Autonomous Supply Chains

An autonomous supply chain is the proposition that a consignment can move from an origin pick to a destination doorstep — warehouse, yard, line-haul, port, deep sea, port, yard, last mile — without a human touching the material flow. Every segment has a plausible autonomy story and a vendor selling it. This brief asks what has been measured, and finds the measured record much stranger and much thinner than the story.

The frame that organises everything here: there are two automations, and they fail for opposite reasons. The decision layer — forecasting, slotting, allocation, sequencing, routing, scheduling — fails by being quietly wrong, and its errors are absorbed downstream as inventory, expedited freight and markdowns. It is cheap, reversible, and it is where most of the measured gains in this literature actually sit. The physical layer — robots, tractors, trucks, hulls — fails by stopping, loudly and locally, and it is a twenty-year commitment in concrete, steel and a floor slab. They share almost nothing: not their failure modes, not their capital intensity, not their reversibility, and not their regulator.

The most interesting finding here inverts the scaling narrative. The one peer-reviewed result that tests warehouse-robot coordination against facility size finds the gain present in small and medium warehouses and absent in large ones, for a stated geometric reason. Section 2 gives that argument properly, because if it holds, the automated megawarehouse is the wrong shape of answer.

Two declared holes, at the front rather than buried. The research behind this brief ran without web search and with arXiv gated — which is where the robotics and autonomy preprint frontier of the last three years actually lives. That frontier is absent here, not summarised. The pack also carries no full-text reads for this subject and twenty-two publisher-deposited abstracts: an abstract evidences what a paper claims, and a verified bibliographic record evidences only that a paper exists. Where a number lives in a body nobody could open, this brief says so. The deep-sea leg, the terminals and the fuels are Future Ports and Shipping; machines on building sites are Robotics in Infrastructure.

Established Start with what the warehouse literature actually measures, because it is almost never labour. Zhao and Zhang, jointly optimising order allocation, sequencing, rack selection and rack sequencing in a goods-to-person robotic mobile fulfilment system with multiple picking stations, report up to 93.3 per cent reduction in computation time against exact optimisation and strategies cutting rack movements by up to 44.8 per cent (Applied Sciences, 2025). Established Yang and colleagues, on order sequencing and rack scheduling in the same architecture, report that it “reduces the robotic tasks by up to 59.8 per cent” (SSRN, 2021). Established Read the denominators: computation time, rack movements, robotic tasks. None is picks per labour hour, and none carries a control group. Frontier El Moussaoui and colleagues, surveying robots, autonomous vehicles and drones in warehouse logistics, find them “positively associated with perceived warehouse operational efficiency, operational accuracy, and warehouse resilience” (Logistics, 2026) — a perception survey, and representative of a large share of the field.

Speculative The one headline efficiency percentage available here is a decision-layer gain wearing a physical-layer badge. Rao, Khan and Wanage report warehousing efficiency rising “20 per cent or more” by combining product demand data with co-ordering affinity patterns, “particularly in facilities employing automated order picking” (Management Practice Insights, 2024). Established The gain is attributed to slotting driven by demand data; automated picking is the context in which it is largest, not its cause. Frontier Quoted as “automated picking delivers twenty per cent” — which is how it will be quoted — the causal claim has been inverted. Handwave It is also the nearest thing to a demand-forecasting result in the retrieved corpus: no forecast-accuracy figure, no inventory-automation trial and no measured bullwhip reduction was obtained for this brief at all. That is a declared hole, not a summary of the field.

Frontier And now the result that inverts the scaling story. Warita and Fujita, studying online planning for autonomous mobile robots on a warehouse commissioning task — posed as spatial task allocation over a Markov decision process, solved with Monte Carlo tree search plus an inter-agent item-exchange method — report that “the achievement rate is improved in small- and medium-sized warehouses. However, the achievement rate did not improve in large warehouses because the average distance from the depot to the items increased” (Information, 2024). Frontier The improvement does not shrink with scale. It disappears. Established And the mechanism the authors name is not algorithmic, which is why it should be expected to travel beyond their method.

Established The geometry, stated plainly. In a goods-to-person architecture the storage medium travels: a robot drives beneath a movable rack, lifts it, carries the whole rack to a picking station at the edge of the floor, and puts it back. Take a facility of area A, roughly square, so its linear dimension grows as the square root of A; expected travel from a fixed station to a randomly located rack grows with it. Established The station does not get faster as the building grows — its rate is set by a human hand, a scanner and a reach envelope, all invariant to the size of the estate behind it. The numerator of every trip grows; the denominator does not. Frontier Three consequences follow arithmetically rather than empirically. Robots per unit of throughput must rise with the square root of area, because each robot completes fewer trips per hour. Frontier Energy per pick rises with it, and is unflattering to begin with, since the mass set in motion is a loaded rack and the mass delivered is one item; a peer-reviewed literature on energy-aware storage assignment in this architecture exists, and this brief can verify only that it exists. Speculative Congestion adds a term the square-root argument omits: a larger fleet in a fixed aisle topology interferes with itself, so effective travel time grows faster than distance does. That term is this brief's extension, not the paper's finding.

Frontier Why this is not a quirk of one simulation. The intuition that automation favours size is imported from software, where the marginal cost of serving one more unit is near zero. Established In a warehouse the marginal storage location is not free: it is by construction further from the station than every location already in use, and it lengthens every trip that touches it. Speculative Any architecture in which the stored medium moves to a fixed service point inherits that penalty, which is why this should generalise past one planner and one paper. Frontier The field's own answer is visible in its formulations — Zhao and Zhang's system is explicitly multi-station, and distributing stations does bound mean travel, but only if station density is held constant as area grows. Established Constant station density means picking capex scaling linearly with floor area rather than being amortised across it, at which point the economy of scale has been spent buying the geometry back. Speculative The hypothesis is cheaply falsifiable with data operators already hold: if gains invert with size, the rational answer is more small nodes rather than fewer large ones. Track mean node size in fulfilment networks over time. If it is falling while node count rises, the industry has conceded this argument without stating it. Handwave Nobody publishes that series.

Established Against all that simulation, the peer-reviewed physical deployment count in this corpus is three. Garcia, Montaña and Montés report “three Survival 1.0 AIMWs in operation in the factory at Ford España S.L. in Almussafes” — an autonomous industrial mobile warehouse that makes part of the central store itself mobile (Applied Sciences, 2024). Established Three units, one plant, named. Frontier Vendor and analyst counts are not comparable artefacts: they generally count units shipped rather than units in continuous revenue service, and they never report units removed. Handwave Installed-base statistics were unreachable here, so none appears in this brief.

Established Autonomous freight: three published platooning fuel savings, and all three are simulations. Liu and Meidani, building surrogate models for drag and fuel consumption over a 161-kilometre corridor of Interstate 57 in Illinois, report “average fuel savings achieved can be up to 10 per cent, depending on the headway between the trucks”, with delivery cost down 30 per cent against conventional line-haul (Illinois Center for Transportation, 2023). Established Chen and colleagues, applying reinforcement-learning platooning control in on-ramp regions, report that “the average energy consumption decreased by 14.8 per cent, and the road occupancy rate decreased by 43.3 per cent” (World Electric Vehicle Journal, 2023). Frontier Huang reports “fuel savings ranging from 10 per cent to 20 per cent” (2024). Established Not one of the three is an instrumented on-road measurement in revenue service with a control group. The apparent convergence is a convergence of modelling assumptions — principally about following distance — rather than of evidence.

Established And the load-bearing condition is the one always lost in transmission. Liu and Meidani say it in the abstract: depending on the headway. Established The aerodynamic benefit of a platoon is a function of the gap between vehicles and grows as that gap closes, and the gap between heavy vehicles on a public highway is not set by aerodynamics. It is set by following-distance law, by the liability regime behind it, and by what a road authority will permit. Frontier Platooning is therefore an unusually clean case of a technology whose headline benefit is written in a rulebook rather than measured in a wind tunnel: the same hardware, the same control stack and the same corridor deliver a different saving under a different statute. Speculative The published figures are less forecasts of what platooning will deliver than statements of what a permissive regulator would be worth, and no source here states its headway assumption in a form that lets a reader convert between the two.

Frontier The economics of driverless freight do not rest on fuel, and this is where the literature stops. A ten to twenty per cent fuel saving is not what an autonomy stack is bought for; the driver is, and every deployment retaining a safety driver has added cost while removing none. Established The economic variable is remote operators per vehicle in revenue service — the supervision ratio, not the disengagement rate — and a system with excellent disengagement statistics and one operator per vehicle has removed no labour cost whatever. Handwave No published figure for that ratio could be located, and the assessment behind this brief is that none exists outside operators' confidential data. Speculative Worse, the two metrics move in opposite directions as the system improves: a 2023 Journal of Safety Research paper addresses sleepiness and takeover performance in prolonged automated driving, and a 2024 Applied Ergonomics paper is titled “Once a driver, always a driver”. Both were verified as bibliographic records only, so this brief states their subjects and not their effect sizes — but the structure stands on its own logic: the better the automation, the less practised the supervisor, and supervised autonomy prices its safety case on that supervisor. Frontier That validation is unsolved is visible in the shape of the literature: Guo and colleagues argue that test cases drawn from naturalistic driving must satisfy both “representativeness” and “coverage, capturing high-risk corner cases” (Research Square, 2025) — criteria that pull against each other, since a representative sample under-weights the rare cases that kill.

Established The deep-sea leg is not blocked on autonomy. It is blocked on law. Fenton and Chapsos, combining regulator interviews with a practitioner survey, conclude that safe integration of maritime autonomous surface ships requires that “the international community needs to work together, and update by consensus the key legal instruments and policy documents” (Frontiers in Computer Science, 2023) — the crewing obstacle stated politely, since consensus is the rate-limiting step. Established Yoshida and colleagues find international competence requirements for remote operators have not been introduced, and build a goal-based gap analysis against the STCW convention (Applied Sciences, 2020); their case study is a focus group of three instructors — elicitation, not measurement. Frontier Kim notes that even a fully autonomous vessel “can effectively and legally avoid approaching ships only when they are steered in compliance with the COLREG”, and attempts to quantify the convention's qualitative terms — “narrow channel”, “restricted visibility”, “best aid to avoid collision” (JMSE, 2023). Frontier Suri argues the Salvage Convention 1989 does not cleanly accommodate autonomous ships (SSRN, 2022): salvage law assumes a master who can accept a contract, and removing the master removes the counterparty for rescuing a drifting hull.

Established And the finding this brief inherits rather than restates: automation does not necessarily help. Future Ports and Shipping carries both independent results — a twenty-port data envelopment analysis concluding that automation alone does not necessarily raise efficiency unless integrated with training and gradual investment, and a Ravenna terminal simulation in which the automated electric tractor configuration increased unloading time. Frontier Both are load-bearing here, because the yard is the seam between warehouse and ship. Established The honest form of the claim is that the productivity case for terminal automation is unproven in the independent literature, not that automation makes terminals worse.

Frontier The one road segment with a public safety record at scale is the passenger one, and it is worth importing precisely because freight has nothing comparable. Driverless passenger services now operate without a safety driver in several American cities, and the leading operator has published crash comparisons against human benchmarks at successive mileage milestones — first at around 7 million rider-only miles, later at tens of millions — reporting large reductions in airbag-deployment and injury-involved crash rates against a matched human baseline. Frontier Every one of those studies is authored or funded by the operator, which is not disqualifying and is the first thing a reader should be told. Established The methodological load sits in the benchmark: a human comparison rate has to be matched to the same cities, the same road types and the same severity threshold, and police-reported human crashes are known to be under-counted at the minor end, which biases the comparison in the direction the operator would prefer. The result is credible and the uncertainty is one-sided, and both halves belong in the sentence.

Established The operational design domain is the product, and it is what makes a mileage record unreadable as a capability claim. The industry taxonomy defines an automated-driving feature only in terms of the conditions it is declared to handle — road types, speeds, weather, lighting, geography — so miles driven are a statement about a domain rather than about competence. Frontier Rider-only services were built first in low-rainfall, grid-planned, snow-free metropolitan areas at surface-street speeds; extension to freeway speeds and to winter is the announced direction of travel and the thing to check rather than assume. Established The domain is also an infrastructure claim: lane markings, sign legibility, work-zone layouts and road geometry are maintained by public authorities to standards written for human perception, and a fleet inherits whatever those authorities fund. Frontier In mixed traffic the interaction runs both ways — published collision breakdowns are dominated by the autonomous vehicle being struck, frequently from behind, which is what a more conservative machine inside a human traffic stream should produce. Handwave Whether that is a safety property or a friction cost that decays as human drivers adapt is established by nothing published.

Established And here the passenger segment supplies the number section 2 says the freight literature does not have. Rider-only fleets are not unsupervised: they run behind remote-assistance desks whose staff answer the vehicle’s questions — confirming a path around an obstruction, authorising a manoeuvre — rather than driving it. Frontier Operator statements made during the 2023 San Francisco controversy put remote contact at roughly one intervention every few miles and staffing at a small number of vehicles per remote operator. Those are company disclosures under press pressure rather than audited figures, and they are the only quantitative supervision ratios in public anywhere in road autonomy. Frontier Read against this brief’s freight argument they are the most useful numbers here that this brief did not generate: they establish that the ratio is finite, that it is above one vehicle per operator, and that nobody reports it on a cadence. Speculative If passenger fleets cannot lift the ratio into the tens, the labour case for driverless line-haul is weaker than its promoters’ arithmetic, because a truck’s remote desk handles a heavier vehicle at higher speed with a longer stopping distance.

Frontier The most telling frontier result is a retreat. Sousa and colleagues develop obstacle avoidance for mobile robots in human-robot collaborative warehouse environments, fusing fuzzy logic with a convolutional network to adjust velocity and yaw continuously around people and moving objects (Sensors, 2025). Established The design target is a floor with humans on it. Speculative A research programme optimising for coexistence has stopped expecting the humans to leave, and that is evidence about the lights-out thesis rather than an unrelated engineering advance.

Frontier The building may set the robot's performance ceiling, and this is genuinely under-appreciated. Cho and colleagues review how floor finishing materials affect autonomous mobile robot performance, deriving five sensing relationships — colour on obstacle detection, texture and transparency on detection and mapping accuracy — and five locomotion relationships including slipperiness and unevenness on speed and path consistency; they describe existing knowledge as “mostly anecdotal and remains fragmented” (Buildings, 2026). Speculative If a meaningful fraction of the existing estate needs floor replacement or relighting before robots work in it, warehouse automation is a new-build phenomenon and the installed-base transition is far slower than deployment announcements imply. Handwave The fraction is unpublished, and it is the cleanest quantitative link between this brief and Robotics in Infrastructure.

Frontier For outdoor and yard robotics the binding constraint appears to be radio, not autonomy. Baruffa and Rugini treat resource assignment for mobile robots with task offloading as latency minimisation across radio links and computation nodes (Future Internet, 2025). Speculative This is the least durable obstacle in the topic: if onboard compute closes the gap within a hardware generation, the constraint dissolves. Speculative What decides it is watts and dollars of onboard compute needed to eliminate offload, set against the cost of guaranteed-latency coverage across a yard.

Frontier Planning is improving, and the improvements are reported in the units of the laboratory. Yilmaz and Kivrak present orientation-aware path planning via phase-portrait dynamics with zero final orientation error, path lengths comparable to Hybrid A*, and planning times reduced 52 per cent on an indoor map and 84 per cent on a playpen map across 28 scenarios (Inventions, 2025). Speculative Deng and Chen claim a 99 per cent navigation success rate in simulation against self-selected baselines, in a preprint (2025). Established Ninety-nine per cent is a one-in-a-hundred failure rate, and a facility performing a hundred thousand moves a day at that rate produces a thousand failures a day. Frontier Autonomy claims in logistics are unreadable as success percentages and should be quoted as failures per operation-day.

Frontier Two architectural frontiers change the shape of the problem rather than the performance of a component. The autonomous industrial mobile warehouse makes the store itself mobile, eliminating the fixed central warehouse instead of automating travel within one — which, read against section 2's geometry, is an attack on the right variable. Frontier And the warehouse stack is being ported outward: Wang, Yi and Tan optimise on-site building material handling with construction robotics, using variable neighbourhood search and particle swarm optimisation for task distribution, routing and team composition (Automation in Construction, 2026). Speculative The methods travel more easily than the machines do.

Frontier On the water, the frontier is translating law into something a machine can check. Hwang and Youn build collision-avoidance validation scenarios from twelve months of AIS data in the west sea of Korea, clustered by navigation situation, and extend that into a graph model of how simple collision-risk situations evolve into harsh ones (JMSE, 2021 and 2023). Frontier Observed traffic replaces arbitrary or convention-derived test cases — the maritime version of the sampling problem the road literature is stuck on. Speculative If the COLREG-quantification programme succeeds it is the single most important enabling step for deep-sea autonomy; if it fails, autonomy stays in coastal and inland waters permanently and the end-to-end thesis loses its longest segment.

Speculative And the most valuable paper in this subject has not been written. Removing the crew removes the accommodation block, the hotel load, the lifeboats and the freshwater plant, and frees hull volume — which is precisely what a low-energy-density fuel needs. Speculative A policy chapter on the autonomy-decarbonisation nexus exists, verified as a record and not read here; no naval-architecture study runs the calculation. Frontier The calculation is specific: for one reference hull, compare cargo capacity of uncrewed-plus-alternative-fuel against crewed-plus-alternative-fuel against crewed-plus-conventional. Speculative If the first beats the third, autonomy and decarbonisation become easier together than either is alone. Handwave This brief prints no fuel density to support that, deliberately: the cluster's corrections register found a factor-of-1000 unit error in the fuel-property table of a 2026 peer-reviewed review, which is exactly the table such a calculation would be built on. Those numbers belong in Future Ports and Shipping, with their heating-value basis stated.

Frontier The passenger segment’s real frontier is not the driving. It is whether a regulator can read the evidence. The United States collects crash reports from automated-driving and driver-assistance systems under a standing general order and publishes them, and the resulting file is close to unusable for comparison: firms differ in how much telematics they can see, narratives are redacted as confidential business information, and a company with better sensing reports more incidents. Established California’s disengagement reports carry the same defect from the other side — operators choose where and when to test, and the definition of a disengagement is theirs. Frontier What would settle it is a reporting rule that specifies exposure denominators by road type and domain rather than collecting incident counts, which is a drafting problem rather than a research problem. Speculative The commercial withdrawal of one major American operator after a single 2023 injury and the suspension of its permit is the clearest evidence available that this sector’s binding risk is licence and reputation rather than crash rate.

Established The workback plan to an end-to-end autonomous chain has twelve links, and the one that binds is not an engineering link. Target state: a consignment moves origin pick to destination delivery with no human intervening in the material flow, under remote supervision far above one-to-one.

Frontier Links one to seven are engineering and permission, and all show incremental movement. A warehouse pick with a bounded exception rate, measured as exceptions per thousand picks and the fraction requiring a human to walk to the location. Frontier Robotics gains that survive scale-up, measured as mean rack-travel distance per pick as facility area grows. Frontier An existing building stock that tolerates robots, measured as the share of estate needing floor or lighting works and its cost per square metre. Established Yard and terminal automation beating a manned baseline, measured as moves per hour with a manned control group — the test the independent port literature says has not been passed. Frontier Driverless line-haul that is legal, measured as lane-miles open to an unoccupied heavy vehicle by jurisdiction. Frontier A supervision ratio far above one-to-one. Frontier And an end-to-end driver-hour reduction rather than a middle-mile one, measured as driver-hours per tonne-kilometre including drayage at both ends.

Speculative That last link deserves separate attention because it is routinely skipped. A hub-to-hub autonomous middle mile with human drayage at each end may remove well under half the driver hours in a movement — the difference between a transformative business case and a marginal one, settled by an accounting boundary rather than by a technology.

Established Links eight to eleven are legal. Machine-checkable collision regulations, evidenced by a classification society certifying an avoidance algorithm rather than publishing a paper about one. Established A remote-operator competence standard, evidenced by a recognised certificate and a count of holders. Frontier A flag state willing to register an uncrewed hull for international voyages, evidenced by hulls on a register. Frontier An insurable casualty and salvage framework, evidenced by a protection-and-indemnity club writing cover and stating a distance-from-shore envelope.

Speculative The flag-state link binds, and binds first, because it is the only one that is a pure sovereign decision with no technical precondition. Everything else needs consensus, a standard or a demonstration; a flag state can act alone and cheaply. Speculative That does not make the chain short. It makes its timeline unforecastable rather than merely long, which is a different and more awkward property, and insurance follows registration within a few years once registration happens.

Frontier Link twelve is the spoiler: the handoffs. Even with everything above complete, a chain whose human interventions are dominated by customs, documentation and exception handling is not autonomous in any sense a shipper would recognise. Handwave The honest metric for this subject is interventions per consignment, located by point in the chain, and nobody publishes the trace. A brief that counts segments will overstate progress; this one tries not to.

Established Liability is the passenger-road bottleneck that freight will inherit, and it is being resolved by statute rather than by litigation. The United Kingdom’s automated-vehicles legislation of 2024 creates named entities — an authorised self-driving entity answerable for the vehicle’s behaviour, and a licensed operator for services with no user in charge — and moves responsibility off the person in the seat while a self-driving feature is engaged. Established Germany’s 2021 statute permits driverless operation inside approved areas under a technical supervisor, and the United Nations regulation on automated lane keeping sets a type-approval envelope that was raised from 60 to 130 kilometres per hour. Frontier None of these has been tested by a contested fatality with a disputed sensor log, which is the event that will price the risk. Handwave Reading the existence of a liability statute as evidence that liability is settled is the routine error.

Established This brief waits on one other brief for a result it cannot produce itself. Whether automated container handling outperforms manned handling on a common measure is the hinge between the warehouse and the ship, and it lives in Future Ports and Shipping. Frontier That brief's answer is currently negative-to-unproven, and every end-to-end autonomy claim inherits it.

Frontier Three dependencies are on measurement disciplines that do not yet exist rather than on results. A published pairing of simulated against realised performance for the same installation, which would make the simulation literature informative rather than decorative. Established A convention for reporting autonomy reliability as failures per operation-day at a stated operation count. Frontier And a common accounting boundary for driver-hours that includes drayage, without which every line-haul saving is reported against a boundary chosen by the party reporting it.

Speculative One dependency runs the other way, from law to engineering. Quantifying the collision regulations is a legal-drafting problem currently being attempted by naval architects, and it will not be finished by them; the same is true of a remote-operator competence standard. Better sensors unblock neither: no amount of capability closes a gap whose content is a definition.

Handwave And a dependency this brief must record as unmet in its own construction. The robotics and autonomy preprint literature of the last three years — where this subject's frontier actually sits — was unreachable when this brief was assembled, along with public intervention datasets and installed-base statistics. Everything above is drawn from journal records and publisher-deposited abstracts. Treat section 3 as a lower bound on the field, not a survey of it.

Established Six experiments would settle most of what is disputed here, and five of them are cheap. They are listed in order of value per dollar, not of difficulty.

Frontier One: picks per labour hour, before and after, same SKU mix and order profile, with a control facility. This is the measurement the entire warehouse-robotics case rests on, it is trivially available to any operator, and it is essentially unpublished. Speculative Its absence is informative: an operator holding a favourable number would publish it.

Frontier Two: mean rack-travel distance per pick as facility area grows, holding station density constant. This tests the geometric inversion directly, distinguishes it from an artefact of one planner, and needs only telemetry every fleet already generates. Speculative Its network-level companion is mean fulfilment node size over time, which tests whether the industry has already conceded the point in its capital plans.

Frontier Three: an instrumented on-road platooning trial in revenue service, with a control group and a stated headway. Every published saving in this brief is a model output. Established The trial must report fuel per tonne-kilometre at two or more headways, so the regulatory sensitivity is visible rather than buried in an assumption.

Frontier Four: the exception trace. For a single physical consignment, count every human intervention from origin pick to destination delivery and record where each occurred. Speculative The expectation this brief holds and cannot verify is that the count is dominated by handoffs, customs and exceptions rather than by movement.

Frontier Five: the estate survey. Sample existing warehouse floors for the sensing and locomotion properties the finishing-materials review identifies, and price the remediation. Speculative That converts an anecdotal engineering complaint into a capital number, and answers whether automation is a retrofit or a new-build technology.

Speculative Six, the expensive one: the uncrewed low-carbon hull comparison. One reference vessel, three arrangements, cargo capacity and delivered cost per tonne-mile for each. Handwave It is a naval-architecture study rather than a policy chapter, and it does not appear to have been done.

Established The engineering requirements here are unglamorous and mostly about the environment rather than the machine. Floors first: flatness, slip resistance, wheel-mark resistance, and optical properties that do not defeat the sensor — colour, texture and transparency all bear on detection and mapping accuracy, and the review establishing this calls the existing knowledge anecdotal and fragmented. Speculative A specification for a robot-ready floor slab does not appear to exist in a form a contractor could build to, and writing one is a standards job, not a research job.

Frontier Second, guaranteed-latency coverage across yards and outdoor sites, or enough onboard compute to make it unnecessary. The published framing is latency minimisation across radio links and computation nodes; the engineering choice underneath is whether to buy radio infrastructure per site or silicon per robot. Speculative Silicon per robot is a falling cost and radio coverage of a paved yard in rain is not, which is a reason to expect this constraint to be designed out rather than solved.

Established Third, energy and mass. A goods-to-person system moves a loaded rack to deliver one item, so energy per pick is set by the mass of the storage medium rather than the goods. Frontier Under the geometric argument that cost grows with the square root of floor area, which makes energy per pick an architectural decision taken years before anyone measures it.

Frontier Fourth, exception handling as designed hardware rather than a person on call. A facility running a hundred thousand moves a day at a one-per-cent failure rate needs a designed recovery path for a thousand events, and staffing is then set by coverage and response time rather than by workload. Speculative That is a queueing requirement with a floor, and it is why the residue does not scale away.

Speculative Fifth, interfaces at the seams. Pallet, container, trailer and hull each have their own handling geometry, and the transitions between them are where the chain currently employs people. Handwave No standard exists for a machine-legible consignment handoff spanning all four.

Established The nearest neighbour is the port. Everything in Future Ports and Shipping about terminal automation, shore power and alternative fuels sits directly upstream and downstream of this brief, and the two share the yard as a boundary object.

Frontier Construction is a closer methodological neighbour than most readers expect. The warehouse stack — task distribution, routing optimisation, team composition for autonomous equipment — is being applied to on-site material handling, and the finishing-materials work that determines robot performance is published in a buildings journal rather than a robotics one. Speculative Both seams point at Robotics in Infrastructure and Automated Construction Systems, and the traffic runs both ways.

Speculative Road infrastructure is adjacent in the load-bearing sense. There is a literature asking whether closely spaced heavy platoons excite bridges differently from random traffic, and whether channelised autonomous wheel paths change pavement wear; this brief verified those records exist, could not read them, and so states the question and no result. Frontier The structure is nonetheless clear: the fuel saving accrues to the carrier and any structural consequence accrues to the road authority, which makes platooning an externality problem as well as an aerodynamics one. The corridors themselves are Continental Transportation Systems and Intercontinental Rail Systems; rail's own automation record is absent from this brief's evidence base and is not supplied from memory, and the same is true of last-mile delivery robotics.

Frontier The decision layer's neighbours are economic rather than mechanical. Forecasting, slotting and inventory optimisation belong with AI-Driven Productivity, where the measurement problem is the one section 2 identifies here: the gain is real, the denominator is chosen by the party reporting it, and labour is rarely the denominator.

Established Four institutions decide whether this subject has a future, and none of them is a robotics laboratory. A flag registry, a maritime standards body, an insurer, and whoever writes following-distance rules for heavy vehicles.

Frontier The flag registry is the most consequential and the least watched. Registering an uncrewed hull for international voyages is a sovereign act requiring no consensus, no new technology and very little money. Speculative The prediction this brief makes, and offers as a near-term testable indicator, is that deep-sea autonomy will arrive, if it arrives, through a single permissive flag state rather than through multilateral instrument reform. Speculative That is a watchable signal: a register entry is public, and it would precede any change in the international framework by years.

Established The standards gap is specific and stated in the literature. International competence requirements for remote operators have not been introduced, and a goal-based gap analysis against STCW exists precisely because the convention assumes a watchkeeper aboard. Frontier A certificate and a count of holders would be evidence that this link had moved; neither exists.

Frontier Insurance is the quiet gatekeeper. Salvage law presumes a master who can accept a contract, and an uncrewed drifting hull has no counterparty. Speculative The operational limit on autonomous shipping will therefore be expressed not as a capability envelope but as a distance-from-shore envelope inside which an underwriter will write cover — a number an insurer could state today and nobody has asked for. Frontier Cybersecurity enters the same conversation as a first-class regulatory constraint rather than an IT concern; a book chapter treats it at that level, verified as a record here and not read.

Established On the road, the institutional variable is following distance. Because the platooning benefit is a function of headway and headway is a legal quantity, the regulator is setting the technology's return on capital directly. Speculative A jurisdiction could double the published saving with a statutory amendment and no engineering at all, and could equally zero it. Frontier The second road-side variable is lane-miles legally open to an unoccupied heavy vehicle — the only permission that matters to the business case, and the only one measurable in a public register.

Established The labour question here is not mainly about job counts. It is about what happens to the jobs that remain, and the field has begun to consolidate that critique: a 2026 monograph on warehouse automation and the dehumanisation of labour exists, verified as a record and not read, which marks the point at which scattered commentary becomes a literature.

Frontier The specific harm this brief can evidence is the supervisor's position. Supervised autonomy asks a person to stay alert through long uneventful periods and take over in the rare moment the machine cannot handle — the exact task human attention is worst at. Speculative The published records point at sleepiness under prolonged automation and at manual habits persisting into takeover; this brief cannot quote their effect sizes, and does not need to in order to say that the safety case for supervised autonomy is priced on a human capability that improves as the automation gets worse.

Speculative The exception workers are the ethically distinctive category. Where automation handles the routine, the human residue is by definition the awkward, the damaged, the mis-scanned and the urgent, arriving at machine tempo. That is a job with the variety removed and the difficulty retained, and it is created by successful automation rather than by failed automation. Handwave No study retrieved for this brief measures it.

Frontier Two distributional points follow from the technical findings above. If the geometric argument holds, automation's advantage is largest at medium scale, which cuts against the assumption that automation necessarily concentrates an industry — a competition-policy conclusion drawn from a travel-distance argument. Speculative And if platooning's benefit is carrier-side while its structural consequences are public-side, the technology transfers value from road authorities to freight operators by default unless someone prices it. Handwave Both claims are structural, neither is measured, and both are cheap to test.

Speculative The strongest civilizational reading of this evidence is that logistics automation decentralises rather than concentrates. If gains invert with facility size, the efficient network is more nodes, each smaller and nearer the customer, and the automated megawarehouse is a category error rather than an endpoint. Speculative That is close to the opposite of the picture the subject usually evokes, and it has a visible signature in mean node size that anyone with the data could check.

Frontier A supply chain with fewer people in it is not obviously more resilient. Automation replaces a workforce that improvises with a system that has a designed response set, and the events that break supply chains — a closed strait, a cyber incident, a customs regime changed overnight — are precisely those outside the designed set. Speculative The resilience literature retrieved for this brief measures perceived resilience, which is not the same thing and should not be read as it.

Speculative The longest-range consequence is jurisdictional. If deep-sea autonomy arrives through a single permissive flag state, the effective law of autonomous shipping will be written by whichever small registry moves first, and everyone else negotiates with a fait accompli. Handwave That is how a good deal of maritime law already works, and it is a reason to watch registries rather than conferences.

Handwave The exotic terminus is a chain with no person in it anywhere — pick, load, haul, tranship, sail, discharge, deliver, all under remote supervision measured in hundreds of vehicles per operator. Nothing retrieved is evidence for it. What the literature does support is a narrower and stranger claim: that such a chain may be available now for a sufficiently uniform catalogue, and never for a general one.

These horizons track institutional permissions and published measurements rather than demonstrations, because in this subject the demonstrations have consistently run ahead of both.

• 10 yr: Frontier Warehouse automation keeps spreading through new-build estates and stalls in old ones, with human-robot collaborative floors the dominant design rather than lights-out. Frontier Hub-to-hub driverless line-haul operates on a growing but jurisdictionally patchy set of corridors; the supervision ratio stays unpublished and is the thing to ask for. Speculative An instrumented on-road platooning measurement with a control group finally appears, and lands below the simulation range. Speculative Picks per labour hour remains unpublished, and the absence continues to be the most informative fact in the sector. Frontier On passenger roads expect domain extension — freeway speeds first, weather later — to be announced faster than it is independently verified, and expect the remote-assistance ratio to stay a disclosure made under pressure rather than a reported metric.
• 25 yr: Speculative A remote-operator competence standard exists and has holders, and at least one classification society certifies a collision-avoidance system against quantified collision regulations. Speculative One or more flag states register uncrewed hulls for international voyages; insurance follows within a few years, expressed as a distance-from-shore envelope rather than a capability claim. Frontier Fulfilment networks have visibly more and smaller nodes, and nobody credits the geometry. Speculative The decision layer, by then unremarkable, has delivered more measured value than the physical layer and is not called robotics.
• 50 yr: Speculative End-to-end movements with no human in the material flow are routine for uniform, high-volume, low-variance catalogues — bulk, fuels, standardised components — and out of reach for general merchandise. Speculative The residual workforce is concentrated at handoffs, exceptions and customs, which is the opposite of where automation began. Handwave Interventions per consignment is a published metric, because a regulator required it rather than because an operator volunteered it.
• 100 / 250+ yr: Handwave A logistics system in which the exception rate has been engineered away by standardising what is shipped rather than by improving the robots that ship it — packaging and product converging on machine-legible forms. Handwave That is a manufacturing and design intervention wearing a logistics label, and it is the only route in view by which the last few per cent stop mattering.

• Depends on This brief depends on Future Ports and Shipping for a result it cannot produce itself: whether automated terminal and yard handling outperforms manned handling on a common measure, with a control group. The independent evidence there is currently negative-to-unproven, and the yard is the seam every end-to-end autonomy claim has to pass through, so the sign of that result propagates into every segment count made here. The same brief holds the fuel-property and volumetric arithmetic that any uncrewed-hull calculation would need, and holds it under a correction notice after a factor-of-1000 unit error was found in the review that would otherwise be its natural source. That is why no fuel density appears in this brief, and why the complementarity argument in section 3 is posed as a calculation to run rather than a result to cite.
• Requires (not on this map) Three things nobody has to discover. A crash database comparable across operators: the American standing order already collects incident reports from automated and assisted driving, but firms differ in the telematics they can see, narratives are redacted as confidential business information, and there is no exposure denominator by road type or domain, so the file cannot support the comparison it exists to support. A liability rule that names the responsible entity: the British statute of 2024 and the German statute of 2021 both do this for passenger vehicles, and no equivalent exists for a driverless heavy vehicle in most of the jurisdictions its corridors would cross. And a supervision ratio published for revenue service: the only figures in public came out of one operator under press pressure, and the number decides whether autonomy in freight is a labour saving or a fuel saving. None of the three is a research result, and all three are the difference between a demonstration and a business case.
• Enables What this brief would enable, if its links landed, is narrower than the subject's rhetoric suggests. A bounded warehouse exception rate combined with a robot-tolerant building stock would make automation a retrofit rather than a new-build technology, which changes the transition rate for the installed estate rather than the capability of the technology. A supervision ratio far above one-to-one would, for the first time, put a labour saving rather than a fuel saving into the autonomous-freight business case, and it is the only development that would make the case transformative rather than marginal. A flag state registering an uncrewed deep-sea hull would enable the longest segment on a timescale nobody can forecast, because the decision is sovereign, cheap and unilateral. And a published intervention trace would give the whole field its first honest end-to-end metric.
• Adjacent Adjacent work runs in both directions with construction robotics, which is importing this field's task-allocation and routing methods while exporting the finding that the building sets the robot's performance ceiling. Road infrastructure is adjacent in the structural sense: platoons are a different excitation from random traffic and channelised wheel paths a different wear pattern, questions this brief can name and not answer. The decision layer belongs with the productivity-measurement literature, where the denominator problem identified here appears under a different name and with the same consequences.

Established “Warehouse robots deliver X per cent productivity gains.” The peer-reviewed measurements available here report computation time, rack movements, robotic tasks and perceived efficiency. Picks per labour hour with a control group is essentially absent. Frontier When a productivity percentage for warehouse robotics is quoted, the question that resolves it is what the denominator was, and the answer is usually not labour. Speculative That is not an accusation of dishonesty; it is what happens when a field publishes what is easy to simulate and an industry quotes what is easy to sell.

Established “Automation scales — the bigger the facility, the bigger the gain.” The one study here that tests it finds no improvement in large warehouses, because depot-to-item distance grows. Established The scaling intuition is imported from software and does not survive contact with a geometry in which the storage medium physically travels. Frontier It is a single study and should be replicated; the mechanism it names, however, is not a property of that study.

Established “Truck platooning saves up to twenty per cent fuel.” Three sources give ten per cent, 14.8 per cent and ten-to-twenty per cent, and all three are simulation or surrogate-model outputs. None is an instrumented on-road measurement in revenue service with a control group. Established The largest savings occur at following distances that regulation, not aerodynamics, sets — the source giving the ten per cent figure says so explicitly, in the phrase “depending on the headway”. Frontier Both conditions are gone by the time the number reaches trade press, and the second is the load-bearing one. Speculative A platooning fuel saving quoted without a headway assumption is not a measurement of a technology; it is an unlabelled statement about a legal regime.

Frontier “Autonomous trucks are nearly ready — look at the disengagement rate.” Disengagements per mile is the wrong metric for a freight business. Established The economic variable is remote operators per vehicle in revenue service, and a system with excellent disengagement statistics and one operator per vehicle has removed no labour cost at all. Handwave No published figure for that ratio was located, and the assessment behind this brief is that none exists outside operators' own data. Speculative Compounding it, supervisor performance appears to degrade as automation improves, so the two headline metrics move in opposite directions as the system matures.

Established “Deployment counts prove maturity.” The verifiable count in the peer-reviewed literature retrieved here is three units, at one plant, named. Frontier Vendor and analyst counts are not comparable artefacts: they generally count units shipped rather than in continuous revenue service, and never report units removed. Speculative A field with a large claimed installed base and no published retirement rate is a field whose survival curve nobody has seen.

Established “Lights-out warehouses are the direction of travel.” The visible research frontier is human-robot collaborative floors, which is a design commitment to keeping people present. Speculative And the strongest form of the sceptical case is not that the last few per cent are hard. It is that the last few per cent are where all the cost lives, for a structural rather than a technical reason. If some fraction of operations needs a human, staffing is set by coverage — someone present, on shift, within a response time — not by workload. Speculative Labour cost is then a step function of the exception rate rather than a linear function of it, and the last person on the floor costs what the first ten did. Established At a hundred thousand moves a day, a 99 per cent success rate is a thousand exceptions a day: not a residue, a department.

Frontier The best counter to that, and this brief thinks it genuinely strong. The exception rate may be a function of catalogue heterogeneity rather than of automation maturity. Speculative A facility handling one product geometry, one packaging and one weight class could plausibly go lights-out today, while a general-merchandise facility never can. Speculative If that is right the debate is misframed: the question is not when lights-out arrives but what it arrives for, and the answer is decided by the product catalogue rather than by the robot. Handwave The test is whether exception rate correlates with SKU count and packaging variance or with years since automation, and it would take one operator one afternoon.

Established “Autonomous ships are a technology question.” They are a registration, certification, liability and salvage question. Established The autonomy stack for coastal operation broadly exists; the international competence standard for the operator ashore does not, the collision regulations have no machine-checkable definition, salvage law has no counterparty without a master, and no flag registry entry exists for a deep-sea uncrewed hull, and none of those four is unblocked by a better sensor. Established Nor is a success percentage readable on its own: autonomy reliability in logistics has to be quoted as failures per operation-day, at the operation count of a real facility, and that should be a convention rather than a note on one preprint.

Frontier And the sceptical overcorrection, stated because this brief has spent most of its length making the sceptical case. Several results here are solid and unglamorous: substantially fewer rack movements, substantially fewer robotic tasks, materially lower road occupancy from platooning. Established The correct critique is that the gains are in the wrong variables to justify the capital, not that they are fictitious. Speculative A brief that overstates the sceptical case will be as wrong as the vendors, and the evidence for that is that the corrections in this cluster do not all point one way — one of them makes a technology look considerably better than the flawed source implied.

Handwave Finally, a fact about how citations decay, which belongs in a brief built the way this one was. Four independent research passes in this project caught their own bibliographic tools silently substituting one document for another — a lookup service returning unrelated records, an abstract attached to the wrong paper, a correctly addressed identifier serving a different article. Established No error, no warning, no low-confidence signal. This cluster's own register additionally found a factor-of-1000 unit error in a fuel-property table in a 2026 peer-reviewed review, caught only by its inconsistency with the same paper two lines away. Speculative Machine-assembled scholarship fails silently, and a supply chain assembled from machine-read documents will fail the same way. Handwave That is not a metaphor about this subject. It is the same failure mode, in the layer of the chain that decides rather than the layer that lifts.