Legate-Yang and Massenkoff build a robot exposure index: robots can already do 74% of US physical tasks but are cost-competitive for just 0.3% of work

The authors propose a four-tier exposure scale based on how controlled an environment a robot needs (E0 cannot perform, E1 purpose-built robotic environment, E2 structured human workplace, E3 unstructured environment) and use it to build an occupation-level robot exposure index. Prior AI exposure work largely targeted cognitive tasks that LLMs could speed up; this work extends measurement to physical tasks that require robots and characterizes automation difficulty by degree of environmental control rather than a binary can/cannot. Built on roughly 900 occupations and about 19,000 task descriptions from O*NET, with Claude searching for specific robots and quoting sources; 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, and results are similar when ratings relying only on demonstrations are omitted.

Robots can perform 74% of physical tasks, or 34% of all working hours; cognitive and interpersonal work makes up 54% of tasks by working time and physical work 46%, split into 12% E0, 23% E1, 10% E2, and 1% E3. The result turns 'what robots can do' from individual demonstrations into an economy-wide, task-level accounting, and shows exposure concentrates in driving and warehousing jobs while nursing and general repair jobs remain unexposed even in highly controlled settings. Tasks are weighted by the number of workers doing them and the fraction of working time spent on them; 9 of the 10 most exposed occupations are vehicle operators, with taxi drivers leading at an index of 2.2.

Robots expand technology exposure from about half of work under LLMs alone to 81%, and the exposed workforce looks opposite to the LLM-exposed one: highly exposed workers are 20 percentage points less likely to be female, 16 points more likely to be Hispanic, 55 points less likely to hold a bachelor's degree or higher, earn about $30 less per hour, and face more than twice the unemployment rate. Prior LLM exposure research pointed mainly at cognitive white-collar tasks; this work shows robots cover physical work that LLMs cannot, so the two technologies affect different populations. Worker characteristics come from the 2020-2024 American Community Survey; Bureau of Labor Statistics occupational requirements show exposed jobs more often require carrying weight and working in extreme heat or near hazardous contaminants.

Although robots can do most physical tasks, they are cost-competitive for only 0.3% of work; if robot prices keep falling at roughly 3% per year, reaching 10% would take about 40 years. Unlike LLM exposure measures that assess capability without cost, this work estimates annual robot deployment cost per exposed task and compares it with labor cost, showing a large gap between capability and affordability. Cost estimates cover the robot itself, accessories, integration and installation, maintenance, software, energy, oversight and operation, insurance, and decommissioning; for packers and packagers, robots cost over $2 million to purchase and install and replace about 14 workers' yearly output, annualized over roughly a 10-year life at an 8% cost of capital to about $45,000 per replaced worker versus about $49,000 in compensation.

The index is meant for readers who want to judge where physical automation appears first: policymakers can monitor exposed occupations such as drivers and warehouse packers for early signs, firms can assess the cost competitiveness of their own jobs, and researchers can update the work periodically to track robot capabilities, as the authors suggest. The results apply to the US economy and to today's robot capabilities, and the authors state their premise explicitly: jobs are more likely to be affected when today's robots already do their work in some settings.

The authors note that AI-powered robots could leapfrog their scale and do work they cannot today, such as learning to climb ladders or use arms and grippers more deftly; the cost 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, while robotics companies may prioritize certain capabilities for other reasons; cost parity need not imply high unemployment or rapid growth, since tasks still done by workers may become bottlenecks and cost estimates often include human supervision, exception handling, or repair; and the authors say they have not considered how AI could affect physical work without robots, for example by better predicting when factory machines need maintenance.