Computer vision in logistics has moved out of the pilot phase at a handful of large operators, and the published record now includes real operational numbers. Maersk cut its inventory count error rate from 4% to zero across four pilot sites and then committed to a network rollout. Amazon reports its damage-detection models catch defects three times more often than a person doing the same check. 

A DHL Supply Chain site in Ohio runs at 99.96% inventory accuracy and cut picker training from two weeks to about an hour. Below, five deployments where the operator disclosed what changed, what the vision system does, and the design decisions behind the result. The pattern across all five: the model was never the hard part.

Why We Selected These Five

Most articles on computer vision in logistics list logos. You get DHL, Amazon, and UPS as evidence that the category exists, a paragraph on how cameras can read barcodes, and no numbers you could take to a CFO.

We applied three filters. The operator had to be named. The result had to be disclosed by the operator or reported by a trade publication rather than sitting only on a vendor's homepage. And the system had to be running in production, not sitting in a proof of concept.

That filter cut most of what gets published. It left five deployments that answer different questions: audit inventory you cannot easily reach, catch damage on rare-event volumes, read labels moving faster than a human eye can track, guide a picker without slowing them down, and measure a physical object well enough to write the number into a WMS.

We also flag where a figure comes from a vendor rather than the operator. Vendor numbers are useful directionally. They tend to describe the best site rather than the estate.

One more note on evidence. The figure everyone quotes about AI failure, MIT Project NANDA's finding that roughly 95% of enterprise pilots produced no measurable P&L impact, comes from a study of generative AI. Computer vision in logistics sits in a different evidence position, because you can walk onto the floor with a caliper and check whether the system is right. That verifiability is the reason this category has case-level numbers at all.

Why Computer Vision in Logistics is Harder Than It Looks

Detection is not measurement

Computer vision sits in a different category from the AI agent work that dominates most vendor roadmaps right now. An agent reasons over text and calls tools. A vision system in a warehouse has to be right about a physical object that someone will pick up thirty seconds later, and a caliper settles the argument.

Most warehouse vision projects answer one question: is there something on the rack? That is a detection task, and a competent team solves it in weeks with an off-the-shelf model and a few thousand labelled frames.

The brief gets much harder when someone asks how big the thing is. Recovering length, width, and height from an image, in centimetres, with a stated tolerance, is metrology built on top of computer vision. Pixels carry no scale. You supply the scale through calibrated scene geometry, and the quality of that calibration sets your accuracy ceiling before the model gets involved.

Our own deployment made this concrete early. We started with bounding boxes, which is where most teams start, and a carton sitting at a slight angle on a pallet reported several centimetres of extra length. The box is axis-aligned; the carton is not. Switching to instance segmentation, so the pipeline works from the object's silhouette instead of a rectangle around it, removed that error class entirely. That single decision mattered more to measured accuracy than any model swap we tried afterwards.

The constraint set decides the architecture

Ask a vendor how their system works, and you get a diagram. Ask an operator what they were allowed to build and you get the real design.

On the 100+ site estate we worked on, three constraints arrived before any technical discussion. Raw video could not leave the site, because of site-level privacy obligations and tenant agreements at several depots. Depot-to-centre bandwidth varied across the estate, with some links unreliable enough that a cloud round trip was never realistic. And the system could not add a single second to anyone's cycle time, because pickers, putaway and receiving staff were already working to a rate.

Those three sentences determined the architecture. Inference runs on a GPU node inside each depot. Only structured measurements and variance events cross the wire, which turns continuous video into kilobytes. Each depot buffers locally through an outage and syncs in order when the link returns, so no central blind spot opens. Capture stays passive, pulling from existing CCTV over RTSP plus a few dedicated cubing-angle cameras at receiving.

We enforced the video boundary at the network layer rather than in policy documents: bandwidth caps and outbound rules block raw frame egress, so a misconfiguration cannot quietly push footage off-site. Policy alone would not survive a busy quarter.

Honest uncertainty beats confident wrongness

A vision system that writes a wrong number into the WMS is worse than no system, because it corrupts the record everyone else trusts. Once a planner catches the system putting bad cube data into slotting, they stop trusting all of it, including the parts that work.

Every measurement in our pipeline carries a confidence envelope rather than a bare number. Below a per-SKU-class threshold, the system marks the measurement advisory and asks a supervisor to confirm with one tap. High-confidence cubes write back to the WMS with no human involved. Confidence drops on irregular loads, which is correct behaviour: a shrink-wrapped mixed pallet with cartons overhanging the footprint deserves less certainty than a standard tote.

Watch for the same pattern in all five case studies below. Every deployment that reached production hands its uncertain cases to a person before anything downstream acts on them.

The Five Case Studies

1. Maersk and Verity: auditing inventory nobody can reach

Operator and scale. Maersk North America, through its Performance Team contract logistics arm, running large pallet-storage warehouses.

The problem before. Counting pallets meant sending staff up on ladders and order pickers to read labels at height. Maersk described the work as difficult, repetitive and tedious, which is also a fair description of conditions under which people make counting mistakes.

What the vision system does. Verity's drones fly autonomously between pallet positions at night or on weekends, with the overhead lights off, reading barcodes at any height through onboard high-resolution cameras. The drones localise themselves visually rather than through infrastructure, return to a charging pad when needed, and the system reconciles what it saw against the WMS before pushing exceptions to a dashboard. Operator training takes one day.

Measured results. DC Velocity reported that the count error rate at Performance Team sites fell from 4% to 0%. After six months across four sites, Maersk decided to roll Verity out to every warehouse in its network with pallet storage. Maersk's investment arm then led a CHF 30 million round into Verity, which is the least ambiguous signal a pilot can produce. Verity's own site claims 200+ facilities at 99.9% accuracy with a 60% stockout reduction; treat those as vendor figures. In a separate three-month pilot with the footwear brand On, Verity flew over 1,500 autonomous missions and performed more than 80 million RFID reads across roughly 1.25 million tags, scanning up to 1,000 tags a second at 99.9% accuracy.

Reading the numbers. A 4% count error rate means roughly one in every 25 pallet positions disagreed with the WMS. In a 20,000-position warehouse, that is 800 locations wrong at any given moment, and each one is a live trap: a picker travels to a slot that is empty or holds a different SKU, the order ships short, and someone spends an afternoon reconstructing what happened. Read the 0% as zero errors on the counts the drones performed, not as a claim that error has been abolished. It is a rounded operational figure, and the honest version is closer to "below the level our audit process can detect."

On the RFID figures, separate read rate from localisation. Reading 1,000 tags a second tells you those tags exist somewhere within antenna range. Knowing which pallet position holds each one is a harder problem, and it depends on the drone knowing its own pose. Vendors quote read rates because they are large. Ask about position accuracy.

The transferable lesson. Maersk did not evaluate a technology. They ran four sites for six months, measured against their own baseline, and let the number decide the rollout. Six months is roughly the shortest honest window for an inventory system, because you need to see it survive a peak and a physical count cycle.

2. Amazon Project P.I.: catching a defect that appears once in a thousand items

Operator and scale. Amazon's North American fulfilment network moves around 8 billion packages a year.

The problem before. Associates checked items for damage while picking and packing, under rate expectations, as a secondary task. Jeremy Wyatt, director of applied science at Amazon Robotics, put the difficulty plainly: the check is cognitively demanding because you are looking for something rare and it is not your primary job.

What the vision system does. Items route through an imaging tunnel, where Project P.I. evaluates colour and monochrome images for physical damage, torn packaging and bent covers. OCR models read expiration dates. Separate models compare the item against catalogue reference images to verify colour and variant. Flagged items go to an associate, who decides whether the item can be resold through Second Chance or donated. Amazon later added a multimodal LLM that describes damage in plain language and summarises recurring damage patterns, which lets the team take specific evidence back to a manufacturer or seller instead of a complaint.

Measured results. Amazon states the system identifies damage three times more effectively than manual inspection, and set out to scan more than 40 million customer products a month across a dozen operations in North America and Europe. Fewer than one in 1,000 items Amazon handles is damaged. Item-level defect detection through imaging tunnels has been rolling out since May 2022.

Reading the numbers. The base rate is the whole story here. At a 0.1% damage rate, an associate at an imaging station sees roughly one damaged item per thousand while working to a rate on a different primary task. Human recall on rare visual signals under time pressure is poor, and it gets worse across a shift.

"Three times more effective" almost certainly describes recall, meaning the share of truly damaged items caught, rather than overall accuracy. Overall accuracy would be a useless metric here: a model that flags nothing scores 99.9%. Run illustrative arithmetic on the disclosed figures. Forty million items a month at a 0.1% damage rate is about 40,000 damaged items. If a person catches a quarter of them, tripling recall moves you from roughly 10,000 intercepted to roughly 30,000. Amazon has not published its baseline, so treat the arithmetic as a way to feel the scale rather than a fact.

The constraint that decides whether this works is false positives. Every flag consumes an associate's time. A model with excellent recall and a 5% false-positive rate would generate two million reviews a month and get switched off. The human review queue, not the model, sets the ceiling.

The transferable lesson. Vision earns its place fastest on tasks people are structurally bad at: rare events, sustained attention, and judgements made while doing something else. If you are looking for a first use case, look for the check your team performs as a side task.

3. PepsiCo and KoiReader: reading labels faster than a person can track them

Operator and scale. PepsiCo distribution, starting at a Dallas-Fort Worth facility and extending across sites.

The problem before. Warehouse labels and barcodes arrive at any size, at any angle, sometimes partially covered or damaged, on conveyors moving at production speed. Conventional scanners need the code presented within tolerance. Anything outside that becomes a manual exception, and exceptions at conveyor speed create backpressure through the whole line.

What the vision system does. KoiReader's AutonomousOCR, built on NVIDIA Metropolis, reads complex labels on fast-moving belts without requiring a controlled presentation angle. PepsiCo then extended the same technology to yard operations, automating identification as tractors and trailers enter and exit the Texas distribution centre, with a further step toward validating customer deliveries against what was picked.

Measured results. KoiReader started with PepsiCo in 2021 and expanded application by application from there. Greg Bellon, senior director of digital supply chain at PepsiCo, framed the goal as throughput: find the right lever and you move it. Once accuracy reached near-perfect, PepsiCo widened the scope to validating human-assisted picking. KoiReader claims a minimum of 99.9% accuracy across use cases and nearly $2 million saved per facility from eliminating picking errors in real time. Both are vendor figures.

Reading the numbers. The gap between 99% and 99.9% on a read task sounds cosmetic and decides headcount. Take a line reading 10,000 labels a shift. At 99%, you generate 100 exceptions a shift, which needs a person standing there resolving them. At 99.9%, you generate 10, which a supervisor absorbs between other tasks. At 99.99% you stop thinking about it. Nothing about the model changed conceptually across those tiers, but the operating model did.

The yard extension is the more interesting number, and PepsiCo did not publish it. Gate processing is where dwell time accumulates. If OCR removes a manual check-in step, you recover minutes per truck, and minutes per truck across a day compounds into dock scheduling that holds.

The transferable lesson. PepsiCo did not buy a warehouse vision platform. They solved label reading on one line, proved the accuracy, then reused the same OCR capability at the gate. Four years of expansion from one narrow problem. The teams that struggle are the ones that scope five use cases at once.

4. DHL Supply Chain: vision picking, and the metric nbobody forecasts

Operator and scale. A DHL Supply Chain facility in Lockbourne, Ohio, averaging 300 orders a day and reaching 30,000 to 50,000 orders a week at peak.

The problem before. The site ran paper-based workflows alongside systems that made workers step through several screens and menus to complete a pick. Both slowed execution and made onboarding harder, and the strain showed as volumes climbed.

What the vision system does. DHL deployed vision picking on wearable smart devices using TeamViewer's Frontline Pick, for both case picking and e-commerce. Instructions appear in the worker's field of view: what to pick, where to go, where to place it. Confirmation happens through the device, and inventory checks fold into the pick rather than running as a separate pass. Courtny Arnold, an operations manager at DHL, described the interaction as scanning the item, being told how many to pick, and seeing on the headset where each one goes.

Measured results. The facility reports 99.96% inventory accuracy. Training dropped from up to two weeks to roughly one hour. The same approach now covers domestic and international operations at the site and can extend to others.

Reading the numbers. Do the arithmetic on 99.96% before you treat it as perfection. At 30,000 to 50,000 orders a week, four hundredths of a percent leaves 12 to 20 problem lines a week. That is the point. DHL landed on an error volume a supervisor can work through by hand, which is the useful way to set an accuracy target: pick the number where your exception queue stays inside one person's day, rather than the number with the most nines in it.

The training figure deserves more attention than it gets. Two weeks of ramp for a seasonal cohort of 200 workers costs roughly 16,000 hours of low-productivity labour plus the experienced staff pulled off the floor to train them. One hour costs about 200. Beyond the money, it changes who you can hire and how late you can hire them, which is the constraint that bites in October. Almost nobody puts ramp time in the business case, and it is often the largest single line.

The transferable lesson. State plainly what this system is. Vision confirms the pick and reads the item; the routing and guidance come from the WMS and the device. Calling it autonomous vision would oversell it. The gain came from removing screens and paper from a worker's hands, and vision made the confirmation step free.

5. Continuous stock visibility through computer-vision measurement: 100+ depots, measured rather than counted

Operator and scale. A US multi-site distribution operator running more than 100 warehouses and depots, with a roadmap past 500 sites. Full case study here. Client identity and exact figures are generalised at the client's request; the architecture and measurement approach are as delivered.

The problem before. Stock visibility ran on cycle counts against a schedule, barcode scans at receiving and putaway, and a WMS that treated those inputs as truth. That playbook holds at smaller scale. At 100+ sites with an expansion roadmap, blind spots accumulated faster than the operations team could close them.

Between counts, three failures repeated. Empty pick-faces went unnoticed until a picker arrived at a gap. Mis-stows, where inbound goods land in the wrong location, surfaced days later after downstream picks against that SKU had already failed. And dimensional data for inbound goods was missing or keyed by hand at receiving, so cubing errors fed straight into slotting and load planning.

What the vision system does. Cameras at each depot feed a five-layer pipeline. Per-camera calibration against fiducial markers at receiving and reference objects already in the workflow (totes, pallet footprints, shelf beams) locks the pixel-to-millimetre relationship and removes lens distortion. Instance segmentation isolates the carton, parcel, tote or pallet load from background and neighbouring stock. Depth comes from the sensor where stereo or RGB-D cameras exist and from three combined signals where the CCTV is monocular, which covers most of the estate: the calibrated ground plane, in-frame reference objects at known scale, and multi-view fusion when an object passes more than one camera. Length, width and height follow from the silhouette under the calibrated geometry; volume follows for cuboids and comes from a reconstructed volumetric model for irregular loads. Every measurement carries a confidence envelope.

Structured events flow to a central reconciliation engine that matches them against WMS state and routes variances to a named operator: pick-face fill state, putaway confirmation, inbound cube, and floor inventory staged outside its nominal location.

Measured results. Against the operator's own pre-deployment baseline, tracked at each rollout wave: hand-keying of cube data dropped to the residual low-confidence cases, with cube accuracy holding against caliper ground truth. Pick-face stockouts fell, because replenishment now prioritises against live state instead of the last cycle count. Mis-stow latency moved from days to minutes, surfacing inside the same shift. Better cube data flowed downstream into slotting and load planning. Percentage figures are withheld at the client's request.

Delivery took an eight-person team over 12 months: a system architect, two ML and CV engineers, backend engineers on the reconciliation engine and WMS integration, DevOps for edge image builds and fleet provisioning, QA running the accuracy harness, and a project manager coordinating rollout waves.

Reading the numbers. We achieved sub-centimetre accuracy on warehouse cartons, and that phrase means very little on its own. Any dimensioning claim needs three parts: the tolerance per axis, the confidence level, and the shape of the error distribution. A system that measures correctly on average and misses by four centimetres on 5% of pallets will break load planning, because load planners work to the worst case rather than the mean. Report the distribution.

What produced the accuracy was not the model architecture. It was calibration quality, segmentation rigour, honest confidence reporting, and a validation harness that samples measurements and compares them against physical re-measurement with calipers or a certified cubing station. Those deltas feed retraining and per-site error budgets. Drift detection watches calibration against known scene geometry and triggers re-calibration on its own, because cameras get bumped, rack beams get moved and lighting fixtures get replaced, and no site is going to stop work so you can re-shoot a calibration target.

The transferable lesson. The measurement discipline transfers across industries; the delivery surface does not. We run the same five-layer pipeline server-side behind a cloud API for consumer-facing sizing work, and in medical imaging for anatomical measurement. Calibration, segmentation quality, depth recovery, multi-view fusion, confidence reporting and ground-truth loops stay constant. Edge nodes versus cloud follows the constraint set.

Five Deployments at a Glance

Key Takeaways

Continuous signal beats better periodic inspection. Every result above comes from replacing a scheduled check with a constant one. Maersk did not count more accurately; they counted continuously. Our engagement did not improve cycle counts; it made cycle counts stop being the only source of stock truth. Gather AI reports that GEODIS scans locations 15 times faster than manual methods, and Taylor Logistics cut cycle counting time 87% while moving accuracy from 97-98% to nearly 99%. Same shape.

Your existing cameras are probably enough. Most positions across the estate we worked on were monocular CCTV. Depth-capable sensors sat only at receiving and a few high-value pick zones. Calibrated scene geometry plus in-frame reference objects recovered depth everywhere else. Budget for calibration discipline before you budget for hardware.

The human in the loop is a design feature. Amazon routes flagged items to associates. PepsiCo validates human-assisted picking. Our pipeline queues low-confidence measurements for one-tap confirmation. Systems that write uncertain values into the system of record destroy the trust that makes them useful.

Track accuracy; do not assume it. A vision system degrades silently. Cameras shift, lighting changes, SKU mix turns over, and nothing throws an error. Sampled physical re-measurement, per-site error budgets and automated drift detection are what separate a production system from a demo that used to work.

Second-order benefits often exceed the primary one. DHL bought accuracy and got a training-time change worth more. We built cube measurement and the largest downstream effect landed in slotting and load planning. Ask what improves two systems away from the one you are fixing.

Constraints come before architecture. Data residency, link reliability and cycle-time pressure decided our edge-first design before anyone opened a model zoo. Write your constraints down first. If a vendor's architecture does not change when you read them out, they have not heard you.

Pilot-to-network is the only test that counts. Maersk ran four sites for six months against their own baseline and then committed the network. That sequence, baseline, bounded pilot, measured comparison, rollout decision, is the pattern worth copying.

Where This Leaves You

Computer vision in logistics works where the problem has a defined boundary and someone stated the accuracy budget in advance. Maersk, Amazon, PepsiCo and DHL each picked one narrow failure, measured what it cost them, and let the numbers decide the rollout. None of them started with a platform.

The projects that stall treat this as a model selection exercise. They skip the baseline, skip the calibration discipline, ask the floor to change how it works, and then wonder why a system that demoed at 99% cannot survive a quarter. Calibration, segmentation quality, honest confidence reporting and a ground-truth loop do the work. The model is the part you can buy.

We build these systems as custom software engagements rather than licensed products, because the constraint set differs at every operator. If you are scoping something in this space and want to pressure-test the approach before committing budget, our CTO takes those calls directly.