Automation across diverse geographies.
A department of transportation owns one of the largest distributed physical estates in public hands, and the recurring work of keeping it open is the largest line in its budget. The question this topic asks is not whether Physical AI can take that work on. It is where on the trajectory each application sits, and which constraint is currently binding it, given that the answer changes with the ground the estate is built on.
The review is organised by terrain and climate class rather than by state, because the physical environment is what decides what the work actually is. An archipelago separates its assets with water. An Appalachian county carries steep grade, narrow valleys and, after Helene, a rebuild backlog. An Alaskan corridor closes for part of the year and the construction season is what bounds it. A tribal route network in the arid southwest runs very low route density over an area larger than ten states, much of it beyond grid power and cellular coverage.
The claim under test is that the binding constraint changes class with the terrain, which would make a single national automation case the wrong unit of analysis. Each hypothesis resolves to a position on the trajectory stated in degrees, a constraint named from the five technical classes or from regulatory and workforce where those bind, and the measured change that would move the boundary. Every figure carries a verification grade, and quantities this review could not locate are recorded as not located rather than asserted absent.
Eleven hypotheses were stated before the evidence was assembled. Five of them bind on regulation, three on workforce, three on engineering implementation, one on material science and one on energy. None binds on AI algorithms, and the one that binds on computation efficiency sits downstream of a compliance rule rather than upstream of a capability. On this estate the question of whether a machine can do the work is settled more often than the question of whether the work can be bought, measured or reached.
One hypothesis was written to be capable of refuting the review's own partition, and it survives in part. Inside the Appalachian class the share of bridges in poor condition runs from 3.43 percent in Virginia to 17.79 in West Virginia against a national 6.68, and two of the four steep-terrain anchor states sit below the national figure. The alternative that hypothesis proposes fails on the same four points: the most consolidated state by ownership has the worst bridges and the most fragmented has nearly the best. The terrain variable that would settle it, a published grade or curvature distribution, was searched for and not located for any of the four states. Terrain is therefore kept as a partition of the physical work, where the measured class separation runs from 3 to 123 times, and is explicitly not claimed as a predictor of asset condition.
TR-2026-34 · Research / Review, preprint v1 · 41 pages · The Hiner Lab with The Charlot Lab
Michael Hiner, Ronnie Gomez and Grant Markhart, Industrial Research Fellows, with Dean David Jean Charlot.
Download the review (PDF)Read online← The Hiner LabThe fellowship →
The finding, made operable.
Pick a terrain class and see the quantity that decides automation there, the constraint in the way, and the test that could have overturned the whole partition.
A transportation estate is not one estate, and that is the finding: the binding constraint changes class with the terrain. One national number averages over places where materials bind, places where energy binds and places where nothing binds except procurement. Read them separately and each becomes a different, tractable piece of work.
One of eight, and only one of them is physics. How we read a frontier →