Robotics in Infrastructure

Robotics in infrastructure is three trades wearing one label. Inspection sends a machine to look at something — a drone over a bridge soffit, a magnetic crawler up a crane leg, a climbing robot on a turbine blade, a tethered vehicle down a penstock. Maintenance and repair sends a machine to do something to the asset once it has been looked at, which is a much harder problem and a much smaller literature. Construction sends a machine to make the asset in the first place, which is where the money and the rhetoric are, and where the measured evidence is thinnest. Established The three differ in kind and not merely in difficulty: inspection returns information, maintenance changes an asset's state, and construction creates the asset, and only the first is a task whose success can be checked by looking at what came back.

The field's centre of gravity is not where its press coverage is. Established What is demonstrably working, in named field deployments with published accuracy figures, is inspection: crawling, climbing and flying platforms that reach surfaces a person should not be asked to reach, and return a georeferenced record. Frontier What is being sold is construction robotics, on the strength of a productivity crisis that has been documented for half a century and never once decomposed into its causes.

That undecomposed crisis is this brief's subject as much as the machines are, because every business case in the sector is written against it. Established The standard framing sets construction's sub-one-per-cent annual labour productivity growth against manufacturing's 3.6 per cent and concludes that construction is uniquely broken. Frontier A separate and well-established literature on offshoring bias argues that the manufacturing number in that comparison is itself inflated by between 0.1 and 0.5 percentage points a year. Speculative The research behind this brief found no instance of that correction being applied inside the construction-productivity literature, which cites the manufacturing comparison constantly. Every point of overstatement in the benchmark is a point of the gap that never existed, and the correction is made here.

Established The productivity record, as the field states it, comes from a small number of figures that are quoted far more often than they are chased. The most recent consolidated statement available to this brief is Bühler, Nübel, Jelinek, Köhler and Hollenbach's 2025 framework paper in Buildings, read in full: over roughly two decades construction productivity has risen about 10 per cent while overall global economic productivity has risen more than 50 per cent; annual labour productivity growth in construction has stayed below 1 per cent against 2.8 per cent for the world economy and 3.6 per cent for manufacturing; and global construction output is expected to reach 22 trillion US dollars a year by 2040. Frontier Two cautions travel with those numbers and are almost always stripped off. First, they are figures the paper attributes to its own numbered references; this brief read the citing paper, not the primaries, and says so. Second, the 3.6 versus sub-1 comparison carries no stated period. A multi-year average without its window is a rhetorical object, not a measurement, and it is the single most repeated pair of numbers in the sector.

Established The decisive study is Goolsbee and Syverson's 2023 NBER paper, and it says something more careful than it is usually cited for. Their contribution is triangulation: physical productivity measures — houses completed per worker-year, which no price deflator can corrupt — alongside materials-to-output efficiency and state-level variation, showing a decline running more than fifty years and identifying market frictions that limit the reallocation of activity toward more productive firms. Frontier Their stated conclusion is that measurement error is probably not the sole source of the stagnation. That formulation is deliberate, and it concedes in its own words that measurement error is a source. Frontier The paper does not decompose the decline into measured and real components and does not claim to. This brief obtained the deposited abstract summary and not the full text; the physical-productivity series and the state-level results are therefore described here at the level the authors summarise them, not read.

Frontier Here is the correction this brief exists to make, and it is a benchmark correction rather than a refutation. Construction is judged against manufacturing. Houseman, writing in the International Labour Review in 2007, states that the growth of domestic and foreign outsourcing “may result in inflated and misleading increases in productivity statistics”; Houseman, Kurz, Lengermann and Mandel's 2012 work on offshoring bias puts a number on it, arguing that failing to capture import-price declines from foreign sourcing potentially overstates measured US manufacturing productivity gains by 0.1 to 0.5 percentage points annually. Established Both figures are read from publisher-deposited abstracts. Frontier Apply the top of that range and the manufacturing comparator falls from 3.6 per cent to roughly 3.1 per cent, which removes about a fifth of the headline gap before anyone argues about construction at all. Speculative The research behind this brief found no instance of this being done in the construction literature. It is not a refutation: Goolsbee and Syverson's physical measures are independent of the manufacturing comparison and survive it untouched. What it damages is the framing — the claim that construction is uniquely and catastrophically bad — which is the claim doing the persuasive work in every robotics business case.

Frontier A second precision about the same paper, which is about how evidence arrives rather than what it says. Syverson had separately published the general case against mismeasurement explanations for productivity slowdowns, in an NBER working paper and a subsequent journal article, before the construction paper appeared. Established Those two items are verified bibliographic records here and were not read. Frontier The point is not that the construction result is wrong — the triangulation stands on its own — but that it arrives from an author with a stated public prior against the alternative hypothesis, and a reader deciding how much weight to put on “probably not the sole source” is entitled to know that. Frontier The statistical-agency treatment that would sit on the other side of the scale, by Sveikauskas and colleagues at the US Bureau of Labor Statistics, is the strongest form of the measurement case and could not be obtained for this brief at all. Its absence is declared rather than papered over: this brief is therefore better sourced on the case that the decline is real than on the case that it is measured, and the imbalance is a property of what was reachable, not of the evidence.

Established What construction robotics has actually measured is less than its literature implies, and the gap has a specific shape. The most-quotable result in the recent record — 84 per cent faster bricklaying, 6.75 against 43.2 days per unit, 94 per cent productivity retention under 30 per cent workforce absenteeism, 15 per cent mortar-waste reduction, payback in under five months — is the output of a simulation of ten hypothetical residential buildings of eighty units under Jordanian codes. Frontier Nothing was built. The manual baseline of 43.2 days per unit is doing most of the arithmetic work in that 84 per cent, and the abstract does not source it. Frontier The stronger paper is Dindorf's 2026 case study of an industrial robot performing bricklaying in Poland, which runs from a Revit model through simulation in ABB RobotStudio to an ABB IRB 4600 and reports laboratory work, on-site demonstrations and commercial wall-laying in residential apartments with a return-on-investment calculation against labour cost. Frontier A rebar-tying robot with two-camera target identification has been field-validated at a named site, the Shenyang Hunnan Science and Technology City Phase IV project, and reports its contribution as innovation rather than as a productivity figure — which is characteristic of the genre. Established So the machines work. What does not exist anywhere this brief could reach is a controlled comparison of square metres of wall per crew-hour, robot against manual, on the same site to the same design.

Established Inspection is the part of the field that is genuinely in service, and its published numbers are modest, concrete and useful. A metamorphic climbing robot for wind-turbine blades, using a ten-revolute folding mechanism that reconfigures between rectangular and hexagonal topologies to hold adhesion on compound curves, achieves a maximum crawling speed of 125 millimetres per second — 0.45 kilometres per hour, which is the honest scale of this technology against a blade tens of metres long. Established A magnetic-adsorption wall-climbing robot for gantry-crane inspection was analysed under pulsating wind and re-optimised for a 99.19 per cent reduction in induced vibration displacement in simulation, wind pressure concentrating on the front of the vehicle body. Established A wall-climbing robot paired with a neural radiance field for dam-surface inspection reports an average peak signal-to-noise ratio of 33.99 and average depth error of 0.027 metres within 2.5 metres — a 27-millimetre accuracy figure of the kind this literature usually omits. Frontier Two wind-turbine tower-climbing prototypes, one on four mecanum wheels with a winding adhesion mechanism and its successor on four rubber wheels with a Bowden-cable winder, were validated indoors across speeds and payloads. Frontier That last detail is the subfield's whole story in miniature: the distance between a climbing robot demonstrated and a climbing robot in maintenance service is where almost all of the remaining work lies. Established Note also the asymmetry between the two halves of this section's title. Inspection has field deployments, named sites and published accuracy figures. Maintenance and repair robotics — a machine that does something to the asset rather than looking at it — has a far smaller literature, and what this brief could reach of it is not robotics at all but scheduling: grouping maintenance actions under variable cost, and forecasting the weather windows in which any intervention, robotic or crewed, is possible. Frontier That is a real signal about where the difficulty sits. Getting a machine to the defect is close to solved; deciding when to send it, and having it do anything useful when it arrives, is not.

Established One historical number should be set beside every robotics business case in this sector. Frank Gilbreth's time-and-motion work reduced the motions required to lay a brick from eighteen to as few as two, raising output from 120 to 350 bricks an hour — a factor of about 2.9, achieved roughly a century ago by method study alone, with no machinery whatsoever. Frontier It is the largest documented productivity gain in the trade that construction robotics proposes to transform, it was obtained by reorganising human work rather than replacing it, and it is almost entirely absent from modern automation business cases. Speculative Its omission is informative: a sector that had already demonstrated a near-tripling through process design, and then did not compound it, has an organisational problem that a faster machine does not obviously address.

Frontier The most interesting move in the current literature is to instrument the construction process rather than argue about how to measure its output. Bühler and colleagues propose a four-tier operational, tactical, strategic and normative framework built on standardised repeatable process steps mapped through value-stream mapping, capturing input efficiency, output effectiveness and first-time quality, aggregated through takt-time compliance and schedule reliability, with sensor-equipped machinery and machine learning automating the data capture and flagging deviations. Frontier If the measurement critique has force, this is a methodological answer to an econometric dispute: stop deflating and start counting. Frontier The same logic runs through real-time bricklaying progress monitoring in which a YOLOv8 detector feeds a building-information-model digital twin so that actual progress can be compared against planned and the schedule updated live. Speculative Note what that automates. It automates measurement, not work — and if the sector's central problem is that it cannot see itself, instrumentation may be worth more than mechanisation. That inverts the usual priority ordering and it is cheap to test.

Frontier On the inspection side the frontier is representation and adhesion. Neural radiance fields are being adopted as the scene representation for infrastructure inspection in place of point clouds and photogrammetry, which changes what a defect record is: not a mesh with photographs attached but a rendering function from which arbitrary views and depths can be recovered. Frontier Metamorphic and reconfigurable climbing mechanisms that change their own topology to keep contact on compound-curved surfaces are the corresponding hardware move. Frontier On-site construction logistics is emerging as a distinct robotics problem from construction task robotics — material handling with task distribution, routing and team composition, solved with variable neighbourhood search and particle swarm optimisation — which is the warehouse stack arriving on a building site. Speculative That reframing may matter more than it looks: moving material is the largest single category of on-site labour that is neither trade-skilled nor design-dependent, and it is the part of a site that most resembles the structured environment robots already work in.

Frontier A finding that looks minor and is not: the building sets the robot's performance ceiling. Work on how finishing materials affect autonomous mobile robots identifies colour, texture, transparency, slipperiness, unevenness and wheel-mark resistance as properties that each bear on sensing or locomotion, and states plainly that the knowledge base here is mostly anecdotal. Speculative Inverted, this is a design brief rather than a robotics problem: buildings and structures could be specified to be robot-legible, with surface, marking and access requirements written into the specification the way fire and accessibility requirements now are. Handwave Taken to its limit, that is the self-maintaining-infrastructure programme — an asset designed from the outset around the machines that will inspect and service it for a century, with docking, power, reference targets and clearances as first-class design constraints. Nobody has written that specification, no code contains it, and this brief could locate no source that has attempted it.

Frontier The live questions, stated as questions. What fraction of the fifty-year decline is measurement, what fraction is product change, what fraction is composition and what fraction is real — nobody has answered this and it decides where the money should go. What is a construction robot's utilisation on a live site, expressed as productive hours over hours on site? What is the sea-state envelope for robotic against crewed offshore intervention, and does robotics add availability or only remove risk? Can a collision-avoidance or inspection-decision algorithm be certified rather than merely validated? Speculative And a structural question the field does not ask itself: whether construction should be benchmarked against manufacturing at all, rather than against other one-off, site-assembled, weather-exposed activities such as shipbuilding, large plant erection or film production. If construction's record resembles theirs, the uniquely-broken framing collapses, and the comparison is testable with sectoral data that already exists.

Frontier One sourcing fact belongs in this section rather than in a footnote. The research behind this brief was conducted with arXiv gated throughout. The robotics and autonomy preprint literature — which is where this field's frontier actually lives, and where locomotion, manipulation, perception and sim-to-real results appear one to two years before journal publication — is therefore entirely absent from the evidence base here. Established What was reachable was MDPI full text and publisher-deposited abstracts through Crossref, which is why the hard numbers above are disproportionately from Buildings, Machines, Applied Sciences and JMSE. Frontier That divergence between the readable literature and the authoritative literature is a real distortion and it runs in a predictable direction: open-access journal papers over-represent completed, packaged demonstrations and under-represent the messy capability frontier. A reader should assume the frontier is somewhat ahead of what is described here, and should not assume the field's evidence about productivity is any better than what is described here, because that evidence is economic and does not live on arXiv.

Frontier The binding constraint in this field is not a technology, and that is an unusual answer. The workback plan for a construction sector with manufacturing-like productivity has ten links, and the one that binds is the third: a decomposition of the decline. Link one is a productivity measure both camps accept — physical output per worker-hour, quality-adjusted for code-driven content such as envelope, seismic, fire, accessibility and services, deflated by an output-price rather than an input-cost index. Link two is a corrected benchmark, with manufacturing productivity growth re-estimated free of offshoring bias and then re-compared. Speculative Both are methodological, both are within reach of two competent economists in about a year, and neither requires a robot.

Frontier Link three is what decides where the money goes, and fifty years of documentation have gone by without it. Split the decline into measurement, product change, compositional shift and real productivity loss. If it is mostly real and technological, buy robots. If it is mostly artefact, fix the statistics and stop panicking. Speculative If it is mostly the reallocation frictions Goolsbee and Syverson name, the remedy is market structure and regulation, and no machine touches it. Frontier Every robotics investment case in this sector is currently being made against an undiagnosed condition.

Frontier The secondary bottleneck, binding only if the decline is real, is robot utilisation on a live site. Field speed is demonstrably achievable. Utilisation is not published by anyone. A construction robot waits for preceding trades on independent schedules, for materials, for weather, for access, and for the site to be cleared of the people it cannot work beside. Established A machine three times faster at 20 per cent utilisation is slower than a person at 70 per cent, and that arithmetic is the whole of the case. Frontier The barrier literature agrees in a different vocabulary: a fuzzy DEMATEL study of Indian construction places high initial cost, rigidity in organisational processes and procedures, interoperability, lack of competency and lack of standardisation in the causal tier, with market uncertainty and ethical concerns downstream of them. Four of those five causal barriers are organisational.

Established On the inspection side the bottleneck is throughput, and on the maintenance side it is weather. At 125 millimetres per second a climbing robot is not competing with a rope-access team on area covered per day once setup and recovery are counted; it is competing on access to places people should not go. Frontier Offshore, the constraint is not the machine at all: subseasonal-to-seasonal forecasting research aimed at wind-turbine maintenance scheduling works to horizons of up to six weeks precisely because the scarce resource is a weather window and a vessel, not a manipulator. Frontier If the sea-state envelope for robotic intervention is similar to the crewed envelope, robotics has removed risk without adding availability — a real safety gain and no productivity gain, and the two are routinely booked as one.

Frontier This brief is waiting on four results, none of which is a robotics result. The first is the Sveikauskas group's statistical-agency work on construction productivity measurement, which is the strongest form of the measurement-artefact case and which could not be obtained here. The second is the full text of Goolsbee and Syverson, in particular the physical-productivity series and the state-level analysis, which are described in summary above and not read. Established The third is any published figure for construction-robot utilisation on a live site. Frontier The fourth is a field-measured robot-against-manual productivity comparison on the same design.

Frontier Two of those four are research gaps rather than retrieval gaps, which is a stronger statement about the field than any finding in this brief. Utilisation is trivially measurable by anyone operating a machine on a site and is not published by anyone. Established The controlled productivity comparison requires one contractor, one building, two crews and a stopwatch. Speculative Their joint absence, in a field whose entire justification is a productivity claim, is the most telling fact available about how seriously that claim is held by the people making it.

Established The dependencies that are technological run in one direction: from sensing and representation into the robot, not the other way. Structural health monitoring, sensor fusion and data reduction for civil infrastructure supply the reference against which a robotic inspection is judged; if the fixed instrumentation is good enough, some inspection missions disappear rather than being automated. Frontier Building information models supply the geometric prior that lets a robot know where it is, and the digital-twin work above closes that loop. Frontier Everything else — locomotion, adhesion, manipulation, path planning — is being advanced primarily in a preprint literature this brief could not read.

Frontier Ten measurements would settle most of what is contested here, and each is specified deliberately tightly. One: physical output per worker-hour for a standardised product — the same house type to the same code vintage — over a long series. Speculative Two: the same series with an explicit hedonic adjustment for code-driven quality content, which is the direct test of whether physical measures are themselves contaminated and therefore the sharpest available challenge to Goolsbee and Syverson's strongest evidence. Frontier Three: construction productivity recomputed with an output-price rather than an input-cost deflator, which quantifies the standard technical objection instead of asserting it.

Frontier Four: manufacturing productivity growth re-estimated free of offshoring bias and then re-compared with construction, which is the benchmark correction this brief argues for and which nobody appears to have run. Speculative Five: the dispersion of productivity across construction firms and the rate at which market share moves from low- to high-productivity firms, which turns the reallocation-frictions claim into a number. Frontier Six: square metres of wall per crew-hour, robot against manual, same site, same design, same code, with the crews swapped between halves of the building to control for learning.

Established Seven: productive robot-hours divided by robot-hours on site, reported with the reasons for each idle interval — trade interference, materials, weather, access, fault. Established Eight: square metres inspected per robot-day against per inspector-day, with setup, mobilisation and recovery time included on both sides, because rope access does not incur a robot's rigging and calibration overhead and comparisons that omit it are not comparisons. Frontier Nine: the sea-state and wind-speed envelopes for robotic against crewed intervention on the same asset class, which decides whether offshore robotics buys availability or only safety. Frontier Ten: the count of code-compliant printed structures carrying certified structural reinforcement, against the count of printed demonstrators — a ratio that is the field's honest maturity measure and that is small.

Speculative One further experiment is cheap and nobody has run it. Instrument a single project end to end and compare takt-time compliance as instrumented against takt-time compliance as self-reported by the same site's management. If the gap is large, the sector's measurement problem is not econometric at all but managerial, and the intervention that pays is a sensor rather than a robot.

Established The engineering problems in inspection robotics are adhesion, endurance and the tether. Magnetic adsorption works on ferrous structures and does not work on concrete; suction and electro-adhesion work on smooth surfaces and degrade on the rough, wet and contaminated ones that actually need inspecting; Frontier the metamorphic reconfiguration approach exists because a wind-turbine blade is a compound curve of continuously changing radius that no single fixed-geometry contact patch handles. Frontier Wind loading on a climbing platform is a live structural problem in its own right, concentrated at the front of the vehicle body and addressable by shape optimisation to a 99.19 per cent reduction in induced vibration displacement in simulation. Established Power and data are almost always solved by a tether, and the tether then becomes the operational constraint: it snags, it weighs, it limits reach, and it is what makes an untethered drone attractive despite its far worse endurance.

Frontier On the construction side the engineering constraint is the site rather than the machine. Surfaces are wet, uneven, provisional and changed daily by other trades; the floor a robot navigates on Tuesday is not the floor it navigated on Monday. Established Finishing-material properties — slipperiness, unevenness, transparency, texture, colour — bear directly on sensing and locomotion and are documented mostly anecdotally. Frontier There are two routes out. Either the robot tolerates the unstructured site, or the site becomes structured — and structuring the site is prefabrication, which is a different industry, which is precisely why prefabrication keeps being proposed as the answer to construction productivity and keeps not scaling.

Frontier Two cross-cutting engineering constraints are usually missed. The first is network: outdoor and yard robotics that offloads computation is limited by guaranteed-latency coverage rather than by its autonomy stack, and whether that constraint survives depends on whether onboard compute closes the gap within a hardware generation. Established The second is mobilisation. A construction robot must be transported, craned, powered, commissioned, supervised and demobilised for each deployment, and every published payback figure this brief could reach is computed on operating speed rather than on realised deployment cost.

Established The nearest neighbour is warehouse and logistics robotics, assessed in Autonomous Supply Chains, which shares this field's central defect: an enormous simulated literature and a nearly empty validation literature. Frontier On-site material handling is now being formulated with the same optimisation machinery used inside distribution centres, so the seam between the two fields is closing from the logistics side rather than the construction side. Speculative The transferable lesson runs the other way too: warehouse automation's performance depends on the floor it runs on, which is the same finding as this field's discovery that the building sets the robot's ceiling.

Established Ports are the crossover case where inspection robotics is already operational infrastructure. Gantry-crane inspection by magnetic wall-climbing robot is a port-automation technology as much as an inspection one, and the terminal-automation productivity dispute in Future Ports and Shipping has the same structure as the construction dispute: vendor claims, an independent literature that does not replicate them, and no controlled comparison. Frontier Offshore wind operations and maintenance is the other operational neighbour, and it contributes the finding that the binding resource is a weather window rather than a machine.

Frontier Autonomous freight is adjacent in a way that cuts against it. Closely spaced heavy vehicles in autonomous platoons excite bridges differently from randomly spaced traffic, and channelised autonomous wheel paths concentrate pavement wear rather than distributing it. Speculative Both are recorded here as verified bibliographic records whose content was not read, so no magnitude is stated. The structural point does not need a magnitude: the operating saving accrues to the fleet operator and the fatigue accrues to the highway authority, which is a straightforward misalignment that nobody is currently pricing. Established Printing, prefabrication and modular delivery are assessed in Automated Construction Systems, and structural health monitoring and asset resilience in Infrastructure Resilience. Site instrumentation shades directly into the sensing questions of Smart Cities, and end-of-life disassembly into Circular Infrastructure Systems.

Frontier Certification and liability for autonomous inspection is the institutional question this field has not answered, and this brief could reach no source that answers it. That absence is stated rather than filled. The concrete version of the question is small and hard: when a drone or crawler collects the data, who signs the inspection report, what exactly does the professional engineer's seal cover, and what is the standard of care when the robot's sensor suite did not see a defect that a person on a rope would have felt? Frontier A second version of the question is evidentiary rather than professional: whether a regulator will accept a robot's record as discharging a statutory inspection duty, or only as supporting evidence for a human judgement. Speculative No certification standard, competence framework or liability case for robotic infrastructure inspection was retrievable in the research behind this brief. Naming that gap is more useful than inventing an answer to it.

Frontier The structurally analogous case that is documented is maritime autonomy, and it is instructive because every obstacle there is institutional. International competence requirements for remote operators of autonomous ships have not been introduced, and the gap analysis identifying that is expert elicitation from a focus group of three veteran instructors rather than a measurement, which is worth stating. The collision regulations are written in natural-language seamanship terms — narrow channel, restricted visibility, best aid to avoid collision — with no machine-checkable definitions, so work is under way to quantify qualitative regulatory language. Frontier Salvage law presumes a master who can accept a contract. And the assessment of what it would take concludes that the international community must update the key legal instruments by consensus, which is the rate-limiting step stated politely. Speculative Map that onto inspection and the same four objects appear: a competence standard for the operator, a machine-checkable definition of what an adequate inspection is, an allocation of liability when the machine misses something, and a body willing to move.

Established Adoption in infrastructure is procured, and procurement is where this brief's dependency on Megaproject Governance bites. Nobody buys an inspection robot; an asset owner writes it into a maintenance contract, or a client writes it into a works package, or neither happens. The causal barriers to adoption identified in the DEMATEL work — capital cost, organisational rigidity, interoperability, competency and standardisation — are all properties of the procuring organisation rather than of the machine, and subsidising capital cost alone addresses one of five. Frontier Canadian evidence on building-information-model adoption, verified as a record and not read here, sits in the same tradition of studying why a demonstrably useful technology diffuses slowly through a fragmented client base.

Established The strongest ethical argument for this technology is also its most honest business case: robots go where people should not. Confined spaces, live plant, energised structures, contaminated ground, radiological environments, subsea and at height in weather — these are the conditions in which human inspection has an injury and fatality record, and in which a slow machine with a camera is unambiguously better than a fast person with a harness. Frontier The case for access is much stronger than the case for throughput and is made far less often, because throughput is what sells.

Frontier The accounting for that gain is systematically confused, and the confusion favours the vendor. If a robot performs an intervention within the same weather or plant-state envelope a crew could have worked in, it has transferred risk and produced no additional availability. That is a real gain and it should be booked as a safety gain, not as a productivity gain. Speculative Almost no published case this brief could reach separates the two, and separating them would change which robots are worth buying.

Frontier The instrumentation agenda has a surveillance edge that its advocates rarely name. Automated data capture that detects productivity deviations on a construction site is, mechanically, continuous monitoring of identifiable workers' output. Established That is not a hypothetical objection: the historical precedent invoked most admiringly in this literature, Gilbreth's motion study, belongs to a tradition whose labour history is contested for exactly this reason. Speculative A site instrumentation programme that measures process and a site instrumentation programme that measures people are the same sensors with different aggregation rules, and the rule is a governance choice made at specification time or not at all.

Frontier Displacement is the weakest of the common ethical objections in this field, on the current evidence. The deployments that exist are small, the utilisation is unmeasured, and the trades most exposed — bricklaying, rebar fixing — are in documented shortage across several markets. Speculative The more likely near-term effect is a change in who is employable rather than in how many are employed: robot supervision, programming and BIM competence are the causal-tier barrier the DEMATEL work calls lack of competency, and the workers displaced by a competence requirement are not the same people as those displaced by a machine.

Established The stock matters more than the flow. Global construction output is projected at 22 trillion US dollars a year by 2040, Established but the built environment already standing is the larger object, and it is ageing on a schedule set by decisions taken decades ago. Frontier Whether a society can inspect and maintain what it has built, at what cost and at what frequency, is a more consequential question than how fast it can build more, and it is the question this technology actually addresses.

Speculative The change worth watching is the shift from periodic to condition-based maintenance. Established Periodic inspection exists because inspection is expensive and scarce; it produces a schedule of visits rather than a state of knowledge. Cheap, continuous robotic and fixed sensing would replace a calendar with a condition estimate, and the value of that is not the labour saved on the visit but the failures caught between visits and the interventions avoided where nothing is wrong. Frontier That is a maintenance-regime change, not a productivity change, and it is not captured by any of the productivity statistics this brief has spent its length arguing about.

Handwave At the exotic end sits self-maintaining infrastructure: assets designed around resident machines, with docking, power, reference targets, clearances and serviceable interfaces specified at design time, so that the structure is continuously surveyed and incrementally repaired over a century rather than inspected periodically and rebuilt. Speculative Nothing in the reachable literature attempts this specification, and the strongest argument for it is not robotic capability but the observation that the maintenance liability of the existing stock is growing faster than the labour force willing to service it.

These horizons track two clocks that run at different speeds: the inspection clock, which is a deployment and certification problem, and the construction clock, which is an economics and organisation problem and has not moved in fifty years.

• 10 yr: Frontier inspection robotics consolidates in access-limited niches — confined space, energised plant, subsea, at height — where the case is access rather than throughput, and the first serious throughput comparisons including setup and recovery time are published. Speculative At least one asset owner or regulator issues a competence and evidentiary standard for robot-collected inspection data, most likely in a sector with an existing hard regulator rather than in general construction. Frontier The benchmark correction argued here gets made somewhere in the construction-productivity literature, because it is available to anyone who reads both literatures. Speculative The decomposition of the decline still does not exist at the end of this decade; that is the prediction this brief is most confident about and least happy to make.
• 25 yr: Speculative condition-based maintenance regimes displace calendar-based inspection for at least one major asset class, and the productivity statistics fail to register it because the gain is in failures avoided rather than in output per hour. Speculative Construction robotics establishes itself in the repetitive, well-defined elements — rebar tying, drilling, layout, surveying, finishing — rather than as general-purpose site labour, and the sector's honest description of it is a tool rather than a workforce. Frontier Site instrumentation becomes normal, and its first consequence is a decade of arguments about what the instruments show.
• 50 yr: Speculative a construction productivity series that both camps accept, if links one to three of the workback plan are ever completed; the alternative is that the dispute is still running on 2023 evidence, which is what the previous fifty years suggest. Handwave Structures specified to be robot-legible appear in codes, initially as an access provision rather than as a robotics provision, in the way that fire-service access provisions arrived.
• 100 / 250+ yr: Handwave self-maintaining infrastructure, in the strong sense of assets whose survey and repair are continuous, resident and largely unattended. Speculative The honest observation is that nothing in the current evidence base bears on whether this is a hundred-year problem or a permanent one, and that the obstacle at every horizon examined here has been organisational rather than mechanical.

• Depends on This brief depends on Megaproject Governance, because infrastructure robotics is not bought by robot users. It is procured, and it is procured inside the appraisal machinery that brief assesses. Three things transfer. First, the appraisal problem: an inspection robot's business case is a whole-life cost claim, and whole-life cost claims are made in an environment where outturns are not systematically audited against forecasts, which means the claim is rarely falsified and therefore rarely disciplined. Second, the fragmentation of the client: the causal adoption barriers identified in the barrier literature — capital cost, organisational rigidity, interoperability, competency, standardisation — are all attributes of the procuring organisation, so adoption moves at the speed of client capability rather than of machine capability. Third, the evaluation vacuum: the utilisation number that decides the construction-robotics case is one an asset owner or contractor could publish tomorrow, and the reason it is not published is the same reason megaproject outturns are not systematically published, which is that nobody is required to and the incentives run the other way.
• Enables Delivers three things to the rest of this map. The first is a measurement discipline: any claim that a technology closes a productivity gap must state the benchmark, the period and the deflator, and this brief's benchmark correction shows what happens when it does not. The second is the access-versus-throughput distinction, which applies unchanged to every proposal to automate a hazardous inspection or intervention task, in ports, in energy corridors, offshore and underground. The third is the risk-transfer-versus-availability separation, which decides whether a robotics investment is a safety measure or a productivity measure and which almost no published case distinguishes.
• Adjacent Adjacent to Automated Construction Systems, which owns the printing, prefabrication and modular programmes and carries the firm-level microdata result that reverses the productivity sign, along with the life-cycle assessment finding that printed concrete carries roughly twice the impact of cast concrete per cubic metre; this brief owns the measurement dispute and the inspection layer, and the two should be read together. Adjacent to Autonomous Supply Chains for the warehouse stack now arriving on site and for the simulation-to-validation asymmetry both fields share, and to Future Ports and Shipping for crane inspection and for a terminal-automation dispute with the same evidentiary shape. Adjacent to Infrastructure Resilience, where fixed structural health monitoring competes with and complements mobile inspection, and to Smart Cities on what instrumentation is actually validated against. Adjacent to Energy Corridors and Circular Infrastructure Systems at the two ends of the asset life.

Frontier “Robotic bricklaying is 84 per cent faster with a five-month payback.” That figure is the output of a simulation of ten hypothetical residential buildings under an assumed 30 per cent absenteeism scenario. Nothing was built. The manual baseline of 43.2 days per unit is the number doing most of the work in the ratio, and the abstract does not source it. A payback period computed from a simulated gain on a hypothetical building under an assumed absenteeism rate is a modelling result three assumptions deep. Established The paper states its method plainly; the error is not the authors' but what happens to the figure in transmission, and the correction is to say “in simulation” in the same sentence as the number, every time.

Frontier “Goolsbee and Syverson proved the construction productivity decline is real and not a measurement artefact.” They wrote that measurement error is probably not the sole source. That formulation concedes that measurement error is a source. Their contribution is triangulation across physical measures, materials-to-output efficiency and state-level variation, which makes a pure-measurement explanation untenable; it is not a decomposition and does not claim to be. Speculative A brief that reports this as “definitely real” and a brief that reports it as “all measurement” are misreporting the same sentence in opposite directions, and the first misreading is far more common because it is the one the field wants.

Frontier “Construction is uniquely bad compared with manufacturing.” The manufacturing benchmark in that comparison may itself be overstated by 0.1 to 0.5 percentage points a year through offshoring bias, because import-price declines from foreign sourcing are not captured in standard indexes. At the top of that range the comparator falls from 3.6 to about 3.1 per cent and roughly a fifth of the headline gap disappears. Speculative And the comparison sets a one-off, site-assembled, weather-exposed activity against a repetitive one in a controlled environment. The honest test is against shipbuilding, plant erection or other project industries, and it has not been run.

Frontier “The problem is that construction never automated.” Bricklaying output tripled from 120 to 350 bricks an hour about a century ago, through motion study alone, with no machinery. Construction has achieved a large productivity gain, and it did so by reorganising human work. Speculative The interesting question is not why the sector never automated but why it did not compound that gain — and the answer to that question is organisational, which is bad news for a remedy that consists of buying machines.

Frontier “Inspection drones and climbing robots will replace inspectors.” The published climbing speed for a wind-blade inspection robot is 125 millimetres per second. Against a blade tens of metres long, and once rigging, calibration and recovery are counted on the robot's side of the ledger, the throughput case is weak. Established The strong case is access: robots inspect what people should not, and that is a sufficient justification which does not require the throughput claim to be true.

Speculative “Adoption barriers are essentially about cost.” The causal tier in the DEMATEL analysis contains high initial cost and organisational rigidity, interoperability, lack of competency and lack of standardisation, with market uncertainty appearing downstream as a consequence rather than a cause. Frontier A subsidy programme aimed at capital cost addresses one of five causal barriers and none of the organisational ones, which is a reasonable prediction of why capital subsidies for construction technology tend to underperform their business cases.

Frontier “Construction robotics is all demonstrations.” This is the sceptic's error and it is also wrong. There is commercial wall-laying in residential apartments with an ROI calculation, and rebar-tying field-validated on a named large project. Established The correct critique is narrower and harder to answer: the productivity numbers are simulated and the utilisation numbers are absent. The machines work. What is missing is evidence that they pay.

Speculative “Digital twins and progress monitoring are peripheral to the real work.” If the sector genuinely cannot measure its own productivity, it cannot manage what it cannot measure, and instrumenting the process may be worth more than mechanising the task. Frontier That inverts the usual priority ordering, it is cheap to test against self-reported takt-time compliance, and it has the unusual property of making the productivity debate itself the object of the intervention.

Frontier “Autonomous freight is free money for the infrastructure it runs on.” Platooned vehicles excite bridges differently from randomly spaced traffic, and channelised autonomous wheel paths concentrate pavement wear instead of distributing it. Speculative Both claims are carried here as verified records whose text was not read, so no magnitude is asserted; the asymmetry does not need one. The operating saving is credited to the fleet and the fatigue is charged to the asset owner.

Speculative “Peer review means the numbers in this field are sound.” The research pass behind this brief and its sibling briefs found, in adjacent peer-reviewed 2025 and 2026 literature, a factor-of-1000 unit error in a fuel-property table, a higher heating value set beside lower heating values in the same comparison, and a founding paper cited across a literature for a stronger claim than it makes. Frontier None of these was detectable without arithmetic on the source's own numbers. Handwave The broader inference — that machine-assembled citation chains propagate errors faster than they propagate corrections, and that a review paper's numeric table is now among the least reliable objects in a literature — is not a proven claim, but it is the working assumption this brief has adopted, and every figure above is attributed to what was actually read rather than to what was cited.

Handwave “Self-maintaining infrastructure is a robotics problem.” It is a specification problem. Nothing in the reachable literature has written a design standard for an asset intended to be surveyed and serviced by resident machines over a century — docking, power, reference targets, clearances, serviceable interfaces — and no code contains one. Speculative The strongest argument for the programme is not capability but demography and stock: the maintenance liability of what has already been built is growing faster than the labour available to service it, and that argument, unlike the productivity argument, does not depend on a contested statistic.