The selling floor and the people on it

The Future of Commerce Using Physical AI

Eleven hypotheses about physical AI in commerce, tested against eighteen graded deployment records, three BLS series, four frontline experiments of which three are randomised, and the physical quantities that set what sensing and manipulation cost. Each resolves to a position on the trajectory, a named binding constraint and the measured change that moves it.

Kat Moore, EdD, MBA, Industrial Research Fellow · The Glass Lab

1. Headline numbers

Physical AI has been installed on the American selling floor at every scale from a single trial store to a chain-wide rollout, and the installed base re-sorted over the decade to 2026. Withdrawals are better documented than installations, because a closure is recorded in an accounting or court filing while an installation is recorded in a company announcement. That asymmetry shapes the evidence base: the strongest evidence in the selling-floor half concerns systems whose scale changed, and the strongest evidence in the workforce half comes from government series and randomised experiments rather than from vendors.

The table below is the short version. It carries the commercial quantities and the physical quantities together, because several of the commercial claims in this report are physical statements written in commercial language and can be checked against published quantities at the grades stated. Each figure is graded at first substantive use, and the grade travels with the number through the tables. Where a number is a company's own claim, the sentence containing it says so.

Table 1. Key figures, with verification grade. Citation numbers refer to the reference list.

2. What is in scope, and how was the evidence graded?

This report is an Industrial Research Fellowship review by Kat Moore, EdD, MBA, hosted by the Glass Lab, the Institute's airspace integration, aviation training and STEM workforce group. The fellow's background is in sales and in education, and the report is organised around that pairing: what embodied AI does to the selling floor, and what it does to the people standing on it. It is a research and review article covering market and usage questions. It is not investment advice, and no security is discussed.

Physical AI here means artificial intelligence embodied in equipment that senses and acts in a store or a fulfilment building: ceiling and shelf cameras that infer what a shopper picked up, mobile robots that photograph shelves, robotic arms in a backroom, electronic labels that change a price on a shelf edge, and the software assistants that frontline staff now carry in a handheld. Online recommendation systems and demand forecasting sit outside the scope except where an experiment on them bears directly on what a human associate does.

Every finding in this report resolves to three things rather than to a verdict. The first is where on the trajectory the application currently sits, stated as what works now, at what scale and under what conditions. The second is which constraint is currently binding, named from five technical classes, material science, actual engineering implementation, energy, computation efficiency and AI algorithms, with regulatory and workforce constraints named as such where those are what bind. Public and reputational exposure is carried inside the regulatory class throughout, because legislation and enforcement are the channel through which that exposure reaches an operator, and no separate reputational class is used. The third is what measured change moves the boundary, stated as a threshold, a factor, a date or a rule. The hypotheses this report tests are stated in section 3 and resolved on those three axes in section 22.

Four grades are used, and they are the Institute's standard. Verified means independently replicated or peer-reviewed, pulled directly from a primary government series, or read from a primary public document of record such as a court filing or congressional correspondence. Reported means a named analyst, operator, agency or news organisation stated it without independent confirmation. Self-published means the number originates with the company selling the thing. Modelled means it was computed rather than measured, including every projection. A modelled figure is never written as a measurement in this report, and where the author has performed arithmetic on other people's figures, the arithmetic is labelled as the author's. One tie-break is needed because the middle two overlap. Where a figure originates with the party running the programme being measured, including a company's own forward-looking target for its own rollout, it is graded self-published regardless of which outlet relayed it. Trade-press repetition of a company statement is not independent confirmation and does not raise a grade. Two further cell values appear in the tables and are not grades. Where a quantity was sought and not found, the grade cell reads not located, which records the absence of a figure rather than ranking one, and it is used only in rows whose value is itself the statement that this review did not locate the quantity. Where a row's content is a reading this author has formed rather than a quantity from any source, the grade cell reads author's judgement, which carries the same standing as modelled without the arithmetic.

One convention governs gaps. Where a quantity was sought and not found, this report says that this review did not locate it, which is a statement about the search rather than a claim about the literature. Two such gaps shape the selling-floor half. This review did not locate a matched-store controlled comparison of any shelf-scanning robot, checkout-free installation or computer-vision loss prevention system published by a retailer or a vendor, so the category's return-on-investment case rests on supplier arithmetic and every vendor effect claim is treated accordingly. This review also found the industry's own shrink benchmark discontinued: the National Retail Federation ended the National Retail Security Survey after 32 editions, the last covering fiscal 2022, so loss-prevention business cases are now written against a baseline for which this review did not locate a current common measurement 36,37.

3. Which hypotheses does this test?

Eleven hypotheses are stated here, seven about the selling floor and four about the workforce. Each is a claim about the market or about usage, not a claim about whether the technology can be built. Each is paired with the reason it matters and with the evidence that would settle it. Sections 4 to 17 test them against the record, sections 18 to 21 set out what a decisive measurement would contain and what this review did not locate, and section 22 resolves every one into a trajectory position, a binding constraint and a threshold.

H1. The variable that sorts the deployment record is how much of the change the customer has to absorb, rather than the technical difficulty of the system. Systems that alter only staff-side work reach chain-wide scale; systems that require the customer to change what they do at the exit reach trial or partial scale and are then made optional or re-sited. This matters because it identifies which applications in the ledger are furthest along their trajectory, and because it tells a workforce planner which roles change first. It is settled by the joint distribution of scale reached against customer-facing change across the ledger in Table 2, and it would be moved by any customer-facing format holding chain-wide under one operator for eight consecutive quarters. That window is set here to exceed the eighteen months inside which four UK operators independently converged on optionality at the exit, and it is the author's choice of observation window rather than a measured interval.

H2. Checkout-free sensing earns where the transaction rate per instrumented square metre is high, because its cost is proportional to floor area and its revenue is proportional to transactions multiplied by basket. This matters because it is the difference between a technology that is finished and a technology that is sited wrongly, and the two sit at different points on the trajectory and are moved by different measurements. It is settled by the sensing load per square foot computed against building electricity intensity, the sensing load being modelled here rather than measured, and by whether the venues that kept the format are the high-throughput ones. It would be moved by a published occlusion-driven camera count, which is what would show how much of the installed density is available for removal, or by a per-stream inference load published for a retail pipeline.

H3. Fixed sensors beat mobile robots on energy per observation above a revisit rate that can be computed, and below that rate mobility is cheaper. This matters because it turns a procurement argument about machines into a measurable quantity, observations per joule, and because the same vendor now sells both. It is settled by the platform power, the observation rate and the shelf-rail camera power, all of which have published or bounded values. It would be moved by a shelf-rail camera drawing about 1 W rather than 4.5 W, which removes the installed-power penalty that is fixed sensing's only disadvantage in this comparison.

H4. Electronic shelf labels are the one system in this ledger reaching chain-wide saturation, and the reason is that the task is repetitive, carries no customer-facing judgement and has an energy budget with roughly two orders of magnitude of headroom. This matters because it is the clearest example in commerce of an embodied system completing its trajectory, and the conditions that let it complete are the conditions to look for elsewhere. It is settled by the rollout count against the company target and by the label's five-year energy budget. It would be moved by completion of the remaining stores on schedule, or by a published labour-hours effect that turns a task-duration claim into a labour claim.

H5. Electronic label hardware carries far more repricing capability than operators have used, so the constraint on dynamic pricing is regulatory rather than technical, with public and reputational exposure carried inside that class. This matters because the public debate about this technology is conducted as though the constraint were the hardware. It is settled by the ratio between the updates the cell supports and the updates operators perform, and by direct measurement of price-change behaviour before and after installation. It would be moved by a replication covering the post-2024 rollout window, with a named chain and at least the 114 stores of the study reported in section 7, showing whether unexplained short-term price spikes rise above the 0.0006 per cent of products measured after installation to mid-2024; and by state legislation on shelf-edge pricing or on camera-based demographic inference at displays.

H6. Centralised robotic fulfilment and store-node fulfilment are separated by achievable local order density rather than by machine capability, so the same technology earns in one geography and is written down in another. This matters because two of the largest US grocers reached opposite conclusions about the same category of machine within about ten months. It is settled by fixed cost per order at achievable local volume, which requires an orders-per-site figure. The record bounds that figure from one side only: the pure-play vendor serving mid-size US grocers reached Chapter 11 at those sites' volumes, which places the break-even above them, and Kroger's expected improvement of about $400 million in e-commerce profitability becomes a per-order figure the moment an order count is attached to it. It would be moved by publication of the orders-per-site-per-week at which a centralised facility matches store-node cost, or by a reduction in fixed capital per automated site.

H7. In manipulation of fresh goods, the binding quantity is reliability and damage rather than cycle time, and the damage threshold is a property of the goods that better perception does not move. This matters because it separates the parts of fulfilment that are close to production from the parts that are not, and because the design variable it identifies, the compliance of the target surface, is cheap. It is settled by measured harvest success, damage rate and cycle time against laboratory picking rates on rigid objects, and by the measured impact thresholds for the fruit. It would be moved by field reliability rising from 66 per cent toward the better than 95 per cent already reached on rigid objects. The two records are not at the same rate and the hypothesis does not claim they are: the fifty-project review's cycle is 33 seconds against under 12 seconds a pick in the laboratory and in one orchard result, so a rate gap sits alongside the reliability gap, and the claim is that reliability and damage are what bind first.

H8. The entry-level commerce role is changing shape rather than disappearing, with transaction handling automated and exception handling, ground truth maintenance and diagnostic selling persisting. This matters because it decides whether the workforce question is retraining the displaced or continuously equipping the churning, and those need different institutions. It is settled by the occupational projections read by task content rather than by headline, and by the openings-to-stock ratio. It would be moved by the 2026-36 projection round, and by BLS publishing the store-against-warehouse split of the stockers and order fillers code.

H9. Assistant effects on frontline workers run in opposite directions by performance rank, so a uniform rollout is not the efficient policy. This matters because the distributional term decides how a rollout has to be designed and measured, and the measurement has to precede the rollout. It is settled by experiments reporting effects by performance rank, of which two now exist at scale, one randomised and one identified from a staggered rollout. It would be moved by a replication on a spoken, mobile, unscripted selling floor, powered to detect removal of the top-quintile reversal rather than an average gain, with the target being a top-quintile effect indistinguishable from zero on the measure that moved at Alibaba, shift-away time up 23.8 per cent; that experiment reached the quintile with 5,940 agents, which sets the order of the sample such a replication needs. It would also be moved by an assistant conditioned on the operator's measured skill.

H10. The dominant displacement channel in commerce and logistics is hiring avoidance rather than separation, which produces no layoff statistics and triggers no publicly funded workforce programme. This matters because every United States workforce instrument of consequence is keyed to a separation event. It is settled by realised headcount against sales growth at the firms whose plans describe the mechanism, and by entry-grade posting volume. It would be moved by a workforce trigger keyed to entry-grade postings rather than to separations.

H11. When equipment turns over faster than a curriculum can be written, accredited and taught, what transfers is a small set of vendor-independent competencies rather than system-specific instruction. This matters because the openings rate in the affected occupations makes any instruction that sits outside the shift undeliverable. It is settled by whether the same competencies recur across systems that differ in every other respect. Table 9 tests that against the ledger records for which this review located an evidence anchor, which are the checkout-free, self-checkout and shelf-scanning entries, together with the frontline experiments, rather than against all eighteen records. It would be moved by one multi-employer agreement on the competency set, and by evidence that a portable record changes hiring at entry grade.

4. Part I. What is actually installed?

The selling-floor half of this report starts with a ledger rather than a thesis. Eighteen programmes are listed below, each with the scale it reached and its status in August 2026. The ledger is not a random sample. It over-represents programmes large enough that their ending was newsworthy, and it under-represents quiet installations that are still running, so it cannot be read as a survival rate for the category.

Sorted by what each system asks of the customer, which is the first test of H1, the ledger separates. Every system in it that changed only in-store staff work, the two shelf-sensing product lines, the shelf-scanning fleet, the electronic shelf label rollout and the committed backroom systems, is operating, expanding or in rollout. Every system that required the customer to change what they do at the exit reached trial or full-store scale and was then made optional, licensed into a higher-throughput venue, or re-sited: checkout-free shopping reached full-store scale under four operators, in the weekly-shop grocery format under its originator alone, and has retreated from it, while the same technology is operating and expanding in venue concessions and breakrooms where the basket is small and the alternative is a long queue. Four records ended without asking anything of the customer, and they ended on other boundaries. Kroger's centralised fulfilment and Takeoff's ended on order density, which is what H6 tests; Walmart's shelf-scanning contract ended on the marginal value of data that pickers in the aisle already produced, which is a workflow condition in Table 6; and Amazon's Blue Jay station was reportedly shut down on a record this report grades weak 39,40.

The arithmetic behind that sort is the status column, and it carries the caveat above. Of the eighteen records, six are operating at or near their maximum recorded scale, one is a commitment rather than an installed base, one entered Chapter 11, and ten were closed, cancelled, converted or reduced in scope. The classification rule is the status cell in Table 2 as written. That count is the author's, over a non-random ledger, and is graded modelled. It supports one planning statement: the iteration cycle for a first-generation installation in this category is measured in years rather than decades, which is a statement about how fast a field in development replaces its own equipment rather than about the durability of the technique.

Table 2. Deployment ledger: physical AI on the US and UK selling floor, 2016-2026. Status as of August 2026.

5. H1 and H2: what does a checkout-free store have to sense?

On 27 January 2026 Amazon said it would close every Amazon Fresh and Amazon Go store in the United States, 72 locations in total, with most shutting that first weekend of February. Amazon's statement frames the decision as a refocus on Whole Foods, where it plans more than 100 new stores, and on delivery; staff were offered redeployment or severance of 90 days at full pay and benefits 1. Store counts of 57 Fresh and 15 Go come from trade and general press on 27 and 28 January 2,3,4, and the grade is reported. Of those 72 stores, the 15 Go stores were the ones checkout-free at close: Just Walk Out had already been removed from all US Fresh stores in April 2024 and replaced with Dash Carts and self-checkout, so the 57 Fresh closures are grocery closures rather than checkout-free closures and are counted that way throughout this report 1,4. This review did not locate per-store economics for the format published by Amazon, nor a disclosed count of employees affected.

What that closure establishes needs care, because Amazon says in the same statement that it continues to expand the technology, and the technical capability is not what changed. Amazon Fresh also carried assortment, siting and price-position problems that have nothing to do with cameras, and the grocery banner was weak overall. The defensible reading is narrower and is a statement about position on a trajectory: sensor-fusion checkout has reached full-store operating scale in the weekly shop under its own originator and has retreated from that format, while continuing to expand in the format where the same technology already earns.

That second format is now the larger installed base. Amazon reports Just Walk Out running in more than 360 third-party locations across five countries, roughly 150 of them added during 2025, concentrated in stadiums and airports, plus more than 40 North American fulfilment-centre breakrooms 1,5,6. Nissan Stadium concessions are one documented example 7. All of these counts originate with Amazon and this review did not locate an independent audit of them, so the grade is self-published. Location is also not equivalent to store, since a concession kiosk counts the same as a shop.

British operators produced the cleanest convergence in the record, and it is the strongest single test of H1. Amazon announced in September 2025 that it would close all 19 UK Amazon Fresh stores, 14 closing and five converting, with around 250 jobs at risk 8,9. Sainsbury's removed Just Walk Out from its Holborn store in February 2025 10. Aldi ended the Greenwich Shop&Go trial, and Tesco moved GetGo to a hybrid format that offers both staffed and self-service tills 11. Four independent operators reached the same design answer inside eighteen months, and the answer was optionality at the exit rather than removal of the technology. UK grocery is unusually price-competitive, and this review did not locate published transaction economics behind any of the four decisions.

H2 states that the format earns where the transaction rate per instrumented square metre is high, and the sensing load makes that testable. Published camera counts for small-format installations cluster near one hundred. A reporter who visited the original 1,800 square foot Amazon Go store in January 2018 estimated about a hundred ceiling units and recorded weight sensors in every shelf 57, and an AiFi-powered Loop Neighborhood Market in Union City carried 109 ceiling-mounted cameras 58. This review did not locate a floor area for the AiFi store, so only the Amazon figure yields a density. At 1,800 square feet, which is 167 square metres, a hundred cameras is one per 1.7 square metres of floor, and the AiFi count corroborates the order of magnitude without contributing to the density.

That density is not set by optics, and the geometry shows why. A 1080p sensor behind a rectilinear 90 degree lens has an effective focal length of 960 pixels, since half of 1,920 pixels divided by the tangent of 45 degrees is 960. The angular pixel pitch is 1.04 milliradians on axis and 0.52 milliradians at the horizontal edge, and the ground sample distance on a plane normal to the optical axis is the range divided by 960, which is 3.13 millimetres per pixel at three metres, 4.17 at four metres, 2.08 at two metres and 0.52 at half a metre. Dividing the full 90 degree field by the pixel count instead returns 0.818 milliradians, which is the frame-average pitch and is the true pitch only for an equidistant fisheye projection; that shortcut understates the ground sample distance by four over pi, a factor of 1.273, at every range, and the figures used here are the rectilinear ones. At three metres a camera with that lens covers six metres by 3.38 metres, which is 20.3 square metres of floor, and at two metres, roughly the height at which a shopper's hands work, it covers nine square metres. One camera per 1.7 square metres is therefore twelve times more cameras than coverage at three metres requires, and 5.4 times more than coverage at hand height requires. What sets the count is occlusion and the number of simultaneous views needed to associate a hand with an item, and this review did not locate either quantity in a vendor publication or in a peer-reviewed source. Mounting height does not change the count: at a fixed recognition standard and a fixed sensor the covered area per camera is fixed, because holding millimetres per pixel constant requires the focal length to scale with range, so 1,920 pixels at 3.13 millimetres is six metres of floor whether the camera hangs at three metres or at six. Camera count rises in proportion to floor area, and ceiling height changes the lens rather than the count.

A published pixel-density standard exists, and the edition matters. IEC 62676-4:2014 anchored four operational requirements, known as DORI, to pixel densities: detection at 25 pixels per metre, observation at 63, recognition at 125 and identification at 250, derived from 40 pixels across a 16 centimetre human face 76,77. Those values were read here from a manufacturer's restatement of the 2014 table rather than from the paywalled standard, and are used at the grade of that restatement 77. The 2025 edition replaced DORI with a seven-grade ladder, overview 20, outline 40, discern 80, perceive 125, characterize 250, validate 500 and scrutinize 1,500 pixels per metre, so the 250 figure survives under the name characterize and the density recommended for identification moved to 500 89,90. Measured against both ladders the ceiling camera sits in a definable place for people: at three metres, 3.13 millimetres per pixel is 320 pixels per metre, which clears characterize at 250 and falls short of validate at 500, and at four metres, 4.17 millimetres is 240 pixels per metre, which falls short of both. The quantity the checkout-free application actually needs is a different one, the millimetres per pixel at which software separates two similar package faces. Both restatements carry the same caveat verbatim, that the specifications hold where a human operator interprets the image and that other definitions apply where the analysis is done by software 77,90, and this review did not locate a machine threshold in any vendor publication or peer-reviewed source. At 3.13 millimetres per pixel a hand of an assumed nominal 100 millimetres is 32 pixels across, and adult hand breadth nearer 90 millimetres lowers that figure. That 32 pixels suffices to follow a hand and does not suffice to separate two similar package faces is the author's judgement rather than a measured threshold, and the corroborating evidence for it is that the installation also weighs every shelf. Camera counts therefore cannot be derived from geometry and have to be taken from installations.

The electrical load follows from the camera count, and it is the quantity H2 turns on. Fixed dome cameras of the class used in retail draw 4.0 W typical against a 10.5 W maximum and 4.8 W typical against an 8.9 W maximum on the two datasheets cited here 59; the working figure of 4.4 W used below is the midpoint of those two typicals and is the author's arithmetic rather than a datasheet value. NVIDIA's DeepStream performance documentation puts a Tesla T4 at 39 concurrent 1080p30 H.264 streams running a primary detector and three secondary classifiers 60, and the T4's board power limit is 70 W 61. Seventy watts over 39 streams is 1.8 W per stream, which is an upper envelope on decode plus inference at 30 frames per second rather than a measured inference power, because the same table reports 31 per cent GPU utilisation at that stream count, so the board is not at its limit and the binding resource is the decode engine. The benchmarked pipeline is an automotive one, a ResNet10 detector at 480 by 272 with three ResNet18 classifiers, inferring at a resolution far below the one the recognition argument turns on, and this review did not locate a published per-stream inference load for a checkout-free retail workload. On that envelope a hundred cameras is about 440 W of sensor and 180 W of inference, which is 0.62 kW carrying 6.2 gigapixels a second, and 0.58 to 0.66 kW across the two datasheet typicals. Held continuously for 8,760 hours a year that is 5.4 MWh, and over a 16-hour trading day it is 3.6 MWh. The 2018 Commercial Buildings Energy Consumption Survey puts food sales buildings at 53.3 kWh per square foot per year 62, so an 1,800 square foot store uses about 95,900 kWh, and the sensing is 5.7 per cent of it run continuously, 5.3 to 6.0 per cent across the camera spread, or 3.8 per cent on a 16-hour day. Scaled at constant density, a 25,000 square foot format needs about 1,390 cameras, 8.6 kW and 75 MWh a year at 8,760 hours, carrying 86 gigapixels a second.

Every one of those quantities is proportional to floor area and independent of how much is sold, while the revenue the cost is set against is transactions multiplied by basket. That is the physical content of H2, and it accounts for the format's position: the sites that kept it are stadium, airport and breakroom venues, whose transaction rate per instrumented square metre during the hours the instrument is powered is the highest on the author's reading of the venue types, since this review did not locate a transaction rate published for any of them. The binding constraint is engineering implementation, because the installed camera density is 5.4 to 12 times what the optics require and the excess is an occlusion allowance rather than a resolution requirement. The threshold that moves the boundary can be stated, and so can what stands between an installation and it. Coverage at hand height alone would be satisfied by about 19 cameras and 0.12 kW at 1,800 square feet, and by about 258 cameras and 1.6 kW at 25,000 square feet. That is an optics-only floor rather than a reachable operating point. The excess over it is the occlusion allowance and the simultaneous-view requirement, and since this review did not locate either quantity, how much of the difference between 100 cameras and 19 is available for removal cannot be established from outside. The gap is the size of the prize rather than the size of a saving. A published occlusion-driven camera count, and a per-stream inference load measured on a retail pipeline rather than an automotive one, are the two measurements that would tell an operator how much of it is real.

Table 3. Physical test conditions on the checkout-free format. Optical figures use a rectilinear lens model with a focal length of 960 pixels for 1080p at 90 degrees horizontal. Rows graded modelled are the author's arithmetic on the sources named and are computed here rather than observed. Where a quantity was sought and not found, the row says that this review did not locate it, which is a statement about the search and not about the literature.

6. H3: where do shelf robots lose to fixed sensors?

Walmart ended its Bossa Nova shelf-scanning contract in November 2020 after piloting in about 500 stores and announcing an expansion to 1,000; Bossa Nova then cut roughly half its staff and wound down retail robotics 12,13. This review did not locate an official reason from Walmart. Reporting at the time attributed the decision to three separate things: online order picking already had associates in every aisle producing the same shelf data as a by-product, simpler human methods worked and cost less, and Walmart US chief executive John Furner was concerned about how shoppers reacted to the machines. Where the data a robot collects is already generated as a by-product of picking work, its marginal contribution is small, and that is a statement about the workflow the machine was placed into rather than about the machine.

The category continued and changed shape. Simbe, on its tenth anniversary, claims deployment with nearly five dozen retail chains across ten countries, 4.7 million autonomous operating hours, 1.8 million kilometres travelled, 44.8 billion shelf photographs and 18 billion price tags scanned 14. Every one of those originates with the vendor and is graded self-published, this review did not locate a third-party audit or a denominator for any of them, and nearly five dozen retail leaders includes trials. Badger Technologies, a Jabil division, reports more than 1,000 robots deployed and markets them as a Digital Teammate 18. Both grades are self-published and the numbers should be read as scale indicators only.

What changed in this category was the sale rather than the hardware. Simbe sells a data subscription in which the robot is the sensor platform, and Badger's naming positions the machine beside the associate rather than in place of them. Simbe's published client outcomes are 98 per cent on-shelf availability, 90 per cent improved price and promotion accuracy, more than 50 hours a week returned to store teams, 42 per cent of detected shelf gaps fixed instantly, and in an earlier statement a 60 per cent drop in out-of-stocks with online fulfilment time halved 14,15,16. This review did not locate a control group, a baseline, a named retailer or a disclosed method for any of them.

One of those claims carries the workforce argument later in this report. More than 50 hours per week returned to store teams is a labour-hours claim that does not state what happened to the hours. Whether they came off the payroll or were redeployed into service, replenishment or fulfilment is the single most important fact for a worker, and the disposition of hours freed by these systems is not stated in any vendor claim located for this review.

H3 states that fixed and mobile sensing are separated by energy per observation, and Simbe's January 2025 launch of Tally Spot, fixed ceiling and shelf cameras watching high-turn and high-shrink zones continuously 17, puts both modes in one vendor's catalogue. The usual physical justification for preferring fixed sensors, that the robot has to move its mass down every aisle, does not survive arithmetic. The Tally 3.0 platform is published at 40 kg on a 2,000 mm mast carrying twelve vision cameras plus a depth camera and an HDR camera 63, and that mass is the current platform's applied to a decade of mixed generations, which the conclusion survives because the bound below already carries a fourfold coefficient margin and mass enters linearly. Rolling-resistance work is the coefficient multiplied by weight and distance, and the coefficient is published rather than assumed. A caster engineering reference gives 0.01 to 0.08 across industrial casters, with 95A polyurethane at 0.025 to 0.035 and 85A at 0.035 to 0.05 86. A manufacturer's white paper states the same quantity in its length form, where force equals coefficient times load divided by wheel radius, giving 0.047 inches for 85A polyurethane at 1,200 pounds per wheel on a steel floor at 3 mph, which for wheels of 75 to 200 millimetres diameter is a dimensionless 0.012 to 0.032 87. The two sources disagree for the same tread material, and this review did not locate a wheel diameter or tread material published for the Tally platform, so 0.02 is taken from the converted band and sits below every polyurethane entry in the first source. The coefficient multiplied by 40 kg, 9.807 m/s2 and 1.8 billion metres is 1.41 times ten to the tenth joules at 0.02, about 3,900 kWh for the whole fleet's ten years of travel. One average United States food sales building at 6,200 square feet and the survey intensity uses 330,000 kWh a year, which is 905 kWh a day 62, so the fleet's entire decade of traction work is 4.3 days of one such building. At 0.08, four times the working value and the top of the published range, the decade is 15,700 kWh, or 17.3 days of the same building. Traction accounts for a negligible share of the energy across the whole published range, which is why the disagreement between the two sources does not change the conclusion.

The cost of mobility is serialisation in time, and the vendor's own aggregate supplies the numbers. Dividing distance by hours gives a mean speed of 0.106 m/s, which is 383 metres an operating hour, and dividing 18 billion price tags by 4.7 million hours gives 3,830 tag reads an hour, which is 10,000 tags for every kilometre travelled 14. These are means over what the vendor calls autonomous hours, and this review did not locate a definition of that term in the release or on the vendor's own site, so whether docking and idling sit inside or outside the denominator cannot be established from outside. The platform's standing load is sensing and compute. The Jetson AGX Orin 64GB module is specified with 15, 30 and 50 W modes and up to 60 W total module power, and the Industrial variant with 15, 35 and 60 W modes and up to 75 W 64. This review did not locate a published power figure for the fourteen cameras the platform carries, so the allowance is bounded rather than sourced: fourteen enterprise dome cameras at 4.4 W is 62 W, an upper anchor, because that class carries a network stack, an encoder and a weatherproof housing that a board-level module does not. The working band used here is 50 to 110 W with 90 W as the working figure, and its endpoints should be read explicitly. The 110 W ceiling is a 50 W module mode plus about 60 W of cameras. The 50 W floor is a module mode with the camera allowance neglected, so the true floor is higher, which raises the 47 J figure below and moves the low end of the crossover envelope up. The 90 W is the author's estimate rather than a manufacturer's specification, and every figure downstream of it scales linearly with it. At 90 W and 3,830 reads an hour, one price-tag observation costs 85 J, falling to 47 J at 50 W and rising to 103 J at 110 W. The figure also inherits the ambiguity in the hours: if the denominator includes time the robot is not reading tags, 85 J is an upper bound rather than a point estimate.

A ceiling camera cannot substitute directly, because at three to four metres it resolves 3.1 to 4.2 millimetres per pixel. That this does not read shelf-edge price text is the author's judgement rather than a measured threshold, since the pixel-density grades are defined for a human operator interpreting faces and not for reading print 77,90, and the corroborating evidence is that the vendor's own fixed product sits on the shelf rather than on the ceiling 17. A shelf-rail camera at half a metre resolves 0.52 millimetres per pixel and covers exactly one metre of frontage by 0.56 metres of face height. Its power depends on the rate at which it infers, and the two are not independent. The 1.8 W envelope is derived from 39 concurrent 1080p30 streams and is therefore a 30 frames per second figure, so charging it in full to a camera that produces one observation a second overstates that camera by a factor of thirty on the inference term: 4.4 W of camera plus 1.8 W is 6.2 W at 30 fps, while 4.4 W plus 0.06 W is 4.5 W at 1 fps. This review did not locate a shelf-rail camera datasheet, so an enterprise dome camera's typical draw stands in for a different product class at either rate. How many tags one such camera sees is the quantity the comparison turns on. Ten tags per metre is tags per metre of robot path integrated over the full 2,000 mm mast, so a single camera covering 0.56 metres of a 2.0 metre imaged face sees about 2.8 of them, and covering the whole face takes 3.6 cameras per metre of aisle.

At 2.8 tags in view a fixed observation costs 2.2 J charged at the full 30 fps envelope and 1.6 J charged at its own 1 fps, against the robot's 85 J, so fixed sensing is 38 to 53 times cheaper per observation and the two are equal at a revisit interval of 38 to 53 seconds on a 90 W platform estimate. Three assumptions widen that interval. Only if one camera saw all ten tags on a metre would the observation cost 0.62 J and the crossover stretch to 137 seconds, or 190 seconds at 1 fps; this review did not locate a published linear density of tags along a bay face, and 10,000 tags per kilometre of path conflates bay height, single-sided against double-sided passes and non-scanning transit, so the ten-tag case is the unsupported end. The 50 to 110 W platform band moves the crossover to 21 to 47 seconds at 30 fps and 29 to 65 seconds at 1 fps. Taking the geometry-supported tag count with both remaining bands gives a crossover of roughly 21 to 65 seconds, with 38 seconds the central figure. The frame-rate correction improves both of the fixed camera's numbers at once, lowering its energy per observation and lowering its installed-power penalty, so it favours fixed sensing on both metrics. What it does not do is change the sign of that penalty. Fixed sensing costs 4.5 to 6.2 W per metre of shelf if one camera suffices and 16 to 22 W per metre for full-face coverage, so instrumenting the 383 metres the robot passes in an operating hour takes between 1.7 and 8.4 kW, against the robot's 90 W spread over the same 383 metres, which is 0.24 W per metre.

H3 therefore resolves with a number attached. Both modes are in production, the same vendor sells both, and the discriminator is revisit rate: above roughly one observation every 21 to 65 seconds fixed sensing is cheaper per observation, and below it mobility is. The binding constraint is energy, specifically installed power per metre of shelf, which is the only term running against fixed sensing. The measured change that moves the boundary is a shelf-rail camera drawing about 1 W rather than 4.5 W, which takes full-face coverage of 383 metres from 6.1 to 8.4 kW down to about 1.4 kW and leaves revisit rate as the sole discriminator. That accounts for the vendor placing fixed cameras on high-turn and high-shrink zones, where the required revisit rate is highest, while keeping the robot for everything else.

Table 4. Energy per observation: mobile against fixed shelf sensing. Rows graded modelled are the author's arithmetic on the sources named. Where a quantity was sought and not found, the row says that this review did not locate it.

7. H4 and H5: what do electronic shelf labels actually change?

Walmart expects electronic shelf labels in all 4,597 US stores by the end of 2026, up from 2,300 stores in mid-2026, against an original 2024 plan that targeted only the 2,300; the supplier is VusionGroup 28,29. Walmart says a price change that used to take two days now takes minutes, which is the company's own figure and describes task duration rather than labour hours or headcount. The 4,597 number is a target for end-2026 rather than a completed rollout, and this review did not locate a shrink or labour-hours effect published by Walmart.

This is the system furthest along its trajectory in the whole ledger, and the reason is visible in what it touches. Changing a paper price tag is repetitive, has no customer-facing judgement in it, and its removal is not something a shopper experiences. Compare that with checkout-free exit or a robot in an aisle, where the labour case is contested and the customer is party to the change. The saturation pattern in Table 2 sorts along that line rather than along technical difficulty, which is the second confirmation of H1.

Display physics is part of why saturation is reachable, and it settles whether energy is the binding constraint on H4. A CR2450 coin cell holds 620 mAh at 3.0 V nominal, which is 6,700 J 65. The VUSION 1.6 BWR label, product code R16S01, is specified with one CR2450 and a five-year battery life at one update a day, over a 27.5 by 27.5 mm active area 66, and other models in the same range carry two, three or eight cells, so the single-cell figure is model-specific. A 2.13 inch electrophoretic panel of the class used in labels draws 9 mW for a three-second full update and 0.003 mW in standby, so one full update costs 27 mJ across an 11.5 square centimetre active area, which is 2.35 mJ per square centimetre 67. The per-update energy and the service life therefore come from different products, and at that energy density the 1.6 inch label's 7.56 square centimetres would cost 17.7 mJ per update. The cell would supply about 248,000 full updates of the 2.13 inch panel if it did nothing else, which is 136 a day for five years, or 377,000 updates and 207 a day at the smaller label's area.

The vendor's own duty cycle gives the cleaner budget. One update a day for five years is 1,825 updates, which at 27 mJ is 49 J out of 6,700 J, or 0.7 per cent of the cell. Two terms belong in that budget. The pigment position is mechanically bistable, so holding the image takes no drive energy and does not relax when the drive is removed, but the panel's own standby draw of 0.003 mW carried over the same five years is 473 J, which is 7.1 per cent of the cell and about ten times the energy of all the updates together. Standby is treated here as unscaled because it is most plausibly driver leakage rather than active-area current, and the alternative is carried so the spread is visible: scaled to the 1.6 inch label's area it is 311 J, giving a display subsystem of 343 J or 5.1 per cent against the unscaled 522 J or 7.8 per cent. Either way the radio duty cycle and the cell's own shelf life carry roughly 92 to 95 per cent of the budget. This review did not locate a measured radio energy per update in a deployed network, nor a published count of labels installed per store.

Hold energy separates the two surviving low-power display classes by about a factor of three, which is not enough to settle which one took the shelf edge. Transmissive twisted-nematic and IPS panels have to be driven continuously, because the molecular alignment relaxes without a field, and they are not in this comparison. The lowest-power liquid crystal class in general use, memory-in-pixel, holds through in-pixel storage at on the order of 10 microwatts static 68, and over five years that is about 1,580 J, roughly a quarter of the whole cell spent on an image that never changes. Computed the same way over the same five years, the electrophoretic panel's own standby is 473 J, or 7 per cent. The comparison is 24 per cent against 7 per cent rather than a comparison against zero. Bistable liquid crystal modes are the counterexample to any universal claim about the technology class. The zenithal bistable device aligns a nematic against a grating surface whose anchoring depends on the applied voltage, giving two stable molecular states 78, and cholesteric liquid crystals were shown to hold a reflective state with the field removed in 1994 80. The zenithal device was carried from concept to a consumer product 79, reached mass production around 2005 with electronic shelf labels as its initial application market, and the technology was sold to New Vision Display in 2019, a span of about fourteen years that is arithmetic on the two published dates 81. A liquid crystal shelf label that holds its image at zero field therefore existed alongside electrophoretic and did not take the market. The reported reason is appearance rather than energy, the trade account stating that buyers moved to electrophoretic displays for the paper-like look 81. That is a single trade-press explanation with no measurement behind it, and this review did not locate a measured account of why electrophoretic displaced bistable liquid crystal at the shelf edge.

H4 therefore resolves as follows. The application is the most advanced in the ledger, at 2,300 of 4,597 US stores under one operator with a company target of completion inside the year, and the energy budget was never the binding constraint, since the whole display subsystem spends 5.1 to 7.8 per cent of the cell over five years. What binds is engineering implementation, meaning label unit cost, radio infrastructure and the installation labour of fitting several thousand labels in each of 4,597 stores. The measured change that completes the trajectory is the remaining 2,297 stores on the published schedule, and the measurement that would upgrade the business case from task duration to labour is a published labour-hours effect with the disposition of the hours stated.

H5 concerns the gap between what the hardware supports and what operators do with it, and the ratio is large. The cell carries 136 to 207 full updates a day against a duty cycle of one, which is roughly two orders of magnitude of headroom, and at chain scale the energy of repricing is negligible: assuming ten thousand labels a store, changing every price in all 4,597 Walmart stores 28 costs about 0.34 kWh of display energy at the 2.13 inch panel's per-update figure, or 0.23 kWh at the 1.6 inch label's, which is 33 or 22 seconds of one average food sales building's electricity 62. This review did not locate a published count of labels installed per store, so that figure moves with the assumption. Nothing physical restrains repricing frequency.

The public debate has been conducted as though the hardware were the constraint. In August 2024 Senators Elizabeth Warren and Bob Casey wrote to Kroger warning that the labels would enable surge pricing and price-setting by consumer profile; Kroger replied that it does not and has never engaged in surge pricing 31. That exchange framed the technology for a general audience and remains the reference point in most coverage.

The first large-scale measurement of what operators actually did supports the retailer. Robert Sanders of UC San Diego Rady, Ioannis Stamatopoulos of UT Austin McCombs and Robert Bray of Northwestern Kellogg studied 114 US grocery stores across four states with more than 180 million product-level observations from October 2022 to mid-2024 and found no evidence that electronic labels enabled surge pricing 30. Unexplained short-term price spikes affected 0.0050 per cent of products before installation and 0.0006 per cent after, a decrease. Price changes stayed chain-wide rather than store-local and tracked wholesale costs, competitor prices and manufacturer promotions. The grade is reported, since it is an SSRN working paper that is not yet peer-reviewed, covering one unnamed chain in a window that closes before most of the current rollout, and it measures what retailers did during a specific inflationary period rather than what the hardware makes possible. The senators' separate concern about camera-based demographic inference at shelf displays is untouched by this data. H5 therefore resolves with the binding constraint named as regulatory rather than technical, public and reputational exposure being carried inside that class because legislation is how it reaches an operator, and with two measurements that would move it. The first is a replication covering the post-2024 rollout window, with a named chain and at least the 114 stores of the study above, reporting whether unexplained short-term price spikes rise above the 0.0006 per cent of products measured after installation. The second is state legislation on shelf-edge pricing or on biometric inference at displays.

8. H6 and H7: what order density does automated fulfilment need?

Kroger closed three Ocado-powered automated customer fulfilment centres at Pleasant Prairie, Frederick and Groveland, cancelled a planned Charlotte facility, and took $2.585 billion of impairment and related charges in fiscal Q3 2025, a figure that includes an accrued cash termination payment of about $350 million to Ocado 19,20,21,22,23. It expects the change to improve e-commerce profitability by about $400 million in 2026 by shifting to in-store fulfilment. Interim chief executive Ron Sargent ordered a site-by-site review in September 2025 before deciding, and Kroger had already paused Ocado development in September 2023 and closed three spoke facilities in March 2025 that did not meet its benchmarks.

The impairment is an accounting entry rather than cash lost, and it bundles other e-commerce changes, so the headline figure should not be read as money that left the building. Kroger's own explanation is order density and suburban delivery distance, which is a property of American geography rather than of Ocado's engineering, and the same technology continues to run in denser markets. What the episode establishes is a boundary condition: at this order density and this geography, the store is the cheaper fulfilment node, and centralised capacity is written down against it.

Kroger's experience was not idiosyncratic. Takeoff Technologies, a grocery micro-fulfilment specialist described in the trade press as a leader in the category, filed Chapter 11 in Delaware on 30 May 2024 after operating losses in 2022 and 2023, owing roughly $12.9 million in trade and unsecured claims, and was sold through a customer-funded 363 process to Woolworths 24,25. The filing states liabilities rather than causes, and the diagnosis that volume per site never reached the level the capital required is trade-press analysis. The pure-play vendor serving mid-size grocers reached Chapter 11 before the largest deployment was written down, which places the break-even volume above the per-site volumes those grocers supplied. That is a one-sided bound and the only one this review located.

Walmart moved the opposite way on the same category of machine. Company guidance puts about 65 per cent of stores serviced by automation and about 55 per cent of fulfilment-centre volume through automated facilities by the end of fiscal 2026, with an expected 20 per cent improvement in average unit cost 26. Symbotic completed its acquisition of Walmart's Advanced Systems and Robotics business on 28 January 2025 for $200 million in cash and up to $350 million in contingent consideration, and signed a commercial agreement under which Walmart pays $520 million for the development programme and, if performance criteria are achieved, is committed to purchasing and deploying systems for 400 accelerated pickup and delivery centres over a multi-year period 27. All of those figures are company guidance rather than audited outcomes, serviced by automation is Walmart's own definition, and secondary reports claiming the 400 units were already deployed by March 2026 conflict with Symbotic's multi-year language, and they are not cited here because this review did not locate a primary statement behind them. The two largest US grocers reached opposite conclusions about the same technology within about ten months, Walmart committing in January 2025 and Kroger withdrawing in its quarter ending November 2025, and the variable that separates them is fixed cost per order at achievable local volume, where Walmart's store density supplies a denominator Kroger's suburban geography did not. H6 therefore resolves with engineering implementation as the binding constraint, specifically fixed capital per site against achievable local order density, and with a single measurement that would settle it for any operator: the orders-per-site-per-week at which a centralised facility matches store-node cost, which this review did not locate published by any party. The record bounds it from one side only, above the per-site volumes Takeoff's mid-size grocer sites reached, and Kroger's expected improvement of about $400 million in 2026 e-commerce profitability becomes a per-order figure the moment an order count is attached to it.

H7 concerns what the machines inside those buildings can pick, and here the binding quantity is a property of the goods. Drop tests on apples put the bruise-free ceiling at an impact velocity of 0.25 m/s, the highest velocity at which none of the tested varieties bruised, with mean maximum contact stress at bruising between 0.4 and 0.5 MPa, measured against a rigid flat plate 69. A free fall reaching 0.25 m/s begins 3.2 millimetres above that plate, since 0.25 squared divided by twice 9.807 is 3.19 millimetres. That 3.2 millimetres is not a release height a manipulator can servo to reliably is the author's judgement rather than a measured limit, and the consequence is that a machine placing produce onto a hard surface has to servo to contact and confirm on every placement.

The compliant case is measured separately and moves the threshold by a factor that depends on which quality criterion is applied. Pendulum impact tests on Jazz apples report three series, each against its own bare-plate baseline 83; the conference companion reports the first two of them 88. The non-bruising impact level is 22 N against a bare aluminium plate, rising to 95 N against 12.7 mm polyurethane foam of 2.1 kPa firmness and to 160 N against foams of 4.8 and 9.7 to 11 kPa, a factor of 4.3 to 7.3. Extra Fancy quality is maintained in 100 per cent of fruit at 50 N bare against 115, 171 and 180 N cushioned, a factor of 2.3 to 3.6. The fresh-market quality threshold is 95 N bare against 189, 230 and 232 N cushioned, a factor of 2.0 to 2.4. The commercial criterion a packing line is held to is the fresh-market one, so the statement of the gain from a compliant target is a factor of 2.0 at the criterion that governs commerce, rising to 7.3 at the criterion of no bruising at all. The fruit-to-fruit case, which is the bed of produce a bin fills with, is measured by the same group on the same pendulum device 84, and a different group independently corroborates the rigid-surface case by free-dropping Golden Delicious apples onto concrete and wood plates 85. The force ratio is not a height ratio, because a compliant target both raises the tolerable peak force and lowers the peak force a given drop produces, so the gain in allowable release height is larger than the force factor. This review did not locate a published release height in millimetres for a named commercial bin liner, so the gain is stated as the measured force ratio rather than converted to a height.

Area-averaged grip pressure is not the binding limit, although the comparison is not like for like. Vacuum cup catalogue forces are quoted at minus 0.6 bar on a dry or lubricated surface with no safety factor, a 90 mm cup being rated 328 N against the 382 N that pressure multiplied by area predicts 70. Minus 0.6 bar applies 0.06 MPa across the cup face, and a catalogue 20 mm cup is rated 15 N at that vacuum 82, which is 7.6 times the weight of an apple taken at an assumed nominal 200 grams. The two pressures are different quantities: the 0.4 to 0.5 MPa figure is a peak dynamic contact stress over a small growing contact patch during an impact against a rigid plate, and the 0.06 MPa figure is a static pressure differential averaged over the whole cup face, so the ratio of about one to seven bounds the average and says nothing about the peak. This review did not locate a published measurement of the local stress under the sealing lip, which carries the reaction load over a narrow annulus. The seal is where grasping on produce is lost. The same technical data notes that quoted values are not reached where the vacuum cannot be established, for example on porous workpieces, which describes mesh bags, netted citrus, leaf surfaces, cold-chain condensate and anything wet.

The deployment record follows those quantities, and it sorts by reliability rather than by speed. A peer-reviewed review of fifty harvesting robot projects found mean localisation success of 85 per cent, detachment success 75 per cent, harvest success 66 per cent, fruit damage 5 per cent and a cycle time of 33 seconds, with no improvement across three decades and none of the fifty commercialised 71. A more recent orchard result, about 85 per cent harvesting success at a 12 second mean cycle on apples and peaches, comes from daytime field trials of the OrBot platform during the 2021 season and is stated as prior work in the abstract of the same group's 2024 nighttime paper rather than measured in it 72; this review did not locate a primary publication of the 2021 trials. The two success figures are not established as the same quantity, because the review's 66 per cent counts localisation failures in the denominator while the OrBot rate may be conditional on fruit already detected, so they are not compared here. Over the same period, laboratory bin picking of up to 25 novel rigid objects reached more than 300 mean picks an hour at better than 95 per cent reliability 73. More than 300 picks an hour is under 12 seconds a pick, which is the OrBot cycle exactly, so there is no rate gap between those two results; the rate gap in these sources is against the review's 33 second cycle, which is 109 an hour. What separates rigid laboratory picking from deformable field picking here is reliability and damage rather than speed, better than 95 per cent in the laboratory against 66 per cent harvest success and 5 per cent fruit damage in the review.

The one large production measurement is on packaged goods and shows how far that class has come. Amazon's robotic stow system logged 224 units an hour against 243 for human stowers in March 2025, which is 92 per cent of the human rate, and over the most recent 100,000 stow attempts recorded 85.86 per cent success, 9.31 per cent unproductive, 3.77 per cent amnesty, 0.24 per cent damage and 0.82 per cent other defects, a breakdown that sums to 100 per cent 74. The largest non-success category is unproductive attempts rather than dropped items. Amnesty is the paper's own term, defined as either the grasped item or an item already in the pod falling to the floor, so it is a per-attempt rate of anything reaching the floor and includes collateral items already stowed. The 10.87 second mean cycle in the same source covers direct inserts and stack behaviors only, at 90.72 per cent success, while 224 units an hour is the end-to-end rate including every other behavior and its overheads, which is why 3,600 divided by 10.87 does not return it. Ocado publishes a theoretical maximum of 630 units an hour for its on-grid robotic pick 75. Carrying the 3.77 per cent amnesty rate onto twenty attempts gives a 54 per cent chance that at least one item reaches the floor, since one minus 0.9623 raised to the twentieth power is 0.536; that calculation is the author's, it is a per-attempt exposure rather than a per-item drop probability, and it applies a self-published rate measured on a different item population to goods whose bruise threshold against a hard surface sits 3.2 millimetres above it.

H7 therefore resolves with two constraints named and two thresholds attached. The binding constraints are material science, because the bruise threshold is a property of the fruit and does not move with better perception, and AI algorithms, because grasp selection under occlusion and variability is what separates 66 per cent field reliability from better than 95 per cent laboratory reliability. The measured changes that move the boundary are, first, target compliance, which is a design variable and raises tolerable peak force by a factor of 2.0 to 2.4 at the commercial fresh-market criterion and 4.3 to 7.3 at the no-bruise criterion, and second, field reliability rising from 66 per cent toward the better than 95 per cent already demonstrated on rigid objects. The cycle times have to be carried separately, because they are not the same. The laboratory result at under 12 seconds a pick and the single orchard result at a 12 second mean cycle match each other, while the fifty-project review's cycle is 33 seconds, so a rate gap of at least 2.75 to one against the review sits alongside the reliability gap. Reliability and damage are what bind first, and speed is the smaller of the two distances.

Table 5. Manipulation: measured damage thresholds and production rates. Rows graded modelled are the author's arithmetic on the sources named. Where a quantity was sought and not found, the row says that this review did not locate it.

9. H1 again: where did the cost of self-checkout move to?

Self-checkout is the largest installed base in this ledger by a wide margin, and it is the system now being re-sited. Dollar General pulled self-checkout from more than 12,000 of its 20,000-plus stores during 2024 and removed it entirely from roughly 300 of its highest-theft locations; Target limited self-checkout to ten items or fewer in March 2024, and Five Below announced in 2024 that it was removing it from its highest-risk stores 33,34. Whether all three policies still stand in 2026 is not established here. A LendingTree survey of more than 2,000 US consumers in December 2025 found 27 per cent of self-checkout users admitting they deliberately did not scan at least one item 35. A 2023 survey by the same publisher reported 15 per cent, but the two are separate online panels and the wording and screening are not established as identical, so they are not treated here as a trend. That figure is self-reported admission in an online panel and is not a measurement of loss; self-report on theft is unreliable in both directions.

The retailers' own causal attribution is theft, and this review did not locate the shrink arithmetic behind the decisions published by any of them. What can be said with more confidence is that the system relocated part of the cost, from payroll into shrink and into the customer's own time at the terminal, and this review did not locate the measurement that would say how much of the payroll saving survived the transfer. A labour-hours saving becomes a profit-and-loss saving only if the two transferred costs stay smaller than the payroll they replace, and several large operators re-sited the system on that arithmetic while keeping it everywhere else. This is the same H1 boundary seen from the operator's side: the format asks the customer to perform the scan, and the surviving design is the one that leaves the customer a choice.

Verifying any of that is harder than it was. The National Retail Federation discontinued the National Retail Security Survey after 32 editions, the last covering fiscal 2022 36. That final report attributed 36 per cent of inventory loss to external theft, 29 per cent to employee theft, 27 per cent to process and control failures and errors, 6 per cent to unknown causes and 1 per cent to other 36. NRF had previously retracted a headline organised-retail-crime statistic after a Retail Dive analysis found an error 37. The widely circulated shrink figure of 1.6 per cent and around $112 billion trace to fiscal 2022 data and should not be presented as current.

Two things follow for a loss-prevention business case. The baseline has no current common measurement that this review located, and process and control failures accounted for 27 per cent of shrink in the last measurement, more than a quarter of the loss, which a camera at the exit is not positioned to see. Everseen, a computer-vision loss-prevention vendor, reports analysing 140,000 self-checkout lanes daily and commissioned a Forrester Total Economic Impact study in September 2024, and this review did not locate the underlying per-retailer shrink deltas in public 38. This review did not locate a controlled measurement of shrink reduction from computer vision at checkout published by a retailer or a vendor. The binding constraint on this application is AI algorithms, specifically at-terminal item verification accurate enough to price against a shrink baseline, and the measurement that would move it is any matched-store shrink delta published with its baseline attached.

10. What does the selling-floor record support?

Eight conditions can be drawn from Tables 2 to 5 with varying confidence, and they are set out in Table 6 with the evidence attached to each. They are siting and engineering conditions rather than laws, and each rests on a small number of operators in two national markets.

One further observation concerns measurement rather than siting. Every effect claim in the selling-floor literature located for this review originates with a vendor, and the industry benchmark that would let a retailer check one has been discontinued. No effect claim located in this category states the store-weeks, the matched comparison stores, the baseline metric, or the disposition of any labour hours freed, which is why none of them can be checked. Section 18 sets out what a decisive measurement would contain.

Table 6. Siting and engineering conditions supported by the 2016-2026 deployment record and by the physical quantities in Tables 3 to 5. The grade column carries the provenance of the figures in each row, on the scheme in section 2.

11. Part II. What do the labour series show, before any argument?

US retail trade employment was 15,459,700 in June 2026, seasonally adjusted, which is below the February 2020 level of 15,519,200 and below January 2019's 15,617,400. Over the same period total nonfarm employment rose from 152,293,000 to 158,984,000. These come directly from the Bureau of Labor Statistics public API, series CES4200000001 and CES0000000001, and are graded verified 42. Retail's share of national employment therefore fell from 10.19 per cent to 9.72 per cent, which is the author's arithmetic on two verified levels rather than a BLS-published series.

The headline is flat rather than falling. Retail is shrinking as a share of American work while the rest of the economy grows, without shedding a visible mass of jobs. Any argument that automation has already emptied the selling floor is not supported by the aggregate series, and this report says so before it says anything else.

Two measurement caveats matter more than usual here. The CES series counts jobs at retail-trade establishments under NAICS 44-45 rather than retail occupations, so a cashier at a hospital gift shop is classified elsewhere. E-commerce fulfilment work has migrated into transportation and warehousing under NAICS 49, which makes retail look worse than the work actually is. Anyone using the retail series as a proxy for selling-floor labour is measuring the wrong denominator by a growing margin.

Pay moved in the direction that complicates the substitution story. Average weekly earnings of all retail trade employees rose from $593.12 in January 2019 to $787.87 in June 2026, a 32.8 per cent nominal increase, again pulled directly from the BLS API, series CES4200000011, and graded verified 43. CPI-U (BLS series CUUR0000SA0) rose 32.7 per cent over the same span, from 251.712 in January 2019 to 333.952 in June 2026, pulled from the same API 43, so the real change is about 0.1 per cent rather than thirty-three, and that deflated figure is the author's arithmetic. Weekly earnings also move with hours, so a shift toward longer scheduled shifts would inflate the nominal figure without any hourly raise. Employers paid a third more in nominal terms for the labour they were said to be replacing, which is the payroll pressure an operator sees, while in real terms the price of that labour held level, so the series records no revaluation of the remaining roles.

12. H8: which retail occupations do the projections move?

The aggregate hides the shape of the change, and the occupational projections show it precisely. The Bureau of Labor Statistics models cashiers falling from 3,157,200 in 2024 to 2,843,600 in 2034, a decline of 313,600 or 9.9 per cent, which is the largest projected numeric decline of any US occupation in the 2024-34 projections (BLS Employment Projections table 1.5), and the agency explicitly attributes it to self-service checkout and online sales 32,44. The 2024 base-year levels are survey estimates and are solid. The 2034 endpoints are model output built on assumed technology-adoption and consumer-spending paths, and this review did not locate the agency's prior-round cashier projection set beside the realised 2024 level, which is what would score the model.

Decline is concentrated in transaction handling rather than across the store, which is the core of H8. Supervision falls only 5 per cent. Stockers and order fillers, which is physical and spatially variable, is projected to grow by 235,000 or 8.5 per cent. That occupation is SOC 53-7065, a transportation and material-moving code whose employment sits in warehousing as well as in stores, and this review did not locate a published store-against-warehouse split, so it cannot be read as selling-floor growth without that qualification. That boundary is the same one the deployment ledger drew: embodied AI has reached production on the repetitive, fixed-position, judgement-free work, and the variable physical work and the persuasive work sit earlier on the same trajectory.

Retail salespersons are projected essentially flat at -0.5 per cent, which the Occupational Outlook Handbook describes as little or no change, with about 586,000 annual openings for retail sales workers, while retail trade as a sector is projected to lose more jobs than any other at -1.2 per cent 32,44. Retail salespersons is a broad occupational code spanning car sales and apparel floors with very different exposure, and flat national employment can conceal large compositional and geographic churn. Taken at face value, the projections say the transactional half of the floor job is being automated and the diagnostic, persuasive half is not, which is what H8 claims. The binding constraint is workforce rather than technical, and the measurements that would sharpen it are the 2026-36 projection round and a published store-against-warehouse split of SOC 53-7065.

Table 7. BLS occupational projections, 2024-34. Base-year levels are survey estimates; 2034 endpoints are modelled. Source: BLS Employment Projections 44.

13. How did thirteen years of automation forecasts score?

Frey and Osborne assigned cashiers a 0.97 probability of computerisation, ranked 657th of 702 occupations, with counter and rental clerks also at 0.97, retail salespersons at 0.92, stock clerks and order fillers at 0.64, and first-line supervisors of retail sales workers at 0.28 45. The prediction can be scored against outcomes only for the period that has elapsed. Cashier employment stood at 3,157,200 in 2024, more than a decade after a 0.97 computerisation probability was assigned to the occupation; the further 9.9 per cent decline to 2034 is BLS model output and cannot serve as an ex-post test. Direction and magnitude score differently. The occupation is declining, and it declined far more slowly than a 0.97 feasibility score was read to imply. Frey and Osborne were explicit that their probabilities express technical feasibility rather than adoption, cost, regulation or consumer preference, so quoting 0.97 as a prediction of job loss misreads their own claim, and the author's judgement, stated as such and carrying no measurement behind it, is that a good deal of retail strategy was built on that misreading. Read correctly, their score is a statement about where an application sits on a trajectory and says nothing about the rate at which it travels.

Changing the unit of analysis changes the answer by a factor of five. Arntz, Gregory and Zierahn applied a task-based measure to the same question and found 9 per cent of US jobs at high automation risk against Frey and Osborne's 47 per cent 46. Their estimate relies on PIAAC self-reported task data, and workers over-report task variety, which biases it downward, so the reading is that the range is wide. The methodological point survives the uncertainty: a retail job is a bundle of tasks with very different exposure, and the unit of analysis decides the answer.

The best available ex-post test found slower growth rather than destruction. Georgieff and Milanez examined 21 countries over 2012 to 2019 and found employment grew everywhere, while employment in occupations at high automation risk grew 6 per cent against 18 per cent in low-risk occupations, with low-educated workers already concentrated in the high-risk group and becoming more so 47. The window closes before generative AI, and the risk classification inherits the limitations of the task-based measure it builds on. Six per cent against eighteen was measured over seven years; sustained, a gap of that size produces a very different labour market without a single layoff, though the extrapolation is the author's and is not something the study measured. Acemoglu supplies the macroeconomic ceiling: a cumulative increase in total factor productivity of no more than 0.66 per cent over ten years, and likely under 0.53 per cent, implying US GDP roughly 0.93 to 1.16 per cent higher in 2034 48. That is a Hulten's-theorem aggregation assuming no new tasks and no reorganisation of production, exactly the channels his own earlier work says matter most, and it is modelled rather than measured.

Two historical analogues are usually offered as reassurance and both need qualification. Basker found that barcode scanning raised retail labour productivity by low single-digit percentages in the years after adoption, and cashier employment grew after scanners spread in the 1980s 55. Larger checkout-time figures circulate in secondary coverage and are not used here. Scanners automated a step inside the cashier's task, while self-checkout removes the cashier from the transaction and hands the work to the shopper, so the mechanism differs and the analogy supports caution about forecasts rather than a prediction about this one. Bessen's ATM case is cited more often still: tellers per branch fell from about 20 in 1988 to about 13 in 2004 while urban branches rose roughly 43 per cent, so total teller employment grew 56. The channel was cost-to-scale, since cheaper branches meant more branches. US teller employment has fallen substantially since the smartphone and US store counts are not rising, both carried here as the author's reading because this review did not locate either series published against the ATM comparison, so the ATM story supports a conclusion only when both halves of its history are carried.

14. H10: what does the entry-level signal show?

Brynjolfsson, Chandar and Chen, using ADP payroll microdata, find workers aged 22 to 25 in the most AI-exposed occupations showing a 16 per cent relative employment decline after conditioning on firm-time effects, while employment for experienced workers in the same occupations held stable 49. In raw terms, 22 to 25 year-olds in the highest-exposure quintiles saw employment fall about 6 per cent from late 2022 to September 2025, against 5 to 13 per cent growth in the lowest three quintiles. Customer service representatives are one of the two named exemplar occupations, which puts the signal adjacent to retail. It is a working paper on a single proprietary payroll panel that over-represents large firms, exposure is an occupational index rather than observed tool use, the authors say they cannot establish causation, and the headline moved from 13 per cent to 16 per cent between drafts.

Set against that, the Budget Lab at Yale finds the occupational mix of US employment has changed about 1 per cent since the release of ChatGPT, within historical range and comparable to the early personal computer and early internet periods, with no discernible economy-wide disruption 33 months in, measured to the study's own endpoint in late 2025 rather than to this report's as-of date 50. Occupational dissimilarity is built to catch reallocation between occupations and is close to blind to task change within one, which is precisely what is happening to cashiers. A null from that measure is not evidence of no effect. Both findings are carried here.

The mechanism H10 names may generate no statistics at all. Internal Amazon documents obtained by the New York Times in October 2025 show the automation team expecting to avoid hiring more than 160,000 US workers by 2027 and more than 600,000 by 2033, saving roughly 30 cents per item packed and shipped, with a stated ambition of automating 75 per cent of operations; executives told the board that robotics could hold US headcount flat while sales doubled by 2033 41. These are leaked planning documents describing an ambition, Amazon disputed the framing, and avoided hires is a counterfactual against Amazon's own growth assumptions that cannot be audited.

Take the mechanism seriously even where the numbers are not auditable, because the institutional consequence is specific. Hiring avoidance produces no layoffs, no WARN notices and no separations. Nearly every publicly funded workforce programme in the United States is triggered by a separation event: WIOA Title I dislocated worker funding, state rapid response teams, and most employer severance obligations 91. A displacement mechanism that never separates anyone is invisible to all of them. The people affected are those who would have been hired, who are disproportionately young and disproportionately entering through the roles the Stanford paper flags. The binding constraint on H10 is therefore regulatory, in the design of the trigger rather than in the size of the appropriation, and the change that moves it is a workforce trigger keyed to entry-grade posting volume. The measurement that would settle the mechanism itself is Amazon's realised US headcount against its sales growth through 2027.

15. H9: what does the assistant do to the worker?

The strongest experimental evidence in this report concerns software assistants rather than robots. Brynjolfsson, Li and Raymond studied 5,179 customer support agents given a generative AI assistant and found issues resolved per hour rose 14 per cent on average, with a 34 per cent improvement for novice and low-skilled workers and minimal effect on experienced high performers 51. The mechanism they identify is a training claim rather than a tooling claim: the assistant propagated the tacit practice of high performers to newcomers. Identification rests on a staggered rollout at one firm rather than random assignment, the setting is scripted text support, and the outcome measured is throughput rather than sales or margin.

Alibaba's field experiment replicates the levelling result at comparable scale under random assignment and adds the distributional term H9 turns on. Ni and colleagues randomised 5,940 after-sales agents, 2,895 treated and 3,045 control, across roughly 2.56 million chats and 0.39 million customer ratings, covering 15.6 per cent of the official customer service workforce 52. Intention-to-treat effects were modest: issue-identification time down 8.2 per cent, chat duration down 1.1 per cent, dissatisfaction down 3.4 per cent against baseline, ratings up 0.042 points. At full usage they reach a 32.3 per cent cut in issue-identification time, 15.4 per cent lower dissatisfaction and a 0.184-point rating gain. Usage fell monotonically from 39.3 per cent among bottom-quintile agents to 18.3 per cent among top-quintile agents, gains concentrated in quintiles one through three, and top-quintile agents got worse on both subjective and objective quality, with higher customer retrials and shift-away time up 23.8 per cent. Almost a fifth of treated agents never used the tool at all.

A universal rollout with universal training therefore applies an intervention that raises the bottom three-fifths of a workforce and degrades the top fifth. The paper is a July 2026 preprint run with the platform whose product is being evaluated, with a co-author employed there, in Chinese after-sales chat rather than on an American selling floor, and the top-performer decline is inferred from shift-away time and retrials rather than observed directly. Discounted for all of that, it remains the largest randomised evidence available and it points the same way as the smaller studies.

Dell'Acqua and colleagues measure the boundary a worker cannot see from the inside. Among 758 BCG consultants working with GPT-4 in 2023, access to the model on tasks inside its capability frontier raised task completion 12.2 per cent, speed 25.1 per cent and human-rated quality 40 per cent, with below-median performers gaining 43 per cent against 17 per cent for above-median performers; on tasks outside the frontier, assisted consultants performed 19 percentage points worse than unassisted ones 53. The researchers constructed the boundary, which is cleaner than any real job has, and the model is a 2023 vintage, so the 19-point penalty is not a current measurement. The transferable finding is that the boundary is invisible from the inside, which places AI training in the category of judgement instruction rather than tool instruction.

How much of this reaches sales is open, and the one direct answer is dispersion. Fang and colleagues ran six randomised generative-AI experiments on an online retail platform and found effects on sales ranging from no detectable impact to plus 16.3 per cent, the largest in the pre-sale service chatbot workflow 54. The authors include platform consultants and an employee of the platform, the setting is online, and none of the six workflows involves an associate on a selling floor. The same technology, in the same firm, produced anything from zero to sixteen per cent depending on which task it touched.

H9 therefore resolves as supported at the largest scale available, with two constraints named. The first is AI algorithms: an assistant trained on aggregate practice encodes something near median practice, which is why it raises the bottom quintiles and degrades the top one. The second is workforce: differentiated assignment requires a performance baseline that most operators do not hold, so measurement of the existing workforce comes before any course rather than after it. The changes that move the boundary are a replication on a spoken, mobile, unscripted selling floor, powered to detect removal of the top-quintile reversal rather than an average gain, with the target being a top-quintile effect indistinguishable from zero on the measure that moved at Alibaba, shift-away time up 23.8 per cent, and with 5,940 agents the order of the sample that reached that quintile; and an assistant whose level of assistance is conditioned on the operator's measured skill, which would remove the quintile reversal rather than manage around it.

Table 8. Frontline AI experiments, ranked by relevance to a selling floor.

16. What happens to a workforce when a format withdraws?

Every change of scale in Table 2 was a workforce event, and none was managed as one in public. When Amazon closed 72 stores, staff were offered redeployment or 90 days of severance at full pay and benefits, a decent package by sector standards, and this review did not locate a disclosure of how many people were affected 1. When Walmart ended the Bossa Nova contract, associates in 500 stores went back to a workflow without a robot in it. When Dollar General pulled self-checkout from more than 12,000 stores, someone had to staff lanes again. The automation and labour literature located for this review is almost entirely about installation. The commerce record says removal and re-siting happen often enough to need their own plan, which is a consequence of a fast iteration cycle rather than of a technology that did not arrive.

Three costs fall on workers when a system is withdrawn. The first is stranded skill. Competence built around a specific system, knowing which exceptions the gate throws on a Saturday afternoon, how to reseat a robot that has parked itself in dairy, which shelf tags the label system silently leaves un-updated, is real expertise, and this review did not locate documentation or credentialing of it. When the system goes, the expertise carries no market value and the worker holds no record that it existed.

The second is role reversion. Systems on the floor do not simply add themselves to an existing job; the job is redesigned around them. A checkout-free store has hosts and exception handlers where it once had cashiers. A self-checkout bank has one attendant covering six lanes and an intervention queue. Reverting that design requires the old role back, and the people who could perform it at speed have moved on or were never hired. Restaffing 12,000 Dollar General stores requires cashiers who can run a lane quickly under pressure, and the pipeline that produced them had been thinned for several years. That last claim is the author's inference from the sequence of events, not a measurement, and this review did not locate published re-staffing data that would test it.

A third cost compounds. In a conventional store the register is where a new hire learns the assortment, the regulars, the price points and the rhythm of the floor, and where supervisors watch performance closely enough to decide who to promote. Automating the register removes the observation post along with the task. The projections show first-line supervisors of retail sales workers declining only 5 per cent 44, so demand for people who can run a floor persists while the route by which they used to be identified narrows. That is a structural change in the internal labour market, and this review did not locate an employment count in which it appears.

Three records can only be made at installation time. The first is what the role becomes, in writing, at the level of tasks and exceptions; the second is a portable record of the competencies the worker acquires; the third is the pre-installation workflow documented well enough that someone who never performed it could restore it, which is a business continuity condition as much as a workforce one. None of the three can be reconstructed afterwards. Whether reversals themselves raise turnover in the affected stores is unmeasured in the sources located for this review, and it is listed among the open questions in section 21.

17. H11: how do you train for equipment that changes yearly?

The training problem in commerce is a throughput problem. BLS models 542,600 cashier openings a year against an average annual net decline of about 31,400, a ratio of roughly seventeen to one, and about 586,000 annual openings for retail sales workers 44. Even in the occupation with the largest projected numeric decline in America, replacement demand exceeds automation by more than an order of magnitude. Openings are dominated by separations and the separations model is calibrated on historical turnover, so if automation produces fewer, better-paid and more skilled floor roles, openings will fall faster than the model implies. Taken as it stands, the number moves the task from retraining the displaced to continuously equipping the churning.

That has an unforgiving consequence for how instruction is designed. At 542,600 cashier openings a year against a stock of about 3.16 million, the entry-level role turns over fast enough that a curriculum sitting outside the work will not be delivered, whatever the training budget says. Instruction has to happen inside the shift, in short units, assessed by doing rather than by seat time. That follows from the openings-to-stock ratio; no median-tenure figure is asserted here, because this review did not locate one for this occupation.

Equipment turnover is the second constraint and the substance of H11. Table 2 contains programmes that launched and ended inside two years, the clearest being Blue Jay, launched in October 2025 40 and reportedly shut down within about six months on a record this report grades weak 39, and vendor interfaces change faster than the equipment does. The ledger publishes a launch-to-change interval for only a few of the eighteen records, and this review did not locate a median programme approval time for any accreditor covering frontline retail instruction, so the statement that a system-specific curriculum has a shorter useful life than the time needed to write it, accredit it and train the trainers is the author's judgement rather than a measured comparison of two intervals. What follows from it does not depend on the size of the gap: teach the invariants and treat the interface as a short, repeatable swap. Six invariants recur across the systems in this report, and together they are the specification in Table 9.

Exception handling comes first. Every deployed system generates exceptions and the human role converges on resolving them: the gate that flags a basket, the terminal that will not accept an item, the robot parked in an aisle, the shelf label that disagrees with the register. This is diagnostic work under time pressure in front of a customer, and it is the most consistently underestimated part of every role redesign in the record. Ground truth maintenance is second. Sensing systems are only as good as the human-maintained planogram, price file and label, and Simbe's claim that 42 per cent of detected gaps are fixed instantly presupposes a person who can act on a detection 14. The competency is understanding what the system believes about the store and how one's own actions change that belief. It is also where the physical quantities in Table 3 land on the worker: a ceiling camera resolving 3.1 to 4.2 millimetres per pixel is following hands rather than reading packages, so what the system knows about an item comes from a shelf weight, a label and a planogram that people maintain.

Boundary detection is third, and the evidence for it is the strongest in this report. Dell'Acqua's jagged frontier shows assisted workers performing 19 percentage points worse than unassisted ones outside the model's competence, with the boundary invisible from the inside 53. A frontline worker cannot tell whether a customer's question falls inside or outside the assistant's frontier. Calibration is trainable through productive failure: route trainees to questions the assistant answers confidently and wrongly, then have them articulate the tell. That transfers across vendors, because the behaviour belongs to the model class rather than to the interface.

Disclosure to the customer is fourth. Staff are asked to explain what a camera does, why a price changed on a shelf edge and what happens to a receipt, usually with no script and no accurate understanding of their own. The Warren and Casey correspondence and the pricing study around it show this is a live public concern even where the evidence favours the retailer 30,31. Fifth is the diagnostic and persuasive half of selling, which the projections say persists: retail salespersons flat at -0.5 per cent while cashiers fall 9.9 per cent 44. It has the strongest existing pedagogy, the weakest existing measurement, and the clearest link to revenue.

Two design decisions follow from the experimental evidence, and they are the least conventional findings here. First, deployment and training have to be differentiated by current performance. The Alibaba result, with usage at 39.3 per cent in the bottom quintile against 18.3 per cent in the top, gains in quintiles one to three, and quality declining in quintile five 52, means a universal rollout applies a treatment that degrades the best performers. Differentiation requires a performance baseline, so measurement of the existing workforce comes before any course rather than after it. That is an instructional design decision and it belongs to whoever owns training.

Second, the credential has to be portable. Turnover means an employer's training investment leaks to competitors within a year, which is the standard reason retail underinvests in it. A vendor-neutral, employer-recognised record of the competencies in Table 9 turns a leaking private investment into a sector asset and makes the churn cheaper for everyone in it. Open-standard verifiable credential infrastructure for this pattern already exists, including the Institute's own, so the binding constraint on H11 is workforce rather than technical: it is agreement on what the competencies are, across employers, rather than the technology to record them. The change that moves it is one multi-employer competency agreement covering the first two rows of Table 9, and evidence that a portable record changes hiring at entry grade.

Table 9. Competency specification for frontline commerce roles under physical AI. Every competency is stated to be independent of any specific vendor system. The evidence anchors cover the ledger records for which this review located one, which are the checkout-free, self-checkout and shelf-scanning entries, together with the frontline experiments, rather than all eighteen records.

18. What would a decisive measurement contain?

Most of the hypotheses in this report are resolvable with data that operators already hold. This section states what a measurement would have to contain to move them, in the order the components depend on each other, because the later components are uninformative without the earlier ones.

The first component is the specification of the comparison itself, fixed before installation: the named comparison stores, the number of store-weeks, the baseline metric and who computes it. The second is the disposition of any labour hours the system frees, meaning whether they leave the payroll or move to another task, which is the single fact that determines whether the technology raises what people can do or removes their hours, and which is not stated in any vendor claim located for this review. The third is a workforce performance baseline, because the Alibaba quintile result makes differentiated assignment impossible without knowing who the top and bottom performers are.

The fourth is a written record of the role as redesigned, at the level of tasks and exceptions, together with a documented path back to the pre-installation workflow. That record is what makes a portable credential possible and what makes a reversal survivable, and neither can be reconstructed afterwards. The fifth is re-measurement at a fixed interval against the same baseline. The first retailer to publish a matched-store comparison of any of these systems will set the standard the rest of the sector is measured against, and this report will be revised against it whichever way the numbers fall.

Three physical measurements would move the selling-floor hypotheses more than any commercial disclosure. An occlusion-driven camera count, published for a real installation, would show how much of the 5.4 to 12 times optical redundancy in Table 3 is required and how much is margin. A per-stream inference load measured on a retail person-and-product pipeline would replace the 1.8 W automotive envelope that carries the sensing-load figures. A published power figure for a shelf-rail camera would replace the enterprise dome camera standing in for it in Table 4 and would fix the crossover interval that separates fixed from mobile sensing.

19. Risk register

The register below covers risks to an operator deploying physical AI in commerce and risks to the workforce affected by it. Likelihood and impact are the author's judgement on the evidence assembled here, and are graded modelled. Leading indicators are chosen to be observable without privileged access.

Table 10. Risk register. Likelihood and impact are the author's judgement, not measurements.

20. Scenarios to 2035

Four scenarios are set out below. They differ in one variable, which is where the machine sits relative to the customer, and they resolve differently for employment and for training. Likelihoods are the author's judgement on the evidence in this report and are graded modelled. All employment figures are anchored on the BLS 2024 base of 3,157,200 cashiers and 3,936,700 retail salespersons 44.

The scenario this report judges most likely is the second, in which sensing and labelling saturate the back of the store and the customer-facing exit stays optional. It is the direction every operator in Table 2 moved once they had run the numbers, and it is consistent with the projections, with the productivity ceiling and with the area-proportional sensing cost in Table 3.

Two signposts would move the judgement quickly. If a large grocer publishes a matched-store comparison showing a shrink or availability effect from computer vision that survives scrutiny, the capital case for full sensing coverage strengthens and scenario three becomes reachable. If Amazon's US headcount flattens while its sales grow, in the pattern its internal documents describe 41, the hiring-avoidance mechanism in H10 is confirmed and the entry-level contraction in scenario four becomes the central case for logistics-adjacent commerce work.

Table 11. Scenarios to 2035. Likelihood is the author's judgement, graded modelled.

21. Which questions are open, and what would settle them?

Several things this report would need in order to be more useful were not located in public. This review did not locate a matched-store controlled comparison of a shelf-scanning robot, a checkout-free installation, or computer-vision loss prevention published by a retailer or a vendor. It did not locate a published disposition of labour hours freed by any of these systems, which is the single fact that determines whether the technology raises what people can do or removes their hours. It did not locate per-store economics for Amazon Go or Amazon Fresh, so the closure of 72 stores cannot be attributed cleanly to the technology.

Four physical quantities are open in the same way, and each carries a numbered figure in this report. This review did not locate an occlusion-driven camera count for any checkout-free installation, so the 5.4 to 12 times optical redundancy in Table 3 stands without a stated requirement behind it. It did not locate a per-stream inference load measured on a retail person-and-product pipeline, so the 1.8 W envelope carrying the sensing-load figures is automotive. It did not locate a published power figure for a shelf-rail camera or for the fourteen cameras on the Tally platform, so the 90 W platform estimate and the 4.5 W rail figure are bounded rather than sourced. It did not locate a machine threshold in millimetres per pixel for separating two similar package faces, which is the resolution requirement the whole checkout-free application rests on.

On the workforce side, four questions are open and answerable with data someone already holds. Does a system withdrawal raise turnover in the affected stores, and by how much? What happens to promotion rates into first-line supervision after the register is automated, given that supervisor demand falls only 5 per cent in the projections 44? Do the quintile effects found in Alibaba's after-sales operation appear on a physical selling floor, where the work is spoken, mobile and unscripted? And does Amazon's actual US headcount trajectory since October 2025 match the hiring-avoidance plan described in the leaked documents 41?

Three items in this report are weaker than the rest and are marked here so they are not carried forward as settled. The Blue Jay shutdown rests on single-outlet reporting accessed second-hand with no Amazon confirmation, and this review did not locate a definition of what shut down means in it 39. The Alibaba top-performer quality decline is inferred from shift-away time and customer retrials rather than directly observed 52. The claim in section 16 that Dollar General's re-staffing was constrained by a thinned cashier pipeline is the author's inference from the sequence of events and has no supporting measurement.

22. Hypothesis resolution

Each of the eleven hypotheses stated in section 3 is resolved below on three axes: where the application sits on its trajectory today, which constraint is currently binding, and what measured change moves the boundary. Binding constraints are named from the five technical classes used in this series, material science, engineering implementation, energy, computation efficiency and AI algorithms, or as regulatory or workforce where that is what binds. No hypothesis resolves to a verdict, because none of these applications is finished and none of the evidence assembled here would support a verdict if they were. A grade column travels with the figures in each row, on the scheme in section 2, so a number met here for the first time is met with its provenance attached, and every modelled figure in the table is computed rather than measured.

Two patterns run across the table. The first is that the binding constraint on the selling-floor hypotheses is engineering implementation more often than it is algorithms or energy, which means the gap is between a capability that exists and an installation that pays for itself at a specific site. The second is that four of the eleven resolve to workforce or regulatory constraints, where the change that moves the boundary is a decision rather than a discovery, and those are the four that could move fastest.

Table 12. Hypothesis resolution. Trajectory position, binding constraint, threshold and figure grade for each of the eleven hypotheses stated in section 3. Modelled entries are computed rather than measured.

23. Conclusions

The selling-floor record between 2016 and 2026 sorts by one variable, which is how much of the change the customer has to absorb, and every application in it can be placed on a trajectory with a constraint named. Electronic shelf labels are furthest along, at 2,300 of the 4,597 Walmart stores the company targets by the end of 2026, and the constraint that remains is installation labour rather than energy, since the display subsystem spends 5.1 to 7.8 per cent of its cell over five years and the hardware carries roughly two orders of magnitude of unused repricing headroom. Checkout-free sensing is commercial and expanding in stadium, airport and breakroom venues, which the area-proportional cost structure predicts are those with the highest transaction rate per instrumented square metre, a reading rather than a measurement since this review did not locate a transaction rate published for any of them, and its cost is proportional to floor area at about 0.62 kW per 1,800 square feet while its installed camera density runs 5.4 to 12 times the optical requirement, so the constraint is engineering implementation and the threshold is a published occlusion-driven camera count. Fixed and mobile shelf sensing are separated by a computable quantity, energy per observation, with a crossover at a revisit interval of roughly 21 to 65 seconds. Centralised robotic fulfilment and store-node fulfilment are separated by achievable order density rather than by machine capability on the operators' own account of the difference, supported by trade-press diagnosis of Takeoff's Chapter 11 filing; the orders-per-site-per-week that would settle it was not located by this review. Manipulation of fresh goods sits earliest on the trajectory, at 66 per cent field reliability on a 33 second reviewed cycle against better than 95 per cent on rigid objects at under 12 seconds a pick, with a damage threshold that is a property of the fruit and a design variable, target compliance, worth a factor of 2.0 to 7.3 depending on the quality criterion applied.

On the workforce side the aggregate is close to flat while its composition moves, and it is the half where a decision rather than a discovery moves the boundary. Retail employment in June 2026 was 15,459,700, below its February 2020 level while national employment grew by 6.7 million 42. BLS models the largest numeric occupational decline in the country in cashiers, and models 542,600 cashier openings a year alongside it 44. Both are true at once, and the second is the operative fact for anyone designing training: half a million people a year enter a role whose content is being rewritten by equipment that this report judges to have a shorter service life than the training cycle, a comparison stated as the author's judgement because this review did not locate a published accreditation cycle time or a published service life for these systems. The role is changing shape rather than disappearing, with transaction handling automated and exception handling, ground truth maintenance and diagnostic selling persisting, and the constraint on that transition is throughput rather than technology.

Experimental evidence says the technology raises what a frontline worker does, and says it does so unevenly enough that a uniform rollout degrades the best performers 51,52,53. That places the design of a rollout, and the performance baseline that has to precede it, with whoever owns training. The competencies that survive equipment turnover are specified in Table 9, the components of a decisive measurement are in section 18, the eleven hypotheses are resolved with their binding constraints and thresholds in Table 12, and the quantities this review did not locate are listed in section 21. Four of the eleven hypotheses, H5, H8, H10 and H11, resolve to a regulatory or workforce constraint alone, meaning the boundary moves when someone decides to move it: a trigger keyed to entry-grade postings rather than separations, a multi-employer agreement on the competency set, a rule on shelf-edge pricing, and the store-against-warehouse split of SOC 53-7065 published alongside the 2026-36 projection round. H9 names workforce alongside AI algorithms, and the matched-store comparison that would move H1 and H2 is a measurement rather than a decision. This report will be revised as each of those arrives, and the Institute will publish the revision whichever way the numbers fall.

References

1. 1 Amazon, corporate statement on refocusing grocery investment on Whole Foods Market and delivery (aboutamazon.com), announced 27 January 2026, page updated 12 May 2026. self-published
2. 2 Newsweek, reporting on the closure of all Amazon Fresh and Amazon Go stores, 27-28 January 2026. reported
3. 3 Axios, coverage of the Amazon Fresh and Amazon Go closures, 27-28 January 2026. reported
4. 4 Supermarket News, store-count reporting on the Amazon Fresh and Amazon Go closures, 28 January 2026. reported
5. 5 SiliconANGLE, Just Walk Out third-party deployment count, 18 December 2025. reported
6. 6 Retail TouchPoints, Just Walk Out third-party location reporting, 2025-2026. reported
7. 7 Axios Nashville, Just Walk Out concessions at Nissan Stadium, 8 January 2026. reported
8. 8 CNBC, Amazon to close all 19 Amazon Fresh stores in the United Kingdom, 23 September 2025. reported
9. 9 CNN, Amazon Fresh UK closure coverage, 23 September 2025. reported
10. 10 Grocery Gazette, Sainsbury's removes Just Walk Out from its Holborn store, 4 February 2025. reported
11. 11 The Grocer, coverage of the end of Aldi's Greenwich Shop&Go trial and Tesco's move of GetGo to a hybrid till format, 2024-2025. reported
12. 12 Wall Street Journal, Walmart ends its Bossa Nova shelf-scanning contract, November 2020; relayed by CNBC, TechCrunch and Retail Dive, 2 November 2020. reported
13. 13 The Robot Report, Bossa Nova layoffs and wind-down of retail robotics operations, 2020. reported
14. 14 Simbe Robotics, tenth-anniversary press release (PR Newswire), 12 November 2025. self-published
15. 15 Fast Company, Simbe client-outcome claims relayed from the company, 2025. self-published
16. 16 Supply Chain Dive, Simbe out-of-stock and fulfilment-time claims relayed from the company. self-published
17. 17 Simbe Robotics, Tally Spot announcement, January 2025; reported by Grocery Dive and The Robot Report. self-published
18. 18 Badger Technologies (a Jabil division), Digital Teammate positioning and deployment count. self-published
19. 19 Kroger, announcement of automated fulfilment network review and facility closures, 18-19 November 2025. self-published
20. 20 Grocery Dive, Kroger Ocado facility closures and impairment, November-December 2025. reported
21. 21 Supply Chain Dive, Kroger-Ocado compensation and 2026 e-commerce profitability guidance, December 2025. reported
22. 22 Forbes, analysis of Kroger's automated fulfilment retreat, 8 December 2025. reported
23. 23 Chain Store Age, Kroger fulfilment restructuring coverage, November 2025. reported
24. 24 Takeoff Technologies, Chapter 11 petition, US Bankruptcy Court for the District of Delaware, 30 May 2024. verified
25. 25 Grocery Dive, Takeoff Technologies bankruptcy and section 363 sale to Woolworths, 2024. reported
26. 26 Walmart investor meeting guidance on automation coverage and unit-cost improvement (Tampa), relayed by Fox Business, Chain Store Age and Retail TouchPoints, 2025-2026. self-published
27. 27 Symbotic, investor relations release on the acquisition of Walmart's Advanced Systems and Robotics business and the 400 accelerated pickup and delivery centre commitment, October 2025. self-published
28. 28 CNBC, Walmart to install electronic shelf labels in all US stores by the end of 2026, 21 March 2026. self-published (company target, via trade press)
29. 29 Grocery Dive and Retail Dive, Walmart electronic shelf label rollout with VusionGroup, 2024. self-published (company target, via trade press)
30. 30 Robert Sanders (UC San Diego Rady), Ioannis Stamatopoulos (UT Austin McCombs) and Robert Bray (Northwestern Kellogg), study of electronic shelf labels and retail price adjustment across 114 stores and 180+ million product-level observations; SSRN working paper, publicised by UC San Diego, 22 July 2025. reported
31. 31 Senators Elizabeth Warren and Bob Casey, letter to Kroger on electronic shelf labels and dynamic pricing, August 2024, and Kroger's response. verified
32. 32 US Bureau of Labor Statistics, Occupational Outlook Handbook entries for cashiers and retail sales workers, Employment Projections 2024-2034, released 28 August 2025. modelled
33. 33 NPR, retailers pulling back on self-checkout, 18 March 2024. reported
34. 34 NBC News, Dollar General, Target and Five Below self-checkout reductions, 2024. reported
35. 35 LendingTree, self-checkout survey of more than 2,000 US consumers, December 2025. reported
36. 36 National Retail Federation, National Retail Security Survey, final edition covering fiscal 2022, and NRF's statement discontinuing the series after 32 editions. reported
37. 37 Retail Dive, analysis of the National Retail Federation's retracted organised-retail-crime statistic and of the survey's discontinuation. reported
38. 38 Everseen, product materials on self-checkout lane coverage, and Forrester Total Economic Impact study commissioned by Everseen, September 2024. self-published
39. 39 Business Insider, reporting that Amazon's Blue Jay station was shut down, relayed by BGR, 5 March 2026. Single outlet, unconfirmed by Amazon. reported
40. 40 Amazon, Blue Jay multi-arm robotic station launch announcement, 22 October 2025. self-published
41. 41 The New York Times, reporting on internal Amazon documents describing automation and hiring-avoidance plans, 21 October 2025. reported
42. 42 US Bureau of Labor Statistics, Current Employment Statistics, series CES4200000001 (retail trade, all employees) and CES0000000001 (total nonfarm, all employees); retrieved from the BLS public API, August 2026. verified
43. 43 US Bureau of Labor Statistics, Current Employment Statistics, series CES4200000011 (average weekly earnings, retail trade), and Consumer Price Index for All Urban Consumers, series CUUR0000SA0 (US city average, all items, not seasonally adjusted); retrieved from the BLS public API, August 2026. verified
44. 44 US Bureau of Labor Statistics, Employment Projections program, 2024-34 occupational projections and occupational separations and openings tables; retrieved from data.bls.gov, August 2026. modelled
45. 45 Carl Benedikt Frey and Michael A. Osborne, 'The Future of Employment: How Susceptible Are Jobs to Computerisation?', Oxford Martin School (2013); Technological Forecasting and Social Change 114 (2017), 254-280. Probabilities read from the appendix ranking table. verified
46. 46 Melanie Arntz, Terry Gregory and Ulrich Zierahn, 'The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis', OECD Social, Employment and Migration Working Paper No. 189 (2016). verified
47. 47 Alexandre Georgieff and Anna Milanez, 'What happened to jobs at high risk of automation?', OECD Social, Employment and Migration Working Paper No. 255, 21 May 2021. Research funded by the Google.org foundation. verified
48. 48 Daron Acemoglu, 'The Simple Macroeconomics of AI', Economic Policy 40(121), 2025, 13-58; NBER Working Paper 32487. modelled
49. 49 Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, 'Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence', Stanford Digital Economy Lab, November 2025 version, using ADP payroll microdata. reported
50. 50 Martha Gimbel, Molly Kinder, Joshua Kendall and Maddie Lee, 'Evaluating the Impact of AI on the Labor Market: Current State of Affairs', The Budget Lab at Yale, October 2025, with subsequent CPS updates. reported
51. 51 Erik Brynjolfsson, Danielle Li and Lindsey Raymond, 'Generative AI at Work', Quarterly Journal of Economics 140(2), 2025, 889-942; NBER Working Paper 31161. verified
52. 52 Xiao Ni, Yiwei Wang, Tianjun Feng, Lauren Xiaoyuan Lu, Yitong Wang and Congyi Zhou, 'Generative AI in Action: Field Experimental Evidence from Alibaba's Customer Service Operations', preprint, July 2026. One co-author is employed by Alibaba Group. reported
53. 53 Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine Kellogg, Saran Rajendran, Lisa Krayer, Francois Candelon and Karim Lakhani, 'Navigating the Jagged Technological Frontier', Organization Science (2025); first circulated as Harvard Business School Working Paper 24-013 (2023). verified
54. 54 Lu Fang, Zhe Yuan, Kaifu Zhang, Dante Donati and Miklos Sarvary, 'Generative AI and Sales Productivity: Field Experiments in Online Retail', working paper, arXiv:2510.12049v5. reported
55. 55 Emek Basker, 'Raising the Barcode Scanner: Technology and Productivity in the Retail Sector', American Economic Journal: Applied Economics 4(3), 2012. verified
56. 56 James Bessen, 'Learning by Doing: The Real Connection between Innovation, Wages, and Wealth' (2015) and 'How Computer Automation Affects Occupations', NBER Summer Institute (2016). reported
57. 57 Devin Coldewey, TechCrunch, first-hand account of the original 1,800 square foot Amazon Go store, January 2018, quoted in 'How Amazon's Convenience Store Uses Surveillance Cameras for a Checkout-Free Experience', Security Sales & Integration, 23 January 2018. reported (a reporter's on-site visual estimate; this review did not locate a camera count published by Amazon for any store)
58. 58 Maeve Allsup, 'How one Silicon Valley startup is using computer vision to make shopping frictionless', Retail Brew, 12 September 2022, reporting 109 ceiling-mounted cameras at the AiFi-powered Loop Neighborhood Market, Union City, California. The article gives no floor area for the store. reported
59. 59 Axis Communications, product datasheets: AXIS M3126-LVE Dome Camera (PoE IEEE 802.3af Type 1 Class 3, typical 4.0 W, maximum 10.5 W) and AXIS P3265-LV Dome Camera (typical 4.8 W, maximum 8.9 W). The 4.4 W working figure used in this report is the midpoint of these two typicals and is the author's arithmetic. self-published (manufacturer datasheets)
60. 60 NVIDIA, DeepStream SDK developer guide, Performance page, giving concurrent 1080p30 stream counts per GPU for a ResNet10 primary detector at 480 x 272 with three ResNet18 secondary classifiers (Tesla T4: 39 H.264 or 64 H.265 streams, at 31 per cent GPU and 5 per cent CPU utilisation). self-published
61. 61 NVIDIA, 'T4 70W Low Profile' product brief, PB-09256-001, stating a 70 W maximum power limit and a 70 W default total board power. self-published
62. 62 US Energy Information Administration, 2018 Commercial Buildings Energy Consumption Survey, food sales buildings: 53.3 kWh of electricity per square foot per year, 163,000 buildings, 6,200 square foot mean size, refrigeration 38 per cent of consumption. verified (primary government series)
63. 63 Simbe Robotics Tally 3.0 published specification (2,000 mm height, 508 mm diameter, 40 kg, twelve vision cameras plus one Intel RealSense depth camera and one HDR camera), as listed in the Leobotics robot specification comparator. reported (vendor specification relayed by a third-party listing; this review did not locate a mass or battery capacity on Simbe's own site)
64. 64 NVIDIA, Jetson AGX Orin Series Modules Data Sheet, DS-10662-001 v1.8, September 2025, Table 1-2 'Jetson AGX Orin SOM Product Summary', total module power: JAOi 15 W, 35 W, 60 W and up to 75 W; JAO 64GB 15 W, 30 W, 50 W and up to 60 W; JAO 32GB 15 W, 30 W, 40 W. self-published (manufacturer datasheet)
65. 65 Panasonic Energy, CR2450 coin-type lithium battery datasheet: 3.0 V nominal, 620 mAh nominal capacity, 0.2 mA standard continuous drain, 24.5 mm diameter. self-published (manufacturer datasheet)
66. 66 SES-imagotag (VusionGroup), VUSION 1.6 BWR electronic shelf label datasheet, product code R16S01: full graphic E Ink display, 152 x 152 pixels at 140 DPI, active display area 27.5 x 27.5 mm, battery 1 x CR2450, battery lifetime 5 years at 1 update per day, radio 2.400 to 2.480 GHz at maximum EIRP 10 mW. self-published (vendor datasheet)
67. 67 Dalian Good Display, GDEM0213B74 2.13 inch electrophoretic display specification: 250 x 122 resolution, 48.55 x 23.7 mm active area, 3 s full refresh, 0.4 s partial refresh, 9 mW refresh power, 0.003 mW standby. self-published (manufacturer specification)
68. 68 DisplayModule, comparative note on display technologies for battery-powered devices, giving memory-in-pixel LCD static power on the order of 10 microwatts against zero static drive power for bistable electrophoretic panels. reported (distributor comparison; no measurement method or test conditions given)
69. 69 Zbigniew Stropek and Krzysztof Golacki, 'The effect of drop height on bruising of selected apple varieties', Postharvest Biology and Technology 85 (2013), 167-172. doi:10.1016/j.postharvbio.2013.05.013. Impacts were made against a rigid flat plate. verified (peer-reviewed)
70. 70 J. Schmalz GmbH, 'Technical Data of Suction Cups' and 'Suction Cup Selection': the specified values are based on a vacuum level of -0.6 bar as well as a dry or lubricated workpiece surface and are given without a safety factor; SUF 90 NBR rated 328 N and SUF 80 NBR 254 N, with the stated caveat that reductions occur where a vacuum level is not reached, for example on porous workpieces. self-published (manufacturer technical data)
71. 71 C. Wouter Bac, Eldert J. van Henten, Jochen Hemming and Yael Edan, 'Harvesting Robots for High-value Crops: State-of-the-art Review and Challenges Ahead', Journal of Field Robotics 31(6), 2014, 888-911. doi:10.1002/rob.21525. verified (peer-reviewed)
72. 72 Jakob Waltman, Ethan Buchanan and Duke M. Bulanon, 'Nighttime Harvesting of OrBot (Orchard RoBot)', AgriEngineering 6(2), 2024, 1266-1276. doi:10.3390/agriengineering6020072. The 85 per cent success and 12 s cycle figures appear in this paper's abstract as prior work from daytime field trials in commercial apple and peach orchards during the 2021 harvesting season; they are not this paper's own measurement, and this review did not locate a primary publication of the 2021 trials. reported (prior-work figures stated in a peer-reviewed abstract)
73. 73 Jeffrey Mahler, Matthew Matl, Vishal Satish, Michael Danielczuk, Bill DeRose, Stephen McKinley and Ken Goldberg, 'Learning ambidextrous robot grasping policies', Science Robotics 4(26), 2019, eaau4984. doi:10.1126/scirobotics.aau4984. Reports better than 95 per cent reliability on heaps of 25 novel objects at more than 300 mean picks per hour. verified (peer-reviewed)
74. 74 Nicolas Hudson, Josh Hooks, Rahul Warrier and colleagues (Amazon Robotics), 'Stow: Robotic Packing of Items into Fabric Pods', arXiv:2505.04572, 7 May 2025. Outcomes are analysed over the most recent 100,000 stow attempts, reported as 85.86 per cent success, 9.31 per cent unproductive, 3.77 per cent amnesty, 0.24 per cent damage and 0.82 per cent other defects, with amnesty defined as either the grasped item or an item already in the pod falling to the floor. self-published (the operator measuring its own production system; not peer-reviewed)
75. 75 Ocado Intelligent Automation, OCADEX/Pick on-grid robotic pick product page, stating a theoretical maximum of 630 units per hour. self-published
76. 76 IEC 62676-4:2014, 'Video surveillance systems for use in security applications - Part 4: Application guidelines', edition 1.0, April 2014, adopted as EN 62676-4:2015 and BS EN 62676-4:2015. This is the edition that defines the DORI operational requirements and their pixel densities. The values used in this report were not read from this document. verified as a document (the standard exists and is identified); the values as used are carried at the grade of the restatement below
77. 77 Axis Communications, 'Pixel density and DORI: Meeting operational requirements in network video', white paper, May 2023, Table 3.1, restating the IEC 62676-4:2014 pixel densities as detection 4 px/face and 25 px/m, observation 10 px/face and 63 px/m, recognition 20 px/face and 125 px/m, identification 40 px/face and 250 px/m, computed on an average human face width of 16 cm, with the note that the specified operational requirements are valid in situations where visual video images are interpreted by human operators and that for video analytics applications or other systems where image analysis is done by software, other definitions apply. The 2014 values quoted in this report were read from this restatement rather than from the paywalled standard. self-published (manufacturer restatement of a standard's table)
78. 78 G. P. Bryan-Brown, C. V. Brown, I. C. Sage and V. C. Hui, 'Voltage-dependent anchoring of a nematic liquid crystal on a grating surface', Nature 392, 1998, 365-367. doi:10.1038/32849. verified (peer-reviewed)
79. 79 J. Cliff Jones, 'The Zenithal Bistable Display: From concept to consumer', Journal of the Society for Information Display 16(1), 2008, 143-154. doi:10.1889/1.2835021. verified (peer-reviewed)
80. 80 D.-K. Yang, J. L. West, L.-C. Chien and J. W. Doane, 'Control of reflectivity and bistability in displays using cholesteric liquid crystals', Journal of Applied Physics 76(2), 1994, 1331-1333. doi:10.1063/1.358518. verified (peer-reviewed)
81. 81 Bob Raikes, 'ZBD's Journey via ESLs to Custom', Display Daily, 1 August 2022, stating that the technology got to mass production with Varitronix by around 2005 with an initial application market in electronic shelf labels, that the ESL market was increasingly moving to EPD displays like E Ink because of the paper-like look, and that in 2019 the ZBD technology was acquired by New Vision Display. The roughly fourteen-year span quoted in this report is arithmetic on those two published dates. reported (trade press; no measurement is given for the stated reason)
82. 82 J. Schmalz GmbH, product data for flat suction cup SUF 20 NBR-55 G1/8-IG, article 10.01.01.14119: holding force 15 N at -600 mbar, nominal diameter 20 mm, maximum diameter 22.8 mm, with the note that the suction force values are theoretical values at -0.6 bar vacuum and a dry, smooth and even workpiece surface, specified without safety factors. self-published (manufacturer product data)
83. 83 Han Fu, Long He, Shaochun Ma, Manoj Karkee, Du Chen, Qin Zhang and Shumao Wang, 'Jazz Apple Impact Bruise Responses to Different Cushioning Materials', Transactions of the ASABE 60(2), 2017, 327-336. doi:10.13031/trans.11946. Three 12.7 mm polyurethane foam cushions over an aluminium plate. Three series are reported, each against its own bare-plate baseline: non-bruising impact level 22 N bare against 95, 160 and 160 N cushioned; Extra Fancy quality maintained in 100 per cent of fruit at 50 N bare against 115, 171 and 180 N; and the fresh-market quality threshold at 95 N bare against 189, 230 and 232 N. The paper converts impact level to an equivalent drop height as a practical measure for designing a catching surface. verified (peer-reviewed)
84. 84 Han Fu, Long He, Shaochun Ma, Manoj Karkee, Du Chen, Qin Zhang and Shumao Wang, 'Bruise Responses of Apple-to-Apple Impact', IFAC-PapersOnLine 49(16), 2016, 347-352. Pendulum-type apple bruising test device, covering the fruit-to-fruit case that a filling bin presents. verified (peer-reviewed)
85. 85 Komarnicki, Stopa, Szyjewicz, Kuta and Klimza, 'Influence of Contact Surface Type on the Mechanical Damages of Apples Under Impact Loads', Food and Bioprocess Technology 10, 2017, 1479-1494. doi:10.1007/s11947-017-1918-z. Golden Delicious apples free-dropped onto rigid plates including concrete and wood. A different group and a different variety from the cushioning work above, providing independent corroboration of the rigid-surface case rather than of the cushioning series. verified (peer-reviewed)
86. 86 CasterHQ, 'Caster Rolling Resistance: Formula and Spec Guide', engineering specifications reference, last reviewed 2 July 2026, defining the coefficient as pounds of push per pound of load and tabulating forged steel 0.01 to 0.015, cast iron 0.015 to 0.02, phenolic 0.02 to 0.03, 95A polyurethane 0.025 to 0.035, 85A polyurethane 0.035 to 0.05, pneumatic 0.04 to 0.06 and 70A soft rubber 0.06 to 0.08, with an overall industrial-caster range of 0.01 to 0.08. self-published (vendor engineering reference; no measurement method or test standard is given)
87. 87 Dave Lippert and Jeff Spektor, 'Rolling Resistance and Industrial Wheels', Hamilton Caster White Paper No. 11. States the coefficient of rolling friction as a length in the form F = f x W / R, giving f = 0.047 inches for 85A polyurethane and 0.019 inches for forged steel at 1,200 lb per wheel on a steel floor at 3 mph, and noting that polyurethane has a range of values depending on the specific material selected. The dimensionless coefficient is f divided by wheel radius, so it depends on wheel diameter. This source and the reference above disagree for 85A polyurethane. self-published (manufacturer technical paper)
88. 88 Han Fu, Long He, Shaochun Ma, Manoj Karkee, Du Chen, Qin Zhang and Shumao Wang, 'Impact Bruise Responses for Jazz Apple on Different Materials of Catch Surface', 2016 ASABE Annual International Meeting. doi:10.13031/aim.20162461461. Reports the same 22, 95, 160 and 160 N non-bruising series and the 115, 171 and 180 N Extra Fancy series as the journal paper above, and is the catch-surface companion to it. verified (peer-reviewed conference paper)
89. 89 IEC 62676-4:2025, 'Video surveillance systems for use in security applications - Part 4: Application guidelines', current edition, superseding IEC 62676-4:2014. Replaces the DORI ladder with seven operational grades: overview 20, outline 40, discern 80, perceive 125, characterize 250, validate 500 and scrutinize 1,500 pixels per metre. The 250 px/m value survives under the name characterize and the density recommended for identification moved to 500 px/m. The values used in this report were not read from this document. verified as a document (the standard exists and is identified); the values as used are carried at the grade of the restatement below
90. 90 Axis Communications, pixel density white paper, 2025 edition, restating the IEC 62676-4:2025 seven-grade ladder and repeating verbatim the note that the specified operational requirements are valid where visual video images are interpreted by human operators and that other definitions apply where image analysis is done by software. The 2025 ladder values quoted in this report were read from this restatement rather than from the paywalled standard. self-published (manufacturer restatement of a standard's table)
91. 91 Workforce Innovation and Opportunity Act, Public Law 113-128: the dislocated worker definition at 29 U.S.C. 3102(15), which requires termination, layoff or notice of either; rapid response activities at 20 CFR part 682 subpart C, triggered by a permanent closure or mass layoff; and the Worker Adjustment and Retraining Notification Act, 29 U.S.C. 2101-2109, triggered by a plant closing or mass layoff. Read for the eligibility trigger only. verified

Market and economic review, not investment advice. Figures carry the verification grade under each reference: verified (peer-reviewed or independently replicated), reported (a named source stated it, not independently confirmed), self-published (the organisation's own figure) and modelled (computed here, not measured). A modelled figure is never presented as a measurement. Corrections are welcome and will be recorded.