# AI/ML models or site for FMCG goods performance Topic radar · 5,298 documents · 2021-01-01–2026-11-02 · updated 2026-10-31 Canonical: https://slate.greyb.com/radar/aiml-models-or-site-for-fmcg-goods-performance/innovations ## In one minute ### Reading a shelf is an alignment problem, not a detection problem Detection is commodity. The filings name their backbones openly — YOLOv5, EfficientDet and ResNet50 in one compliance system, YOLO localisation followed by a CNN or vision transformer classifier in another [1] [2]. What the documents spend their claims on is the comparison against the intended layout. AbleAI borrows sequence alignment from bioinformatics, running a modified Needleman-Wunsch over the detected row and expanding the classic match/missing/added verdict into eight states that include position change, remote placement and position shifting, at roughly 190 ms per image [2]. The point of eight states rather than three is to stop a shifted product being read as a sold one. Tata Consultancy goes the other way, matching feature points against product templates, scoring proposals with a classifier, then using graph-based non-maximal suppression to discard ambiguous detections before assembling a "realogram" [3] [4]. Target stitches several frames into one shelf image, paints every bounding box a common pixel value and subtracts a reference background, so an anomaly is literally what remains [5]. A published system groups detections into rows with DBSCAN and assigns them to planogram slots with the Hungarian algorithm [6]. Every one of these is a different comparison operator, and none is comparable with the others. That is why the reported accuracy figures in this area cannot be ranked against each other. 1. [WO2025251114A1](/?paper=WO2025251114A1&utm_source=radar&utm_medium=dashboard) 2. [KR102909279B1](/?paper=KR102909279B1&utm_source=radar&utm_medium=dashboard) 3. [US12307420B2](/?paper=US12307420B2&utm_source=radar&utm_medium=dashboard) 4. [EP3767527B1](/?paper=EP3767527B1&utm_source=radar&utm_medium=dashboard) 5. [US12136061B2](/?paper=US12136061B2&utm_source=radar&utm_medium=dashboard) 6. [10.61591/jslhu.22.714](https://doi.org/10.61591/jslhu.22.714) ### Brand owners publish plant numbers, not shelf numbers Where a manufacturer can point at equipment it owns, the numbers are specific. Unilever's ice-cream supply chain runs 35 factories and roughly three million freezers across 60 countries, and the company credits AI weather analysis with cutting waste through adjusted sales forecasts and data from 100,000 connected freezers with lifting retail orders and sales by up to 30%. Kellanova put $4–5m into sensors, machine learning and a digital twin of the dough on the Pringles line at Kutno, and reports 10% more capacity from existing equipment and a return above 40% after four years of development with Siemens. A smaller pilot through Unilever's accelerator cut machine cleaning time 20% and utility use 10% at one foods factory, worth €100,000 a year. ### One conglomerate is assembling the software layer alone SoftBank Group's 2026 applications, most of them titled simply "System", arrive as a single bloc with a common construction: retail data goes in, a generative model sits in the middle, and a recommendation comes out. Demand forecasting for restaurants and retailers blends sales history with weather, footfall and store-camera demographics. A companion filing proposes menus and issues order instructions against live inventory; another wraps the same loop in supply-chain management with an emotion engine reading customer facial expressions. An LLM sets and then dynamically adjusts inventory levels. ### Identity is taken from the shelf edge, not the product The harder recognition step is deciding which of two near-identical packs a box contains, and the industry filings answer it by reading the label rather than the packaging. Walmart extracts text from shelf tags and attributes from the items and maps the two together, catching a mismatched price tag as a distinct anomaly class. Urstore matches against GTINs using image embeddings from a MobileNet-based similarity class, with the whole model configured to run offline on a handheld. Trax disambiguates lookalikes by their price ranges, and Salesforce skips annotated training data entirely by segmenting shelves and reading product names off the image. ### The store-sensing layer is drawing in the parties that own the radio Turkcell's filings treat the store as a sensing problem rather than a retail one. One uses millimetre-wave fixed wireless units as radar, reading channel state to track customer movement and derive placement and marketing metrics. Another puts microphones and cameras at the counter to read speech and facial expression and prompt staff with a suggested response. A third builds campaign targeting from online, offline, telecom and sensor data combined, and the same filer has a supply-chain recommendation system classifying supplier and sentiment risk. ### The price tag is becoming the one place where shelf sensing and pricing meet A cluster of filings puts the electronic shelf label at the centre of the store rather than at its edge. One matches products to labels from shelf images and runs a reinforcement-learning price on inventory, sales, competitor prices and shelf-space share, then flags the gap between the recommended price and the displayed one as a "price lag" and updates the tag. Another positions the monitoring cameras themselves with a learned model, and detects missing products and mismatched labels from the same image. Toppan generates display content from past promotion performance, Trax drives transparent displays from detected availability, and a Korean filing puts the camera and facial analysis inside the tag itself. ### The loss function is where the operational judgement gets encoded Several documents are less about accuracy than about how many alerts a system is allowed to emit. A shelf-void detector uses an asymmetric cost function that penalises a missed gap more heavily than a false alert, reaching mAP@0.50 of 0.875 at about 41 ms per frame. An end-to-end empty-shelf pipeline reports a mean average F1 of 68.5% and argues the binding constraint is annotation quality rather than architecture. Trax patents the gate directly: a notice-override condition suppresses an alert when earlier frames show the same state, and product models are updated automatically when candidate confidence falls below a threshold. Zebra fires the camera only when a shelf sensor crosses a threshold; Fujitsu fuses image and depth through four separate determination indices before declaring a stockout. ### Retail software's deliverable is an equipment instruction Blue Yonder's allocation work computes a constrained quantity from a reinforcement-learning reward over margin, carrying cost and opportunity cost, then transmits instructions to automated warehousing equipment at a distribution centre to retrieve the product; a later filing models allocation as a state-action process and again hands the retrieval instruction to DC automation. Its planogram line ends the same way, with simulated annealing over layout options passed to robotic shelving that executes the arrangement. ## Overview The five areas turn out to be one chain rather than five markets: what will sell, what state the product is actually in, what it should cost, what to order, and which shopper sees which offer. Follow it end to end and the pattern is consistent. Each joint has a workable technique, the techniques are widely published, and the value has migrated to the two ends of the chain — the record the model learns from and the rule that turns its output into an act. ## What's new Two commercial launches, two peer-reviewed reviews, a worked measurement paper and one simulation that measured its own parts arrived in four weeks, and they all address the same question: what evidence stands behind a claim. The month's load-bearing documents describe, test and audit systems rather than propose new ones, and the filings that usually carry a month on this topic were thinner than the argument. ## Research ### Shelf-life models are fed by kinetics, not by spoiled data The freshness filings have converged on the same construction: use a physical degradation model to produce the number, and a network to interpolate it. Tata's rate-kinetic method fits zero-, first- and second-order reaction equations, picks the order by coefficient of determination, and takes the minimum shelf life across attributes; a later filing bins storage conditions into weather zones by lookup and derives Arrhenius-form equations from them. For millet flour the kinetic model generates the entire training set by curve-fitting rate constants from literature data, and an LSTM learns that. A Chinese filing integrates a modified Arrhenius model over the temperature trace to get nominal life, corrects it with an LSTM, and feeds the result into an outbound priority queue that re-orders shipments when a later batch is degrading faster than an earlier one. Walmart recalculates sensitivity to temperature, humidity and ethylene deviations with machine learning and compares shipment imaging against a control load. Bohai corrects hydrogel label colour for glare and reflectance before any of it reaches the prediction stage. ### Pricing is a demand model plus a policy, behind a confidence gate Walmart's markdown filing is the clearest statement of the architecture: one model estimates demand from sales observed at experimental price points, a reinforcement-learning policy maps state to markdown action and optimises revenue minus cost, and a confidence threshold decides routing — above it the price goes live, below it the item is sent to test stores. The gate is the deployment design, not an afterthought. ### The forecast's deliverable became a distribution, and the policy simulator took over The forecasting papers no longer stop at a point estimate. One feeds LSTM residuals into a nonparametric error distribution and carries that uncertainty into quartile-based safety stock. Another uses LightGBM quantile regression and runs the output through an (s, S) policy optimised by simulation, finding policy parameters and cost assumptions moved service levels more than forecast accuracy did. A third couples a random forest to an inventory-risk layer whose selected parameters shift when the assumed stockout penalty is raised. On the sparse end, a survey finds Croston variants still competitive with deep models on intermittent series and tests a hybrid of classification and regression trees. ### Personalisation is cohort-then-individual, and it is moving onto the customer's device ByteDance's two-stage design is the reference pattern: a first model infers attributes for a consumer from cohort-level parameters, then a cohort-specific tree ensemble is applied to that consumer's own values. Groupon filed essentially the same architecture earlier, using a random forest to rank attribute importance per cohort. Target takes the transfer route instead — a base network trained across many purchase categories, tuned per promotion task, with relevancy predictions then combined with advertising revenue to score the promotion. ### What gets claimed is the switch, not the model Two enterprise filings state the pattern plainly. One converts predicted scenarios into executable business rules that drive decisions; the other builds the rule set itself by multi-objective optimisation, trading black-sample precision against asset-loss recall with a weighting hyperparameter. Put alongside the shelf-notification overrides and the pricing confidence gates, the patented artefact across this field is consistently the threshold that converts a score into an action. ## Market ### Brand money is going to the far end of the chain Greyparrot launched Deepnest to give brands direct access to their post-use packaging data, with Unilever, Asahi and Amcor named as initial triallists. A Consumer Goods Forum and Bain study found roughly three in ten businesses already using AI somewhere across the packaging lifecycle, with design the clearest case. ### What is being sold is advice before it is being sold as software The pricing case is argued from complexity and regulatory exposure rather than demonstrated. BCG frames AI pricing as potential weighed against retail complexity, and a Grocery Dive column makes the same point with an explicit warning about regulatory moves and consumer sentiment. The transformation side reads the same way: consultancies on where margin goes and vendor pages promising yield and line utilisation. ## Strategy ### The forecasting effort has moved onto the data, and off the model The research is no longer arguing about architectures. A hybrid clustering framework segments products by demand structure and lifts accuracy by adapting the model per segment, using 3.8 million weekly records from over twelve thousand products. A high-resolution brick-and-mortar comparison finds tree ensembles on non-imputed local groups beating N-BEATS, NHITS and the Temporal Fusion Transformer, and a separate study reaches the same verdict with LightGBM over LSTM and TCN once promotions and competitor prices enter the features. ### Freshness sensing is being designed as an input to markdown, not as a quality instrument The freshness filings are not being written as measurement products. Etag builds a freshness model from stacked sample images of similar produce, then uses the estimated grade and sales curve to adjust price tags automatically, with a second neural model over historical freshness triggering replenishment or price-change orders. A fresh-food system corrects for freshness decay before it outputs pricing parameters and inventory instructions. A stewed-bird's-nest filing derives a freshness risk index from remaining shelf-life distribution and folds it into a replenishment trigger that trades stockout risk against depreciation. ## Areas - Supply Chain & Inventory · 1,599 documents - Consumer & Marketing Intelligence · 1,303 documents - Product Quality & Freshness · 1,000 documents - In-Store Retail Operations · 578 documents - Pricing & Revenue Optimization · 265 documents - Enterprise Data & Digital Transformation · 506 documents