measures (arXiv:2610.24267, October 16, 2026)

The exposure index is a familiar tool by now. Write down what a technology can do, write down the abilities people use on the job, hold the two lists against each other, and turn the overlap into a score. The larger the overlap, the closer that technology sits, technically, to the tasks in that occupation. Most of the estimates circulating about generative AI came out of this frame.

On the robot side, no index built that way existed. Studies of how robots affect employment reached instead for patent counts and installed industrial robot units. What those numbers show is the robot of the past several decades, bolted to a factory floor and repeating one motion. The machine that rolls through a warehouse and hands a part to the worker beside it barely registers in that record.

Sebastian Sartor and Neil C. Thompson of MIT FutureTech, with Robert Seamans of NYU Stern School of Business, have filled that blank with a capability-based measure. They keep the frame Felten and colleagues used in 2018 and 2021 for the AI occupational exposure index and swap out only the technology-side list. It produces three measures: ROE, or Robotics Occupational Exposure, scored across 774 occupations; RIE, or Robotics Industry Exposure, scored across 250 industries under the North American Industry Classification System; and RGE, or Robotics Geographic Exposure, which carries the industry values down to the county by way of local employment mix. This is the first robot exposure measure to link occupation, industry and place in one frame.

The names at the top of the occupation list show quickly what the index picks up. Pile-Driver Operators lead, followed by Sewing Machine Operators; Sewers, Hand; Logging Equipment Operators; and Pressers, Textile, Garment, and Related Materials. Once the occupations are grouped into families, exposure gathers in construction, farming, transportation and production. Among industries, Logging comes first, with Truck Transportation and Coal Mining right behind it and Warehousing and Storage inside the top twenty.

The bottom of the ranking looks nothing like the top. Of the 250 industries, the last place belongs to Legal Services, with Agents and Managers for Artists, Athletes, Entertainers, and Other Public Figures and then Junior Colleges just above it. Among the 774 occupations, the final two are Foreign Language and Literature Teachers, Postsecondary and Political Science Teachers, Postsecondary. Work done with the body sits at one end and work done with words at the other. The same logic puts Urban Transit Systems and the Charter Bus Industry into the top ten alongside freight trucking. Sitting at a wheel and running a fixed route, seen through this index, is not far from working a factory line.

One point invites confusion with a similarly named measure. The robot exposure index Anthropic put out in late September has different authors and a different method, and it asks mostly about cost and about when automation arrives. That story is covered in a separate piece. The ROE discussed here is an open measure from university researchers, and what it asks about is not timing but rank order.

Not one measurement of an actual robot's performance went into this index. What went in was documents. The researchers collected 78 academic papers, industry reports and government or standards documents and had a language model read them. Google Gemini ran the work in two stages: first each PDF was processed on its own to pull out the robot capabilities described in it, then the extracted material was consolidated into a single structure. What came out was 16 capabilities spread across eight domains.

The eight domains are Manipulation; Locomotion and Mobility; Navigation and Spatial Awareness; Perception and Sensing; Cognition and Planning; Human-Robot Interaction; Multi-Agent and Collective Behavior; and System Maintenance and Autonomy. The 16 capabilities under them include grasping and handling objects, turning something precisely within the hand, moving over rough terrain, building a map and knowing one's position on it, and working physically alongside a person.

A different model did the scoring. Matching each of the 16 capabilities to each of the 52 work abilities in O*NET, the US occupational information database, gives 832 pairs. The open-weight model GPT-OSS-20B gave each pair a whole number from 0 to 4, and those values were divided by 4 to land between 0 and 1. The model was not asked once and left alone. It was given four personas in turn — robotics engineer, occupational therapist, industrial engineer, and workplace safety and ergonomics inspector — and sampled five independent times under each, so every pair averages 20 ratings.

The researchers also recomputed the index with one capability removed at a time. The ones that move the ranking most are physical collaboration and assistance, terrain-based locomotion, grasping and handling, and task and functional reasoning. Removing any single one of them, however, did not overturn the ordering. The paper concludes that robot exposure rests on the breadth of capabilities rather than on any one of them.

The paper lays the index over three attributes of the occupations themselves: wages, required education and creative weight. Sort occupations into wage bands and the lowest band averages a standardized ROE of +0.23 against −0.65 in the highest. By required education the lowest band sits at +0.85 and the highest at −1.22. The widest spread of the three belongs to how much weight creative abilities carry in the work, running from +1.29 down to −1.07. As correlations, that is −0.42 for wages, −0.71 for education and −0.85 for creative weight.

Read the same occupation table through the generative AI exposure index and the order comes out nearly reversed. At the level of occupational families, the correlation between the two rankings is −0.93. Construction and extraction stands at the top for robot exposure and near the bottom for AI exposure. Computer and mathematical occupations do the opposite.

The contrast stands out because the employment debate of the past few years has effectively been run on one of these indices alone. The claim that office work and credentialed professions get shaken first is the direction the generative AI measure points. When physical AI is measured on its own, the people in the line of fire change. If both currents are moving at once, a picture drawn from either index by itself is half a picture.

Lay industry exposure over the employment mix of a place and a map appears. The method counts what percentage of everyone working in a state holds a job in a high-exposure industry. High exposure means a four-digit industry whose RIE sits more than 0.5 standard deviations above the mean, and low exposure means one that sits the same distance below it. The same calculation runs on state aggregates and on the 3,217 counties and county equivalents left after metropolitan groupings are stripped out.

▲ The four states with the highest share of workers in high-exposure industries and the two lowest. The percentages come from arXiv:2610.24267; the right-hand column is a Pebblous note on each area's widely known employment mix.

Those percentages need a national baseline beside them. Across the United States, 18% of workers are in high robotics-exposure industries. With the same threshold on the AI side, 43% fall into high-exposure industries. Wyoming's 31% sits 13 percentage points above the national baseline, and the 7% in Washington, DC sits 11 points below it. That gap also says the two measures cover different amounts of ground to begin with. The robot measure reaches less far, and what it does reach is packed into a handful of states.

Take New York City on its own and it becomes clear how differently the two measures read one city. Of the people working there, 62% are in industries with low robot exposure and 58% are in industries with high AI exposure. The two shares are not slices of the same group but counts made on different criteria, which is why they add to more than 100. The authors write that the two measures diverge most in large coastal cities, and note the same pattern in Washington, DC, Boston and the San Francisco Bay Area. So this is not one city's oddity. The direction is unmistakable either way: the places that look safe on the robot map sit in the middle of the AI one.

The authors drew their own line on how far these numbers may be pushed. What the index measures is technical relatedness, not readiness, cost or adoption. In the paper's own terms, the measure should be read as an upper bound on where physical automation could affect work. Wyoming's 31% does not mean three in ten workers in that state will hand their jobs to robots; it means that on technology alone, the overlap is largest there. Whether machines actually arrive is settled separately by price, regulation and what people will accept.

What held our attention longest in this index was not the result but the material. The places where ROE comes out high are not places someone observed robots performing well. They come out high because the capability list written into 78 documents has that shape. Documents tend to record demonstrations that worked and research that succeeded. Picking up a box in a warehouse, walking over rough ground, handing an object to the person beside you — these come up often in papers and conference material, and so they settle into the capability list under clear names. Work that nobody has yet managed, and that therefore appears nowhere, is not on the list at all.

Saying that the material rather than the subject of measurement set the conclusion is not a criticism. It is what the authors put down as their second limitation. For anyone looking at physical AI from the data side, though, that limitation turns into a different question. If the capability list for robots could be written from field records instead of documents, how would the list change? And where are those records piling up right now?

That the answer overlaps the ROE map is the point this article was heading toward. The data that trains physical AI is gathered in warehouses, on farms, at construction sites and in logistics terminals. Those are the same industries the index scores highest. One set of sites is both where the data comes from and where the machines built on that data arrive first. The people working there today are producing the data and standing where the index points hardest, at the same time.

Making data usable includes recording whose hands it came from. With robot data this record goes missing especially easily, because training runs fine on trajectories, sensor values and success flags alone. If which site, which working conditions and whose motion were observed is not written down beside them, there is later no way to retrace where that data may be used and where it may not.

Thanks for reading this far. The facts here were checked against the paper and abstract published on arXiv, and the figures were compared twice in separate passes. If you gather robot data in the field, we would be glad to hear how you record where that data came from.

Pebblous Data Communication Team
October 21, 2026

• 1.Sartor, S., Seamans, R., Thompson, N. C. (2026). "How exposed is work to robotics? Occupational, industry, and geographic measures." arXiv:2610.24267.
• 2.Felten, E., Raj, M., Seamans, R. (2021). "Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses." Strategic Management Journal, 42(12), 2195–2217. The generative-AI exposure (AIOE) methodology that ROE reuses wholesale, swapping in a robotics capability list.