cs.ROOct 16, 2026
SaveHow exposed is work to robotics? Occupational, industry, and geographic measures
Organizations: MIT · MIT FutureTech · NYU
Abstract
Exposure indices that map technical capabilities onto occupational abilities have become a standard tool for anticipating how AI may reshape work, but existing measures either cover software AI or proxy robotics through patents and installation data that reflect the industrial robots of past decades. We construct the first abilities-based exposure measures for robotics and physical AI, and the first to span occupations, industries, and geographies. Following the framework of Felten et al. (2018, 2021), we build a robotics capability taxonomy of 16 capabilities in eight domains from 78 academic, industry, and standards sources, use a large language model to score the relatedness of each capability to the 52 ONET work abilities, and aggregate to Robotics Occupational Exposure (ROE) for 774 occupations, Robotics Industry Exposure (RIE) for 250 NAICS industries, and county-level Robotics Geographic Exposure (RGE). We find that robotics exposure concentrates in physically intensive work such as construction, farming, transportation, and production, and in industries such as logging, mining, trucking, and warehousing. Unlike generative-AI exposure, ROE is higher in occupations with lower wages, lower education requirements, and less creative-ability content. Physical collaboration and assistance, terrain-based locomotion, grasping and handling, and task and functional reasoning move the exposure ranking most, yet no single capability carries the index. Because the measures capture technical relatedness rather than readiness, cost, or adoption, they should be read as an upper bound on where physical automation could affect work.
Figures & tables
| Domain | Capability | Definition |
| Manipulation | Grasping & Handling | Direct contact, containment, and secure physical engagement with objects. |
| Dexterous & In-Hand Manipulation | High-degree-of-freedom interactions including reorientation and fine motor control. | |
| Tool Use & Process Interaction | Applying mechanical operations, force-controlled tasks, or specialized material processing. | |
| Complex Object Manipulation | Interacting with non-rigid, soft, or high-uncertainty environments. | |
| Locomotion & Mobility | Terrain-Based Locomotion | Ground-based movement using various physical mechanisms. |
Table 1: Robotics capability taxonomy: 16 capabilities grouped into eight domains.
| Robotics capability | O*NET ability | |
|---|---|---|
| Highest five | ||
| Aerial & Aquatic Locomotion | Control Precision | 1.000 |
| Complex Object Manipulation | Manual Dexterity | 1.000 |
| Dexterous & In-Hand Manipulation | Finger Dexterity | 1.000 |
| Dexterous & In-Hand Manipulation | Manual Dexterity | 1.000 |
| Expressive & Communicative Interaction | Oral Expression | 1.000 |
Table 2: Illustrative capability–ability pairs with the highest and lowest normalized relatedness scores ( ). Many pairs tie at 0 or 1 after averaging 20 LLM ratings; rows shown are the first five under alphabetical tie-breaking.
| Rank | Occupation |
|---|---|
| 1 | Pile-Driver Operators |
| 2 | Sewing Machine Operators |
| 3 | Sewers, Hand |
| 4 | Logging Equipment Operators |
| 5 | Pressers, Textile, Garment, and Related Materials |
| 6 | Industrial Truck and Tractor Operators |
Table 3: Occupational robotics exposure (standardized O*NET-based ROE, ).
Figure 1: Robotics exposure by job family. Panel (a) reports , where is the number of six-digit occupations in family , unweighted by employment counts. Panel (b) reports , where is national employment from BLS May OEWS.
| Rank | Industry |
|---|---|
| 1 | Logging |
| 2 | Truck Transportation |
| 3 | Coal Mining |
| 4 | Urban Transit Systems |
| 5 | Charter Bus Industry |
| 6 | Support Activities for Crop Production |
Table 4: Industry robotics exposure (standardized RIE, NAICS industries).
Figure 2: Geographic robotics exposure, contiguous U.S. Panel (a): standardized RGE by county; the color scale is capped at the 1st and 99th percentiles, and gray counties have no matched employment. Panel (b): workers by industry exposure class, one dot per 250,000 workers. A state with only one or two dots, such as Wyoming, can show only its largest classes.
Figure 3: Robotics exposure and labor-market correlates (ONET ROE). Each point is the mean of standardized ROE within an equal-count bin on the horizontal axis. Required education is on the ONET 1–12 scale (2 = high school diploma, 4 = some college, 6 = bachelor’s, 8 = master’s, 11 = doctorate; full list in footnote 2 ). Creative-ability weight is defined in footnote 3 .
Figure 4: Robotics exposure and demographic composition of occupations.
Figure 5: Robotics exposure (ROE), AI exposure (AIOE), and LLM task success by job family (matched sample, occupations). After z-scoring each measure on the matched sample, panel (a) averages within family with equal occupation weights; panel (b) uses OEWS employment weights. LLM task success is the share of a family’s O*NET tasks with at least 10% LLM time-saving potential times the acceptance rate on the surveyed subset of those tasks, pooled over 41 models ( Mertens et al., 2026 , Figure 2a and Table 2) .
Figure 6: Robotics versus AI exposure by job family (matched ROE–AIOE sample). Axes show mean standardized ROE and AIOE. Bubble area is proportional to BLS May OEWS national employment summed over matched occupations in each family. Numbers inside bubbles match rank by workforce size in the table. Bubble color shows the robotics-minus-AI gap (red: robotics-heavy; blue: AI-heavy).
Figure 7: Geographic robotics and generative-AI exposure. Top: county-level robotics exposure (RGE) and generative-AI geographic exposure (AIGE) on a shared color scale capped at the 1st and 99th percentiles, with equal weight per county. Bottom: each dot represents 250,000 workers (2019 QCEW), colored by whether the worker’s industry has low (below SD), medium, or high (above SD) robotics exposure (RIE, left) or AI exposure (AIIE, right); legend shares are the fraction of matched workers in each class. Dots are allocated and placed as in Figure 2(b) .
Figure 8: Robotics ROE, language-modeling AIOE, and image-generation AIOE versus wage, education, and creative ability. Required education is on the O*NET 1–12 scale (2 = high school diploma, 4 = some college, 6 = bachelor’s, 8 = master’s, 11 = doctorate; full list in footnote 2 ). Creative-ability weight is defined in footnote 3 .
Figure 9: Robotics ROE, language-modeling AIOE, and image-generation AIOE versus demographic shares.
| Capability | Domain | pct | 10% | 25% | 50% |
|---|---|---|---|---|---|
| Physical Collaboration & Assistance | Human–Robot Interaction | 2.5 | 12% | 6% | 4% |
| Terrain-Based Locomotion | Locomotion & Mobility | 1.9 | 12% | 6% | 3% |
| Grasping & Handling | Manipulation | 1.5 | 6% | 3% | 3% |
| Task & Functional Reasoning | Cognition & Planning | 1.5 | 8% | 2% | 2% |
| Tool Use & Process Interaction | Manipulation | 1.2 | 8% | 3% | 2% |
| System Health & Self-Management | System Maintenance & Autonomy | 1.2 | 5% | 2% | 2% |
Table 5: Effect of removing each robotics capability from the exposure index. Removal drops the capability’s row from the relatedness matrix and re-standardizes. pct is the mean movement of an occupation in the exposure ranking, in percentile points. Churn is the share of the most exposed 10%, 25% and 50% of occupations no longer in that group after removal.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
| Ability | Category | Definition | |
|---|---|---|---|
| Spatial Orientation | Cognitive | 11.75 | The ability to know your location in relation to the environment or to know where other objects are in relation to you. |
| Speed of Closure | Cognitive | 8.44 | The ability to quickly make sense of, combine, and organize information into meaningful patterns. |
| Depth Perception | Sensory | 8.35 | The ability to judge which of several objects is closer or farther away from you, or to judge the distance between you and an object. |
| Problem Sensitivity | Cognitive | 7.80 | The ability to tell when something is wrong or is likely to go wrong. It does not involve solving the problem, only recognizing that there is a problem. |
| Control Precision | Psychomotor | 7.39 | The ability to quickly and repeatedly adjust the controls of a machine or a vehicle to exact positions. |
| Response Orientation | Psychomotor | 7.12 | The ability to choose quickly between two or more movements in response to two or more different signals (lights, sounds, pictures). It includes the speed with which the correct response is started with the hand, foot, or other body part. |
Table 6: ONET 30.3 work abilities used in the relatedness matrix and ROE index (52 abilities). Sorted by ability-level relatedness , the sum of normalized relatedness across the 16 robotics capabilities (range 0–16). Definitions follow the ONET Content Model.
| Persona | Description |
|---|---|
| Robotics engineer | Hardware and control: actuators, sensors, end effectors, and physical variables the capability must sense or regulate. |
| Occupational therapist | Bodily demands of work: muscles, joints, senses, postures, and physical strain on the worker. |
| Industrial engineer | Production processes, material flow, task sequences, and physical operations in a job’s workflow. |
| Workplace safety and ergonomics inspector | Physical hazards and load: environmental exposure, force, repetition, and strenuous or risky job elements. |
Table 7: Evaluation personas for the relatedness matrix.
Figure 10: Pearson correlations between relatedness scores from the four evaluation personas. Panel (a) correlates raw matrix cells; panel (b) correlates ROE built from each persona’s matrix using the main-sample aggregation.
Figure 11: GPT-OSS-20B versus Qwen3-14B. Panel (a) compares the mean relatedness rating (0–4, averaged over personas and samples) for the 832 capability–ability pairs; panel (b) compares standardized ROE for 774 occupations built from each model’s matrix; panel (c) compares unweighted mean standardized ROE across the 22 job families. The dashed line marks equality.
Figure 12: Distribution of individual relatedness ratings (0–4) for GPT-OSS-20B and Qwen3-14B, pooled over all 832 capability–ability pairs, four personas, and five samples (16,640 ratings per model).
Figure 13: Human ratings versus GPT-OSS-20B relatedness scores for the 50 surveyed capability–ability pairs. Panel (a) reports the mean absolute difference between the human mean and the LLM mean, by the spread of the 20 LLM ratings (colored) and by the level of the LLM mean (gray); panel (b) plots each pair against the 45-degree line, colored by LLM spread, with human ratings jittered vertically to separate ties; panel (c) reports the Pearson correlation between human and LLM means for the same subsets.
Figure 14: Spearman correlation matrix for harmonized occupation-level exposure ( six-digit SOC codes).
Figure 15: ROE versus the Anthropic robot exposure index [ Legate-Yang and Massenkoff, 2026 ] . Panel (a) plots standardized ROE against the Anthropic index (0–3) for all 774 ROE occupations; panel (b) plots unweighted means across occupations for the 22 job families. Panels (c) and (d) standardize both measures across the plotted units; the solid line is the least-squares fit, whose slope equals the Pearson correlation, and the dashed line marks equality.
| Domain | per capability | Summed | Averaged |
|---|---|---|---|
| Manipulation | 1.00, 1.00, 1.00, 1.00 | 4.00 | 1.00 |
| Human–Robot Interaction | 0.20, 0.97 | 1.18 | 0.59 |
| Perception & Sensing | 0.38, 0.17 | 0.55 | 0.28 |
| System Maintenance & Autonomy | 0.26, 0.17 | 0.44 | 0.22 |
| Navigation & Spatial Awareness | 0.03, 0.24 | 0.26 | 0.13 |
| Locomotion & Mobility | 0.12, 0.10 | 0.23 | 0.11 |
Table 8: The two aggregation rules applied to one O*NET ability, Manual Dexterity.
| Domain | pct | 10% | 25% | 50% |
|---|---|---|---|---|
| Capabilities summed within a domain (published ROE) | ||||
| Manipulation | 11.6 | 36% | 27% | 16% |
| Locomotion & Mobility | 3.8 | 24% | 11% | 5% |
| Human–Robot Interaction | 2.9 | 14% | 6% | 5% |
| Navigation & Spatial Awareness | 2.1 | 9% | 3% | 3% |
| System Maintenance & Autonomy | 2.0 | 9% | 2% | 3% |
Table 9: Effect of removing each robotics capability domain from the exposure index. Removal drops all of the domain’s capabilities and re-standardizes. The upper panel sums the capabilities in a domain, which reproduces the published ROE; the lower averages them, so every domain counts the same however many capabilities it holds. pct is the mean movement of an occupation in the exposure ranking, in percentile points. Churn is the share of the most exposed 10%, 25% and 50% of occupations no longer in that group after removal.
Figure 16: ROE versus an O*NET-only version of ROE in which the 19 physical and psychomotor abilities receive a relatedness score of one and all other abilities zero ( ; both standardized). The dashed line marks equality; labeled occupations are among those that deviate most.
Figure 17: Robotics (ROE) and AI (AIOE) exposure by job family, radar view of Figure 5 . Each spoke is a job family, ordered clockwise by unweighted mean ROE; the radius is mean standardized exposure, and the dark ring marks zero.
Figure 18: Effect of removing each robotics capability from the exposure index, across 774 occupations, as reported in Table 5 . Each bar drops the capability from the relatedness matrix, re-standardizes, and shows the mean absolute change in an occupation’s exposure percentile.
More in this topic
AI and Labor
AI and labor studies how artificial intelligence changes jobs, wages, skills, hiring, and the organization and quality of work.
See all 95 papers