Methods for data analytics based on RFID signals

By clustering RFID tags into home zones and applying machine learning, the system addresses challenges in retail inventory management, accurately tracking items and reducing manual verification, thus improving inventory accuracy and operational efficiency.

WO2026241520A1PCT designated stage Publication Date: 2026-11-19AUTOMATION INC(US)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/US2026/028187
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-15
Filing Date
2026-05-14
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Conventional RFID inventory systems face challenges in retail environments, including difficulty in determining whether items have actually left the store versus being misplaced, inaccurate positioning of tags near walls or barriers, accumulation of stale tag data, and labor-intensive manual verification processes.

Method used

The methods involve clustering RFID tags into 'home zones' based on expected locations, using machine learning to predict item exits, implementing intelligent filtering of position data, and accounting for physical retail space realities to improve inventory accuracy and operational efficiency.

Benefits of technology

The system accurately tracks inventory, identifies misplaced items, and detects when items have exited the store, reducing manual verification needs and enhancing inventory accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure US2026028187_19112026_PF_FP_ABST
    Figure US2026028187_19112026_PF_FP_ABST
Patent Text Reader

Abstract

A system may transmit, by an RFID tag reader, a series of interrogation signals to the RFID tags over a measurement period; receive, by the RFID tag reader, a series of replies to the series of interrogation signals from the RFID tags over the measurement period, the series of replies including EPCs of the RFID tags; estimate, based on the series of replies, respective median locations of the RFID tags over the measurement period; group the RFID tags into clusters based on the respective median locations and the EPCs, each cluster including at least a predetermined number of RFID tags for a corresponding type of object, and each cluster being separated from each other cluster for the corresponding type of object by at least a predetermined distance; and designate each cluster as a home zone for the corresponding type of object.
Need to check novelty before this filing date? Find Prior Art

Description

Attorney Docket No. RADR-026W001 METHODS FOR DATA ANALYTICS BASED ON RFID SIGNALS CROSS-REFERENCE TO RELATED APPLICATION(S)

[0001] This application claims the priority benefit, under 35 U.S.C. § 119(e), of U.S. Application No. 63 / 806,410, which was filed on May 15, 2025, and is incorporated herein by reference in its entirety for all purposes.BACKGROUND

[0002] Radio-frequency identification (RFID) tags are low-cost devices that can be attached to objects and offer the promise of automated tracking, locating, sales check-out, and inventory of the objects among other commercial and medical applications. An RFID tag stores information, typically in the form of a unique electronic product code (EPC) or other unique tag identification number or alpha-numeric sequence. When an RFID tag is queried by an RFID tag reader, also called a reader, interrogator, or sensor, the RFID tag can reply by transmitting its EPC. The EPC can be used to identify and track an object to which the RFID tag is attached.

[0003] There are passive, semi-active, and active types of RFID tags. Passive RFID tags have no battery and are therefore typically lower cost than semi -active and active RFID tags. Passive RFID tags are powered by wireless signals, such as the interrogation signal from a sensor. A passive RFID tag responds to the interrogation signal by modulating (e.g., with the passive RFID tag's EPC) and backscattering continuous-wave energy from the sensor. The sensor detects the modulated, backscattered energy as the passive RFID tag's reply. The sensor may also estimate the passive RFID tag's location from the angle of arrival (AO A) or received signal strength indication (RSSI) of the tag's reply.SUMMARY

[0004] Various embodiments described herein relate to methods and systems for tracking RFID tags in retail environments and using machine learning for inventory management. In some embodiments, the methods include determining home zones for products by clustering RFID tags, detecting misplaced items, predicting whether items have been removed from the retail environment, and maintaining accurate inventory records. The inventors have recognized these methods or systems can allow retailers to improve inventory accuracy, reduce labor costs associated with manual inventory checks, and improve the efficiency of product restocking.Attorney Docket No. RADR-026W001

[0005] In some aspects, the techniques described herein relate to a method of tracking radiofrequency identification (RFID) tags, each RFID tag storing a unique electronic product code (EPC) encoding attributes of an object to which that RFID tag is attached, the method including: transmitting, by an RFID tag reader, a series of interrogation signals to the RFID tags over a measurement period; receiving, by the RFID tag reader, a series of replies to the series of interrogation signals, the series of replies including the EPCs of the RFID tags; estimating, based on the series of replies, respective median locations of the RFID tags over the measurement period; grouping the RFID tags into clusters based on the respective median locations and the EPCs of the RFID tags, each cluster including at least a predetermined number of RFID tags for a corresponding type of object, and each cluster being separated from each other cluster for the corresponding type of object by at least a predetermined distance; and designating each cluster as a home zone for the corresponding type of object.

[0006] In some aspects, the techniques described herein relate to a method, wherein grouping the RFID tags into clusters further includes forming clusters for a given type of object on opposite sides of a wall.

[0007] In some aspects, the techniques described herein relate to a method, wherein the wall separates a front-of-house portion of a retail environment from a back-of-house portion of the retail environment, a first one of the clusters for the given type of object is in the front-of-house portion of the retail environment, and a second one of the clusters is in the back-of-house portion of the retail environment.

[0008] In some aspects, the techniques described herein relate to a method, further including, after the measurement period: transmitting, by the RFID tag reader, an interrogation signal to an RFID tag attached to an object; receiving, by the RFID tag reader, a reply to the interrogation signal from the RFID tag; estimating, based on the reply, a location of the RFID tag; determining, based on a distance between the location of the RFID tag and the home zone for the object, that the object is misplaced; and providing, to a user, an indication that the object is misplaced.

[0009] In some aspects, the techniques described herein relate to a method, wherein at least one type of object has more than one cluster designated as a home zone.

[0010] In some aspects, the techniques described herein relate to a method of tracking an object with a radio-frequency identification (RFID) tag attached thereto in a retail environment, the method including: transmitting, by an RFID tag reader, a series of queries to the RFID tag;Attorney Docket No. RADR-026W001receiving, by the RFID tag reader, at least one response from the RFID tag to the series of queries; extracting, with a machine learning model from the at least one response, features associated with movement and position of the RFID tag, the features including: (i) a time since a last changed position of the RFID tag, (ii) a distance from the RFID tag to a home zone for the RFID tag, (iii) a number of times the RFID tag was within a first predetermined distance of a checkout of the retail environment, (iv) a distance traveled by the RFID tag within a first predetermined period since a last known position change of the RFID tag, and (v) a number of movements of more than a second predetermined distance within a second predetermined period before the last known position change of the RFID tag; and determining, with the machine learning model based upon at least one of the features, a stock status of the object, the stock status indicating whether the object is available on a sales floor, in a stockroom, or out of stock.

[0011] In some aspects, the techniques described herein relate to a method, wherein the features further include a time of day of the last known position change of the RFID tag.

[0012] In some aspects, the techniques described herein relate to a method, wherein the features further include a last known position of the RFID tag.

[0013] In some aspects, the techniques described herein relate to a method, wherein the object is an article of clothing and the features further include a period since the last known position change of the RFID tag and a last visit of the RFID tag to a fitting room.

[0014] In some aspects, the techniques described herein relate to a method, wherein the first predetermined distance is one meter.

[0015] In some aspects, the techniques described herein relate to a method, wherein the second predetermined distance is one meter.

[0016] In some aspects, the techniques described herein relate to a method, wherein the first predetermined period is five minutes.

[0017] In some aspects, the techniques described herein relate to a method, wherein the second predetermined period is five minutes.

[0018] In some aspects, the techniques described herein relate to a method, wherein the features further include a period between the last known movement of the RFID tag and a most recent event associated with the object.Attorney Docket No. RADR-026W001

[0019] In some aspects, the techniques described herein relate to a method, wherein the most recent event includes a sale of the object.

[0020] In some aspects, the techniques described herein relate to a method, wherein the features further include a number of visits by the RFID tag to the checkout within twenty-four hours of the last known position change of the RFID tag, wherein a visit includes detection of the RFID tag within the first predetermined distance of the checkout and separated from other detections by at least ninety minutes.

[0021] In some aspects, the techniques described herein relate to a method, wherein the features further include a distance traveled by the RFID tag within a third predetermined period since the last known position change of the RFID tag.

[0022] In some aspects, the techniques described herein relate to a method, wherein the third predetermined period is ten minutes.

[0023] In some aspects, the techniques described herein relate to a method, wherein the features further include whether the RFID tag was last detected within a predetermined distance of an e-commerce packing area with greater than a threshold probability.

[0024] In some aspects, the techniques described herein relate to a method, wherein the threshold probability is 0.5.

[0025] In some aspects, the techniques described herein relate to a method, wherein: determining the stock status includes determining a probability that the RFID tag has been removed from the retail environment since the RFID tag reader received the at least one response from the RFID tag.

[0026] In some aspects, the techniques described herein relate to a method, further including: notifying a user to replace the object attached to the RFID tag in response to the probability exceeding a threshold probability.

[0027] In some aspects, the techniques described herein relate to a method, further including: extracting, by the RFID tag reader, information from the at least one response; transmitting the information from the RFID tag reader to an interrogator controller; and executing the machine learning model on the interrogator controller.

[0028] In some aspects, the techniques described herein relate to a method, further including: receiving, by the interrogator controller, an indication of an event associated with the RFID tag, wherein determining a probability is based on a time and / or type of the event.Attorney Docket No. RADR-026W001

[0029] In some aspects, the techniques described herein relate to a method, wherein the event includes a sale of the object, a theft of the object, and / or a loss of the object.

[0030] In some aspects, the techniques described herein relate to a method, further including: determining whether to restock the retail environment based on the stock status.

[0031] In some aspects, the techniques described herein relate to a method of training a machine learning model for detecting exits of items tagged with respective radio-frequency identification (RFID) tags from a retail environment, the method including: identifying clusters of retail environments having similar characteristics; collecting ground truth data from at least two retail environments from each cluster, including performing morning and evening hand scans to confirm tag presence by verifying whether each tag is physically present without determining precise tag coordinates of the RFID tags in the at least two retail environments across multiple days; collecting RFID tag data from the RFID tags with fixed RFID tag readers between the hand scans; extracting features from the RFID tag data collected between the hand scans; generating exit labels based on the ground truth data; and training the machine learning model using the extracted features and exit labels.

[0032] In some aspects, the techniques described herein relate to a method, wherein the similar characteristics include at least one of: store size; store layout; customer traffic pattern; product assortment; sensor positioning; and sensor layout.

[0033] In some aspects, the techniques described herein relate to a method of maintaining RFID tag state accuracy in a retail environment, including: detecting, with an RFID tag reader, signals from RFID tags in the retail environment; extracting features from the signals including at least location data representing estimated positions of the RFID tags and timing data representing timestamps of tag detections; determining, using a machine learning model, exit probabilities for the RFID tags based on the extracted features; setting states of RFID tags with exit probabilities above a threshold to lost; identifying RFID tags with exit probabilities below the threshold; and generating an audit list including the identified RFID tags.

[0034] In some aspects, the techniques described herein relate to a method of processing RFID tag data in a retail environment by an interrogator, including: receiving RFID tag read data including tag identifiers, signal characteristics, and read timestamps from multiple RFID tag readers; filtering the RFID tag read data to exclude RFID tag position changes of less than one meter; and using the filtered tag read data to generate features for training a machine learning model for predicting a state of tags identified by the filtered tag read data.Attorney Docket No. RADR-026W001

[0035] In some aspects, the techniques described herein relate to a method of cleaning up stale RFID tag data, including: identifying RFID tags without recorded movements for longer than a threshold period; determining whether identified RFID tags were last detected in a front-of-house location or a back-of-house location; applying an exit probability threshold to the RFID tags based on last detected locations of the RFID tags, wherein a different exit probability threshold applied is for the front-of-house location than the back-of-house location; setting states of tags with exit probabilities above the exit probability threshold to lost; and monitoring rates of state changes from lost back to available, wherein a state change from lost to available is triggered when a previously-lost tag is detected again.

[0036] In some aspects, the techniques described herein relate to a method of validating RFID tag exits, including: collecting point-of-sale (POS) events for items in a retail environment; detecting signals from the RFID tags attached to the items; comparing timing between last detected movements of the RFID tags and POS events, wherein the movements are inferred from changes in position data derived from RFID tag read data; determining exit probabilities of the RFID tags based at least in part on the timing comparisons; and updating inventory data based on the determined exit probabilities.

[0037] In some aspects, the techniques described herein relate to a method of determining exit status of radio-frequency identification (RFID) tags in a retail environment, the method including: identifying, by a processor, a neighborhood of RFID tags based on spatial proximity of most recently detected positions of the RFID tags, the neighborhood including a target tag and a plurality of neighboring tags; determining, by the processor, a corresponding time-since-last-read for the target tag and for one or more neighboring tags of the plurality of neighboring tags; determining, by the processor, whether a number of tags in the neighborhood have not been read within a threshold period; in response to determining that the number of tags in the neighborhood have not been read within the threshold period, classifying the neighborhood as a blind spot; and in response to determining that at least one neighboring tag has been read within the threshold period, comparing the time-since-last-read of the target tag to the respective time-since-last-read values of the one or more neighboring tags and determining an exit probability for the target tag based at least in part on the comparison.

[0038] In some aspects, the techniques described herein relate to a method, wherein identifying the neighborhood of RFID tags includes identifying, for the target tag, whichever RFID tags have most recently detected positions closest to a most recently detected position of the target tag, regardless of absolute distance between the target tag and the neighboring tags.Attorney Docket No. RADR-026W001

[0039] In some aspects, the techniques described herein relate to a method, further including: determining a trajectory of the target tag based on a sequence of detected positions of the target tag prior to a last read of the target tag; and adjusting the exit probability for the target tag based on whether the trajectory leads toward a region previously classified as a blind spot.

[0040] In some aspects, the techniques described herein relate to a method, further including: applying a first trained machine learning model to determine the exit probability when the target tag is located in a front-of-house portion of the retail environment; and applying a second trained machine learning model, different from the first trained machine learning model, to determine the exit probability when the target tag is located in a back-of-house portion of the retail environment.

[0041] In some aspects, the techniques described herein relate to a method, wherein the first trained machine learning model incorporates a cashwrap proximity weighting such that tags detected near a cashwrap receive a higher exit probability weighting.

[0042] In some aspects, the techniques described herein relate to a method, further including: maintaining a history of locations in the retail environment where detection errors have occurred; and in response to determining that the target tag is proximate to a location in the history, reducing the exit probability for the target tag.

[0043] In some aspects, the techniques described herein relate to a method of improving product placement in a retail environment using radio-frequency identification (RFID) tag data, the method including: collecting, by a processor, co-occurrence data indicating which products tagged with respective RFID tags were detected together within a fitting room zone of the retail environment during respective fitting room sessions over a specified period; generating, by the processor, product embeddings for a plurality of products based on the co-occurrence data, wherein products with higher co-occurrence frequencies receive higher similarity scores in an embedding space; computing, by the processor, pairwise similarity scores between the product embeddings; constructing, by the processor, a hierarchical grouping structure based on the pairwise similarity scores, the hierarchical grouping structure defining consideration sets of products at a selected similarity threshold; and generating, by the processor, a floor plan for the retail environment by assigning products within each consideration set to proximate fixture locations and assigning consideration sets with higher inter-set similarity to adjacent regions.

[0044] In some aspects, the techniques described herein relate to a method, wherein generating product embeddings includes at least one of: passing product identifiers through a projectionAttorney Docket No. RADR-026W001layer and training with a contrastive loss function using the co-occurrence data as training labels; encoding product descriptions and attributes through a language model encoder and training with a contrastive loss function using the co-occurrence data as training labels; and encoding product images through a convolutional neural network encoder and training with a contrastive loss function using the co-occurrence data as training labels.

[0045] In some aspects, the techniques described herein relate to a method, further including: identifying traffic hot spots within the retail environment based on RFID tag movement data indicating locations of higher customer traffic density; and assigning products classified as high trial frequency and high conversion rate to fixture locations at the identified traffic hot spots.

[0046] In some aspects, the techniques described herein relate to a method, further including: computing, for each product pair, a co-trial-without-co-purchase probability indicating a probability that both products are brought to the fitting room zone during a same session and at most one is purchased; and classifying product pairs with the co-trial-without-co-purchase probability above a substitution threshold as substitutes.

[0047] In some aspects, the techniques described herein relate to a method, further including: determining that a product classified as having few substitutes based on the hierarchical grouping structure is out of stock or near an out-of-stock status; and generating a replenishment alert for the product.

[0048] In some aspects, the techniques described herein relate to a method, further including: generating the floor plan for a test group of stores while maintaining a current floor plan for a control group of stores; monitoring RFID tag data in the test group to verify adherence to the improved floor plan; and comparing key performance indicators between the test group and the control group over a measurement period.

[0049] All combinations of the foregoing concepts and additional concepts discussed in greater detail below (provided such concepts are not mutually inconsistent) are part of the inventive subject matter disclosed herein. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are part of the inventive subject matter disclosed herein. The terminology used herein that also may appear in any disclosure incorporated by reference should be accorded a meaning most consistent with the particular concepts disclosed herein.Attorney Docket No. RADR-026W001 BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The skilled artisan will understand that the drawings primarily are for illustrative purposes and are not intended to limit the scope of the inventive subject matter described herein. The drawings are not necessarily to scale; in some instances, various aspects of the inventive subject matter disclosed herein may be shown exaggerated or enlarged in the drawings to facilitate an understanding of different features. In the drawings, like reference characters generally refer to like features (e.g., functionally similar and / or structurally similar elements).

[0051] FIG. 1 depicts an RFID environment with several ceiling -mounted RFID tag readers, also called readers or sensors, for interrogating and locating passive RFID tags.

[0052] FIG. 2A shows a bottom perspective view of a reader with a four-element, square antenna array.

[0053] FIG. 2B depicts a perspective cross-sectional view of the reader of FIG. 2A.

[0054] FIG. 3 illustrates a flowchart of an example method of clustering (home zone determination / creation based on object type).

[0055] FIG. 4 illustrates a flowchart of an example method of machine learning exit detection.

[0056] FIG. 5 illustrates a flowchart of an example method of training a machine learning model for detecting exits of items.

[0057] FIG. 6 illustrates a flowchart of an example method of maintaining RFID tag state accuracy in a retail environment.

[0058] FIG. 7 illustrates a flowchart of an example method of preparing filtered data for training a machine learning model for predicting tag state.

[0059] FIG. 8 illustrates a flowchart of an example method of cleaning up stale RFID tag data.

[0060] FIG. 9 illustrates a flowchart of an example method of validating RFID tag exits.

[0061] FIG. 10 illustrates example tables of experimental results relating to lost availability for two different locations.

[0062] FIG. 11 illustrates an example chart of lost availability for two locations.

[0063] FIG. 12A illustrates example tables of key performance indicators for two locations.

[0064] FIG. 12B illustrates an example table detailing information relating to FIG. 12 A.Attorney Docket No. RADR-026W001

[0065] FIG. 13 A illustrates an example table where a single SKU was misplaced an entire day.

[0066] FIG. 13B illustrates an example map indicating the location of such a single SKU.

[0067] FIG. 13C illustrates an example table including one SKU that is misplaced throughout a part of the day.

[0068] FIG. 14 illustrates example maps illustrating trade-offs regarding a distance threshold.

[0069] FIG. 15A illustrates an example table indicating two SKUs misplaced for part of a day.

[0070] FIG. 15B illustrates an example map indicating locations of the two SKUs of FIG. 15 A.

[0071] FIG. 15C illustrates an example table providing time-based information about the two SKUs of FIG. 15 A.

[0072] FIG. 16A illustrates an example table indicating one SKU was in a "stockroom" for part of a day.

[0073] FIG. 16B illustrates an example map indicating locations throughout a day relating to FIG. 16 A.

[0074] FIG. 16C illustrates an example table providing time-based information relating to FIG.16 A.

[0075] FIG. 16D illustrates an example table where one SKU is in the stockroom.

[0076] FIG. 16E illustrates an example map indicating the scenario of FIG. 16D.

[0077] FIG. 17A illustrates an example table relating to a shared wall.

[0078] FIG. 17B illustrates an example map relating to the shared wall of FIG. 17A.

[0079] FIG. 18 illustrates a chart of a feature engineering pipeline according to some embodiments.

[0080] FIG. 19 illustrates an example plot of the time between a tag's last movement and its closest event.

[0081] FIG. 20A illustrates an example fitting room analytics dashboard displaying key performance indicators by department, category, and product, with a product ensembles table.

[0082] FIG. 20B illustrates an example fitting room analytics KPI trends time-series chart showing fitting room session counts plotted over a date range.Attorney Docket No. RADR-026W001

[0083] FIG. 21 illustrates an example trial and conversion segment analysis table and quadrant diagram.

[0084] FIG. 22 illustrates example product embedding generation architectures using fitting room co-occurrence data and a three-step customer decision tree construction process.

[0085] FIG. 23 illustrates an example customer decision tree with a hierarchical dendrogram and adjustable similarity threshold for grouping products into consideration sets.

[0086] FIG. 24 illustrates an example floor plan improvement process and store layout comparison showing current versus improved product placement.

[0087] FIG. 25 illustrates a flowchart of an example method of product substitution analysis for lost sales minimization.

[0088] FIG. 26 illustrates a flowchart of an example method of neighborhood-based blind spot detection for determining RFID tag exit probability.

[0089] FIG. 27A illustrates an example availability analytics dashboard showing sales floor availability metrics, a KPIs-by-day-and-hour heat map, and a widespread out-of-stocks table.

[0090] FIG. 27B illustrates an example availability analytics KPI trends time-series chart showing sales floor availability percentage plotted over a date range.DETAILED DESCRIPTION

[0091] Conventional RFID inventory systems face several challenges in retail environments. These include difficulty in determining whether items have actually left the store versus being misplaced, inaccurate positioning of tags near walls or barriers, accumulation of stale tag data, and labor-intensive manual verification processes. The inventors recognized and appreciated these challenges and developed methods and systems that analyze data from RFID tags on items (and can leverage machine learning) to improve inventory accuracy and operational efficiency.

[0092] The systems and methods described herein solve these problems through several techniques, such as but not limited to the following: (1) creating "home zones" by clustering products based on their expected locations; (2) utilizing machine learning to predict whether items have exited the store; (3) implementing intelligent filtering of position data to reduce data volume while preserving relevant information; and (4) accounting for the physical realities of retail spaces, such as shared walls between front-of-house and back-of-house areas.Attorney Docket No. RADR-026W001

[0093] An RFID tag reader can track the movement of an RFID tag throughout a store, retail establishment, or other environment by interrogating the RFID tag at different times and estimating the RFID tag's locations at those times based on the RFID tag's replies. An interrogator controller or other processor coupled to the RFID tag reader processes this location data to determine product positioning, detect misplaced items, predict exits, and maintain accurate inventory.

[0094] The methods and systems described herein enable retailers to accurately track inventory, identify misplaced items, detect when items have exited the store, and maintain inventory accuracy without constant manual verification. By clustering tags into home zones and applying machine learning to movement and position data, the system can distinguish between items that have actually left the store and those that are simply misplaced or in low-read areas.

[0095] The methods and systems described herein leverage the tag's recently measured locations and movement patterns to predict or estimate whether an item to which the RFID tag is attached has been sold and used to trigger restocking. By analyzing features such as time since last movement, distance traveled, interaction with key areas like cashwraps (or checkout counters), and correlation with POS events, the system can improve accuracy in determining inventory status.System Overview

[0096] FIG. 1 depicts a multipath RFID environment 100 (viewed from above) with a dense population of passive RFID tags 101 and one or more RFID tag readers 150, also called RFID readers, tag readers, interrogators, or sensors, which communicate with the RFID tags 101. The RFID tags 101 can be attached to objects, such as articles of clothing, which can be identified and tracked using the RFID tags 101 and readers 150. The RFID environment 100 can be in a retail store or warehouse, for example, though other settings are possible. There can be furnishings 120 in the environment that affect RF signals (e.g., block, attenuate, and / or scatter the RF signals). The furnishings 120 can include shelves, racks, cabinets, etc. that may be used to hold the objects to which at least some of the RFID tags are attached. For example, some furnishings 120 may comprise metal shelving that holds one or more retail items (not shown in FIG. 1) that are tagged with the RFID tags 101. The furnishings 120 can be arranged in rows in some settings, with aisles separating the rows to allow access to objects tagged with the RFID tags 101.Attorney Docket No. RADR-026W001

[0097] Some stores may have external stockrooms located outside the main store premises, without sensors, electricity or internet connectivity. The system may account for inventory movement to and from these external stockrooms when tracking item status and location.

[0098] The RFID environment 100 can be bounded by the ceiling, floor, and walls 110, 112, any of which can reflect or scatter RF signals from the readers 150 and / or RFID tags 101. There may be one or more cameras 130 installed in the RFID environment to capture images of at least portions of the environment. The readers 150 and cameras 130 can be communicatively coupled to an interrogator controller 140, which can receive and process data from the readers 150 and the cameras 130.

[0099] In some embodiments, the readers 150 are installed in the RFID environment 100 such that together they can communicate with every RFID tag 101 in the RFID environment 100. The readers 150 may communicate via one or more Ethernet cables (not shown) with the interrogator controller 140, which may also supply electrical power from the interrogator controller 140 to the readers 150. The interrogator controller 140 may be a computer, laptop, smart phone, or purpose-built device with a processor and non-volatile computer-readable memory adapted to communicate with the RFID readers 150 and issue commands recognizable to the RFID readers 150. The interrogator controller 140 can also receive signals from the RFID readers 150.

[0100] The RFID environment may also include the cashwrap 170, which may include a POS system 172, which may be communicatively connected to the interrogator controller 140. For example, the connection may be a wired or wireless connection, and the connection can be local or an Internet / cloud-based connection system (e.g., over cloud 160).RFID Readers / Sensors

[0101] RFID tag readers 150 can be used in retail stores, warehouses, supply rooms, libraries, museums, galleries, or other environments for tracking objects with RFID tags 101. Generally, the readers 150 are mounted or hung from the ceiling, e.g., at a height in a range from 8 feet to 20 feet above the floor, so that they emit interrogation signals downward, toward RFID tags 101 on shelves, tables, clothing racks, or other storage units. The RFID readers 150 are arranged so that they provide adequate coverage of the RFID environment 100, e.g., on a grid with a spacing chosen so that all or substantially all of the volume of the RFID environment 100 can be interrogated by the RFID tag readers 150. The RFID tag readers 150 can be arrayedAttorney Docket No. RADR-026W001in a regular (periodic) fashion as shown in FIG. 1 or distributed irregularly across the RFID environment.

[0102] FIGS. 2A and 2B depict an example of an RFID reader 150 that includes a phased antenna array having four antenna elements 205. The illustration of FIG. 2 A is a perspective view from the bottom of the reader 150. The reader 150 also includes a ground plane 220 inside an enclosure or housing 280 that extends across and behind the antenna elements 205 with respect to the emission direction of an RF beam from the reader.

[0103] The enclosure 280 is roughly disc shaped and can be made of acrylonitrile butadiene styrene (ABS), polyvinyl chloride (PVC), vinyl, or another suitable material that is substantially transparent at the carrier frequencies of the interrogation signals and replies. The enclosure 280 may include thinner sections 282 adjacent to the antenna elements 205. These thinner sections 282 can be 1 / 8" thick or less so that they transmit in-band RF energy to and from the antenna elements 205 without significant attenuation of the transmitted RF energy. The enclosure 280 may also have visible fiducial markings or may have a shape (e.g., an asymmetric shape) that facilitates determining the installed reader's pose or orientation as explained below.Locating RFID Tags with RFID Tag Readers / Sensors

[0104] Referring again to FIG. 1, the interrogator controller 140 and the readers 150 can locate the RFID tags 101 based on the tags' replies to signals from the readers 150. For example, the interrogator controller 140 can issue a command to inventory the RFID tags 101 (and attached items) in the RFID environment 100 or to locate one or more RFID tags 101 (and attached item(s)) in the RFID environment. In response to these commands, one or more of the readers 150 transmit signals to the RFID tags 101 and receive the tags' replies. The RFID readers 150 transmit the tags' replies or indications of the tags' replies (e.g., received signal strength indicators (RSSIs), AO As, channel estimates, EPCs, and / or other information derived from or encoded in the replies, etc.) to the interrogator controller 140. The interrogator controller 140 uses this information to identify and locate the RFID tags 101 that replied to the readers 150.

[0105] In some cases, one RFID reader 150, called an interrogator, transmits an interrogation signal to a tag 101 and the interrogator and other RFID readers 150 within range, called listeners, receive the tag's reply. Because the listeners are not transmitting, they do not experience self-interference and so can detect responses at lower power levels (longer ranges).Attorney Docket No. RADR-026W001And because the interrogators and listeners are at different locations, they measure different angles of arrival (AO As) and channel estimates for each response. The interrogator controller 140 can use these simultaneous measurements to locate each tag faster and with higher fidelity. For more on listeners, please see U.S. Pre-Grant Publication No. 2024 / 0193381 Al, entitled "RFID Tag Readers Switchable Between Interrogator and Listener Modes," which is incorporated herein by reference in its entirety for all purposes.

[0106] The readers 150 and / or interrogator controller 140 locate the tags 101 in an environmental coordinate system or reference frame fixed with respect to the environment 100. In FIG. 1, the environmental coordinate system is a Cartesian coordinate system with the origin in one corner and x and y axes (lower left) extending parallel to the floor and ceiling. The z axis (not shown) extends perpendicular to the x and y axes, i.e., into and out of the plane of FIG. 1. Other coordinate systems are also possible, including Cartesian coordinate systems centered at other points within the environment or cylindrical or spherical coordinate systems with appropriately selected origins. The interrogator controller 140 provides information to sales associates, managers, customers, and / or other users about the estimated location(s) of the tag(s) 101 in the environmental coordinate system or with reference to points registered to the environmental coordinate system (e.g., entrances, exits, sales counters, etc.).

[0107] The accuracy of the estimated locations depends in part on how accurately the sensors' locations and orientations are known in the environmental coordinate system. This is because the readers 150 and / or interrogator controller 140 use location estimation methods based on measurements of the tags' locations relative to the readers 150. For instance, the interrogator controller 140 can triangulate a given tag's location from the AO As of replies from that tag 101 as measured by two or more different readers 150. The interrogator controller 140 can also estimate a tag's location from the intersection of a ray along the AOA measured by a single reader 150 and a plane that contains or is assumed to contain the tag 101, e.g., at a height of 1 meter above the ground. The interrogator controller 140 can estimate the tag's location using multilateration from estimates of distance (e.g., as represented by RSSIs) from the tag to several different readers 150 or from an estimate of distance between the tag 101 and one reader 150 and an assumption about the tag's height. Signature matching involves correlating a channel estimate derived from a reply detected by a given reader 150 to previously recorded channel estimates between that reader 150 and known locations in the environment 100. For more on signature matching, see, e.g., U.S. Pre-Grant Publication No. 2026 / 0016590 Al, andAttorney Docket No. RADR-026W001entitled "Channel Estimation for Locating RFID Tags," which is incorporated herein by reference in its entirety for all purposes.

[0108] If the RFID readers' locations and orientations in the coordinate system are known, then the interrogator controller 140 or reader(s) 150 can translate tag location estimates relative to the readers' locations into absolute tag location estimates — that is, tag location estimates given in coordinates of the environmental coordinate system. Translating the tag location estimates into the environmental coordinate system enables the interrogator controller 140 to supply the coordinates of a tag's estimated location to an app on a smartphone, tablet, or other device that shows the tag's estimated location on a map of the environment 10, or relative to a wall, doorway, fixture, or other readily identifiable reference point in the environment 100. Providing absolute coordinates or coordinates with respect to a fixed reference point instead of a reader 150 makes it easier to corroborate or average tag location estimates derived from measurements by different readers 150, and to locate tags 101 and any corresponding objects.Example Techniques for Locating RFID Tags

[0109] As discussed above, the RFID readers are capable of estimating an RFID tag's position. Methods for doing so include but are not limited to AOA and signatures (channel estimates).Angle of Arrival

[0110] AOA is a process of using the azimuth and elevation of a tag's reply in relation to a sensor to locate the tag with respect to that sensor. A pattern file may be used at a specific resolution. In some embodiments, the system uses 5-degree resolution. Some embodiments use diagonal loading as the method to map channel estimates to azimuth and elevation. High elevation can create extreme outliers; to remove these outliers, the system may limit the search area of the pattern file to 120 degrees. Once the relative AOA from the sensor to tag is known, the system can form position estimates using Single Sensor fixed - Z or Multi-Sensor Triangulation processes.Single Sensor fixed - Z

[0111] If we know the relative elevation and azimuth of the tag to a sensor, we still do not know the x, y, z position and we do not know the distance, r, from the sensor to the tag. However, if we set the z coordinate to 1 m (a good average position of where tags could be expected to be on a salesfloor), then we can set a plane at z=l and calculate the x, y position from where the given elevation and azimuth intersect that plane.Attorney Docket No. RADR-026W001 Multi-Sensor Position

[0112] When reads from two different sensors are available and the azimuth and elevation are obtained from each read, the system performs triangulation to find the intersection of these two beams. For this method to be effective, two consecutive reads should occur over a short time period. Some embodiments may use listener reads, but the inventors have recognized that much greater error may occur in listener reads than in interrogator reads. Signatures / Channel Estimates

[0113] When the ground truth location for a set of EPCs is known along with their mean normalized channel estimates for a given sensor, sector, and frequency combination (these all affect the channel estimate), the system stores those signatures as a mapping for each grouping (sensor, sector, frequency), with the normalized channel estimate as the key and x,y,z position as the value. When a tag is read, the channel estimate for the given sensor, sector, and frequency of that read is compared against all stored signature channel estimates for that grouping, and the closest channel estimate is selected with its corresponding position. For more on channel estimates, please see U.S. Pre-Grant Publication No. 2026 / 0016590 Al, entitled "Channel Estimation for Locating RFID Tags," which is incorporated herein by reference in its entirety for all purposes.Example ArchitectureRadar Store

[0114] In some embodiments, a Radar store, besides the RFID tagged retail items, hosts a single central entity called an appliance (e.g., a hardware device that hosts the interrogator controller as a software component) and multiple sensors to provide the desired (e.g., complete) RF coverage. In some embodiments, some or all the sensors and the appliance are connected by an Ethernet link to a gigabit switch.

[0115] The appliance hosts multiple software components, including one or more of:• Interrogator Controller, which is the coordinator of the sensors, and may be hosted as a software component by the appliance or may be its own hardware, as discussed elsewhere herein;• WebApp gateway, which services the cloud (e.g., cloud 160 in FIG. 1) and mobile apps providing accurate location of all the tags in the store;• Multiple submodules like polyhedron service, tag state, and retail backend lite, which aid in locating the RFID items; andAttorney Docket No. RADR-026W001• A router or logical mechanism to communicate between the sensors and the controller and also internally between different modules of the appliance.Interrogator Controller[00 J .16] The interrogator controller is the master coordinator of all sensors and runs a scheduler module, which does this co-ordination.

[0117] In some embodiments, every given interval (e.g., between 300 ms and 500 ms, such as 400 ms), the interrogator controller sends commands to each sensor. The tag population may determine physical parameters (frequency, sector, antenna) to operate on.

[0118] At the end of the interrogation / reply sequence, or hop, the interrogator controller receives messages regarding the hop from each sensor.

[0119] In some embodiments, the sensors send the FPGA messages to the interrogator controller, and the IC extracts the tag information from the messages. Alternatively, the sensors can process the FPGA messages and send a consolidated message for each tag, called "tag measurements."

[0120] The IC does additional computation to determine the location of the tag in space from the channel estimates.SchedulerExhaustive Scheduler

[0121] In some embodiments, the scheduler works on an event list, which is used to send commands to the sensors and receive completion messages. The event list may be regenerated during every sensor add / delete operation.

[0122] A generator may loop through different combinations of sensors, frequencies, sectors, and physical and antenna configurations. In some embodiments, at any given moment, only one sensor is picked as the interrogator, and the other sensors within range of the tags being interrogated act as listeners.Table Scheduler

[0123] In some embodiments, the scheduler also works on an event list, but the list of events is generated by a higher level entity. This can get inputs from either the exhaustive scheduler or from the configuration file. For example, the table scheduler has an option to prune the list of events based availability of current sensors and / or valid configurations.Attorney Docket No. RADR-026W001 Multi-mode Scheduler

[0124] The multi-mode scheduler introduces the concept of time during the scheduling operation. For example, the multi-mode scheduler may provide a start and end in local time for the schedule within which the list of events will have to be scheduled. The inventors recognized this is useful because different options for scheduling may apply during peak hours of a store where rapid transitions in tags need to be covered, as compared to overnight scheduling when the schedule can cover all possible configurations to get better reads. The multi-mode scheduler may also have a feature that allows overlapping schedule configurations. In this case, the multimode scheduler may switch in round-robin fashion across the individual schedulers.Dynamic Scheduler

[0125] In some embodiments, the dynamic scheduler takes in a Graphics Library Transmission Format (GLTF) model and the sensor configuration as inputs. From the GLTF model, the dynamic scheduler figures out the special sensors, including the entrance, exit, cashier, stockroom, and changing room.

[0126] Sets of co-located sensors are created such that the sets can be independently scheduled. An example is to have the exit sensor in Set 1, rest of floor sensors in Set 2, and stock room sensors in Set 3.

[0127] A separate set of floor sensors is created so that it does not have any interference to the exit sensors. This will be used during the time the exit sensor is enabled.

[0128] The number of simultaneous interrogators is a physical parameter which can be set as configuration or can be hard coded.

[0129] Based on the priority of each sensor, a certain combination of sets is used during each schedule.Location / Position EstimatorsAngle of Arrival Based Position Estimation

[0130] The AO A based position estimation can be more accurate than alternative systems. In some embodiments, this estimator is run at the end of every hop after tag measurements across the hop are collected.Attorney Docket No. RADR-026W001

[0131] Each tag measurement may be sent through the AO A computation logic, which uses a co-efficient bank for each frequency and antenna configuration and computes an elevation and azimuth angle.

[0132] The Per Tag Data will hold AOA alone with the sensor and time, which is associated with this angle.

[0133] This data is then sent to the position estimator that will compute the x, y, z position. There are 2 different estimators for the position (Single sensor and Multi -sensor). The single sensor uses an approach of approximating the Z value to be a given distance (e.g., Im) and then computing the position. If there are multiple reads from a single sensor, only the last position (the most recent one) is being used.

[0134] In some embodiments, this module outputs the Tag EPC, position, and an error metric to a router, which then feeds to the other modules.Averaged Interrogator Estimation

[0135] This is a simplified scheme that is a coarse estimator but does not take too much CPU time. The processing is very similar to the AOA scheme. However, this scheme uses the interrogator sensor position as the location of the tag. This scheme uses a moving average across all the tag measurements of a hop if there are multiple measurements.Best Sensor Position Estimation

[0136] During Retailer A store 1 deployment, the inventors recognized position "jumpiness," most of which was attributed to two different sensors picking up the position estimates at slightly different locations. To mitigate this, some embodiments include a best sensor position estimation. For more details on best sensor position estimation, see PCT Application No. WO2024196835A1, titled "Best Sensor / Measurement Selection for Locating RFID Tags" by Webb et al., which is incorporated herein by reference in its entirety for all purposes.Polyhedron Service

[0137] In some embodiments, the polyhedron service extracts the volumes and rooms from a file, (e.g., a GLTF file) and extracts the convex hull of each shape. The perimeter of the store is computed by the union of all the convex hulls. The tag information is streamed to this service from the IC and the polyhedron service binds the tag position to the closest perimeter of the space.Attorney Docket No. RADR-026W001

[0138] In some embodiments, the service does not maintain any database, but rather continuously streams the inputs to the outputs.Tag State Machine

[0139] The tag state machine's main functionality is to annotate a tag state (e.g., Available, Sold, Stale) to the tag reads.

[0140] It may store the tag state on disk by maintaining a per EPC tag information.

[0141] As the tags are streamed through the tag state machine:• If the current tag state was sold, and the tag is found at the entrance, the tag continues to be marked as sold.• If the tag is found anywhere else, the tag is marked as available.

[0142] Periodically, all the tags in the storage are read, and the following steps are taken:• If a tag was last seen at the entrance volume and has not been read for 15 minutes, the tag is marked as sold. In some embodiments, the timeout is 2 minutes, which the inventors recognized can help entrance / exit key performance indicators (KPIs).• If a tag was not seen for 24 hours, but was not at the exit, the tag is marked as stale. • A change in tag state is then sent to the downstream module.

[0143] In some embodiments, the storage is never cleared. In other embodiments, the tag states are cleaned up to avoid consuming too much disk space.

[0144] In some embodiments, a log buffer is used to store the tag state transitions for 50 iterations. Disk space is monitored due to this logging.Retail Backend Lite

[0145] As tag reads are streamed, the retail backend updates copies in its own storage. In additional stages, ignored and missing tags are used in addition to the tagging from upstream modules.Inventory Reporting

[0146] Periodically (e.g., at around 7 am every day), the inventory list of all available items is created. The report can be made to meet a specific start / stop or decide on number of items in the query.Attorney Docket No. RADR-026W001 Space Server

[0147] The space server runs a HTTP server that can be used for dual purposes: an internal API which the sensors use to post HTTP requests to the controller; and an external API that is usable in the local network.SensorStore-Controller Client

[0148] The store controller client handles the primary configuration / status coordination between the sensors and the appliance.During Initialization

[0149] The sensor checks if an IP address is obtained and pings the appliance. Then the sensor checks if the NTP is synced up. If not, the sensor may be set for a reboot.Thread Loop on Interval (e.g., every 60 seconds)

[0150] The sensor obtains the configuration using a HTTPS GET request and configures both the RFID and, if present, the optional camera subsystems.

[0151] The sensor checks its hardware status (uptime, status, temperature, firmware version) and sends status information to the appliance.

[0152] There is a separate thread that the daemon spawn and uploads all the syslog (e.g., once the syslog reaches IK bytes) from the sensor to the appliance.Example FlowchartsClustering in Home Zones

[0153] FIG. 3 illustrates an example method of clustering (home zone determination / creation based on object type), which may be implemented via a system as described herein (e.g., the system shown in FIG. 1). For example, the system may include one or more processors hosted by the appliance, such as CPUs (central processing units), GPUs (graphics processing units), NPUs (neural processing units), custom ASICs, or any other suitable architecture. The system may include any suitable number of RFID tags, each of which may store a unique EPC encoding attributes of an object to which that RFID tag is attached. In some embodiments, the method may begin at step 310.Attorney Docket No. RADR-026W001

[0154] At step 310, the system may transmit interrogation signals to RFID tags over a measurement period. For example, as described herein, the system may transmit the signals via an RFID tag reader. The method may then proceed to step 320.

[0155] At step 320, the system may receive a series of replies including EPCs from RFID tags. For example, as described herein, the system may, via the RFID tag reader, receive the series of replies (which may include the EPCs of the replying RFID tags) to the series of interrogation signals over the measurement period. These replies may include information that allows the system to determine the location and identity of each tag (e.g., based on signal strength, angle of arrival, and / or channel estimates), as discussed herein. The method may then proceed to step 330.

[0156] At step 330, the system may estimate respective median locations of the RFID tags. For example, as described herein, the system may estimate, based on the series of replies, the respective median locations of the RFID tags over the measurement period. This approach may be empirical, such as by relying on actual measured positions rather than predetermined expectations. In some embodiments, the system may estimate these median locations by processing multiple position readings for each tag over time, which builds a data-driven understanding of tag locations that helps filter out temporary positioning errors and provides more stable location data for clustering. The measurement period may be during periods of minimal tag movement, such as overnight, to improve accuracy (e.g., of the empirical baseline). The method may then proceed to step 340.

[0157] In some embodiments, the clustering process may be dynamic, responding automatically to changes in the retail environment as staff rearrange merchandise. When products are moved to new display locations during floor resets or merchandising updates, subsequent measurements will reflect these changes, causing the home zones to naturally migrate to the new locations during the next clustering cycle. This dynamic adaptation ensures that home zones continuously reflect the current store layout without requiring manual reconfiguration.

[0158] At step 340, the system may group the RFID tags into clusters based on the respective median locations and the EPCs of the RFID tags. For example, as described herein, each cluster may include at least a predetermined number of RFID tags for a corresponding type of object, and each cluster may be separated from each other cluster for the corresponding type of object by at least some predetermined distance.Attorney Docket No. RADR-026W001

[0159] In some embodiments, more than one cluster may be designated as the home zone for a given type of object. For example, a type of jeans may have two or three or even more home zones where that type of jeans is expected to be in an environment.

[0160] For example, the system may identify that a particular style of jeans has three different clusters in a retail environment: one in the men's department, one in the women's department, and one in a featured display area. These may be designated as separate home zones if they are more than the predetermined distance apart and each contains at least the predetermined number of RFID tags.

[0161] In some embodiments, step 340 may optionally include step 342, where the system may form clusters for a given type of object on opposite sides of a wall. For example, in some embodiments the wall may separate a "front-of-house" (FOH) portion of a retail environment from a "back-of-house" (BOH) portion of the retail environment. In some embodiments, a first of the clusters for the given type of object is in the FOH portion, and a second of the clusters is in the BOH portion.

[0162] The method may then proceed to step 350.

[0163] At step 350, the system may designate each cluster as a home zone for the corresponding type of object. For example, the system may store the geographic coordinates of each cluster center along with the product information, such as style and color, associated with that cluster. This information is then used to determine the expected location of products throughout the retail environment. In some embodiments, the system may also calculate and store the radius or boundaries of each home zone based on the distribution of tags within the cluster.

[0164] The method may then end or repeat, as appropriate.

[0165] Alternatively or additionally, the method may optionally proceed from step 330 to step 360.

[0166] At step 360, the system may determine whether object(s) are misplaced. For example, the system may transmit an interrogation signal from the RFID tag reader to an RFID tag attached to an object. The system then may receive a reply to the interrogation signal from the RFID tag via the RFID tag reader and estimate the location of the RFID tag based on the reply. This may allow the system to then determine, based on the distance between the location of the RFID tag and the home zone for the object, whether the object is misplaced.Attorney Docket No. RADR-026W001

[0167] The method may then optionally proceed to step 370.

[0168] At step 370, the system may indicate to the user whether the object is misplaced. For example, the system may provide an indication to the user that the object is misplaced. In some embodiments, the system may also provide an indication to the user if the object is not misplaced. The system may provide this indication through any suitable interface, such as SMS message, an email, a notification via desktop or mobile application, etc., which may show the current location of the object relative to its home zone on a store map. The system may also prioritize misplaced items based on factors such as how long they have been misplaced, their distance from the home zone, and their sales velocity.

[0169] In some embodiments, the system may generate reports showing misplaced item trends over time, allowing retailers to identify problem areas in the store where items are frequently misplaced, as described herein. Because the system continuously may gather empirical measurements and updates clusters based on real-world tag positions, these reports can provide an evidence-based view of store operations that evolves as staff rearranges merchandise, seasonal displays change, or store layouts are modified. The empirical and dynamic nature of these clusters allows the system to maintain accuracy without relying on static, predefined zones that would quickly become obsolete in the constantly changing retail environment.

[0170] The method may then end or repeat, as appropriate.

[0171] The clustering process described above can be seen in practice in the examples illustrated in FIG. 17A and FIG. 17B, which demonstrate how the system handles the "shared wall problem." As shown in FIG. 17B, the system has identified a home zone in the FOH area (left side) while also recognizing product clusters in the BOH area (right side). This demonstrates the system's capability to form clusters on opposite sides of a wall as described in step 342.

[0172] The real -world benefits of this clustering approach are further evidenced in FIGS. 13A through 13C, which show how the system can detect when items are misplaced by measuring their distance from their designated home zones. For example, FIG. 13B maps the location of a misplaced item that was consistently detected far from its home zone throughout the day, illustrating an outcome of the process described in steps 360-370.Attorney Docket No. RADR-026W001 Machine Learning Exit Detection (MLED)

[0173] FIG. 4 illustrates an example method of machine learning exit detection or MLED, which may be implemented via a system as described herein (e.g., the system shown in FIG. 1). In some embodiments, the method may begin at step 410.

[0174] At step 410, the system may transmit queries to an RFID tag. For example, as described herein, the system may transmit a series of queries to the RFID tag via an RFID tag reader. The method may then proceed to step 420.

[0175] At step 420, the system may receive response(s) from the RFID tag. For example, the system may receive at least one response from the RFID tag via the RFID tag reader. A response may include the RFID tag's EPC and other information that can be used to determine the tag's location, such as signal strength and phase angle data, which the system may process to build a history of the tag's movements and positions over time, as discussed below. The method may then proceed to step 430.

[0176] At step 430, the system may extract movement / position features from the response(s). For example, as described herein, the system may extract, with a machine learning model from the at least one response, features associated with movement and position of the RFID tag. For example, the features may include any of the following: a time since a last changed position of the RFID tag, a distance from the RFID tag to a home zone for the RFID tag, a number of times the RFID tag was within a first predetermined distance of a checkout of the retail environment, a distance traveled by the RFID tag within a first predetermined period since a last known position change of the RFID tag, and a number of movements of more than a second predetermined distance within a second predetermined period before the last known position change of the RFID tag.

[0177] In some embodiments, the first predetermined distance may be one meter, and / or the second predetermined distance may be one meter. In some embodiments, the first predetermined period may be five minutes, and / or the second predetermined period may be five minutes (e.g., the distance traveled in five minutes before the last position, such as distance_traveled_5_min_last_timestamp discussed herein).

[0178] Some embodiments relate to a specific timing between events and a last read. For example, the features may include a period between the last known movement of the RFIDAttorney Docket No. RADR-026W001tag and a most recent event associated with the object. The most recent event could be, for example, a sale of the object, such as in a retail environment.

[0179] In some embodiments, the features may include a time of day of the last known position change of the RFID tag (e.g., the hour of last move, such as hour last move discussed herein). Alternatively or additionally, the features may include a last known position of the RFID tag.

[0180] Some embodiments relate to defining checkout visits more precisely (including 24-hour and 90-minute parameters). For example, the features may include a number of visits to checkout within twenty-four hours of the last known position change of the RFID tag. A visit can be detection of the RFID tag within the first predetermined distance of the checkout and separated from other detections by at least ninety minutes.

[0181] Some embodiments relate to a 10-minute travel distance. For example, the features may include a distance traveled by the RFID tag within a third predetermined period since the last known position change of the RFID tag. This third predetermined period may be ten minutes.

[0182] Some embodiments relate to e-commerce packing location detection. For example, the features may include whether the RFID tag was last detected within a predetermined distance of an e-commerce packing area with greater than a threshold probability. The threshold probability may be 0.5.

[0183] In some embodiments, the object may be an article of clothing, and the features may include a period since the last known position change of the RFID tag (e.g., mins_since_last_position as discussed herein) and a last visit of the RFID tag to a fitting room.

[0184] The method may then proceed to step 440.

[0185] At step 440, the system may determine stock status based on certain features. For example, the system may determine, with the machine learning model based upon at least one of the features, a stock status. For example, the system may classify the RFID tag as "available" (in its expected location), "misplaced" (in the store but not in its expected location), "in stockroom" (not on the sales floor), or "not in store" (likely sold or otherwise removed from the retail environment), as discussed below.

[0186] In some embodiments, step 440 may optionally include step 442, where the system may determine the probability that an RFID tag was removed from the environmentAttorney Docket No. RADR-026W001(e.g., "not in store" mentioned above). For example, in some embodiments, determining the stock status may include determining the probability that the RFID tag has been removed from the retail environment, after the RFID tag reader received the at least one response from the RFID tag.

[0187] In some embodiments, step 440 may also optionally include step 444, which follows step 442. At step 444, the system may compare the probability determined in step 442 with a threshold probability. For example, the threshold probability may be 95%, which provides a balance between accuracy and coverage. If the probability is not above the threshold probability, the method may return to step 420 or any other suitable step, or may end or repeat, as appropriate.

[0188] If the probability is above the threshold probability, the method may optionally proceed to step 446, where the system may notify the user to replace the object attached to the RFID tag in response to the probability exceeding the threshold probability.

[0189] Alternatively or additionally to other steps within step 440, step 440 may optionally be followed by step 448, where the system may determine whether to restock the retail environment. For example, determining the stock status may include determining whether to restock the retail environment, such as based on the number of available items of a particular SKU, the sales velocity of that SKU, and predefined minimum inventory thresholds.

[0190] Alternatively or additionally, the method may optionally proceed from step 430 to step 450.

[0191] At step 450, the system may transmit extracted information to an interrogator controller. For example, the system may, via the RFID tag reader, extract information from the at least one response, similar to step 430, and then transmit the information from the RFID tag reader to an interrogator controller. In some embodiments, the information transmitted to the interrogator controller includes tag identifiers, position data, movement history, and other metrics that are used by the machine learning model to determine exit probability, as described herein.

[0192] The method may then optionally proceed to step 460.

[0193] At step 460, the system may execute a machine learning model on the interrogator controller. This may relate to the minutes between an event and last read (e.g., such as mins btwn event last read discussed herein). For example, the system may receive, via the interrogator controller, an indication of an event associated with the RFID tag.Attorney Docket No. RADR-026W001Determining the probability may be based on a time and / or type of that event. The system may analyze the temporal relationship between tag movements and events, with tags that have events (particularly sales) occurring close to their last movement having a higher probability of having exited the store.

[0194] Some embodiments may relate to minutes between a sale event and the last read (e.g., such as mins btwn sale event last read discussed herein). For example, the event may include a sale of the object, a theft of the object, and / or a loss of the object.

[0195] The method may then optionally proceed to step 470.

[0196] At step 470, the system may receive indication of a sale / theft / loss of the object. For example, the system may receive POS data indicating that the item has been purchased, or data from security systems indicating potential theft, or inventory audit data indicating a loss. In some embodiments, the system may correlate this information with the tag's last known movement and position to further refine the exit probability calculation.

[0197] The method may then end or repeat, as appropriate.

[0198] A practical application of this MLED approach is demonstrated in FIG. 19, which shows a plot of the time between a tag's last movement and its closest event in time (i.e., most recent event). The chart visually confirms the importance of features such as mins btwn event last read and mins btwn sale event last read from the feature extraction process described in step 430. As shown in the chart and accompanying table in paragraph 421, approximately 81.2% of exited tags had their last movement occur within 60 minutes of a related event.

[0199] This model's effectiveness is further detailed in the tables accompanying FIG.12A and FIG. 12B, which show key performance indicators for two different retail locations, including metrics on items that were likely removed from the store according to the machine learning model's predictions. The inventory accuracy improvements resulting from the MLED process demonstrate that the model can achieve significant improvements in inventory tracking.Training Machine Learning Model via Exit Labels

[0200] FIG. 5 illustrates an example method of training a machine learning model for detecting exits of items from a retail environment, such as by using exit labels, which may beAttorney Docket No. RADR-026W001implemented via a system as described herein (e.g., the system shown in FIG. 1). In some embodiments, the method may begin at step 510.

[0201] At step 510, the system may identify clusters of retail environments having similar characteristics. For example, as described herein, similar characteristics may include any of the following: store size, store layout, customer traffic pattern, product assortment, sensor positioning, and sensor layout. The method may then proceed to step 520.

[0202] At step 520, the system may collect ground truth data from two or more retail environments from each cluster of retail environments. For example, as described herein, the system may retrieve data collected by associates performing morning and evening hand scans of the RFID tags in these retail environments across multiple days. The method may then proceed to step 530.

[0203] Hand scans are a manual inventory process where store staff (or optionally, automated robotic devices) use handheld / portable RFID readers to scan each tagged item in the retail environment. These scans provide definitive verification of items physically present at a specific time, serving as ground truth data for the system.

[0204] Hand scans are typically performed in pairs (morning and evening), creating a controlled time window for data collection. By comparing these scans, the system can more definitively determine which items have exited during the day. Items present in the morning scan but absent in the evening scan may be considered confirmed to have exited, providing reliable training data for an MLED model.

[0205] For new store implementations, multiple rounds of hand scans are conducted over several weeks. For model training, hand scans are performed in at least two stores from each identified store cluster, capturing inventory across multiple days. This paired scanning approach helps establish certainty about inventory status and movement, enabling the system to correlate RFID readings with actual inventory presence.

[0206] At step 530, the system may collect RFID tag data from RFID tags with fixed RFID sensors between the hand scans. For example, as described herein, the system may continuously record tag positions, movements, and other metrics during normal store operations between the morning and evening hand scans, building a comprehensive dataset of tag behavior. The method may then proceed to step 540.Attorney Docket No. RADR-026W001

[0207] At step 540, the system may extract features from the RFID tag data collected between the hand scans. For example, as described herein, the system may calculate metrics such as time since last movement, distance from home zone, cashwrap visits, distances traveled within specific time periods, and correlations with POS events for each tag.

[0208] The method may then proceed to step 550.

[0209] At step 550, the system may generate exit labels by comparing the ground truth data. For example, the system may label tags that were present in the first hand scan but absent in the second hand scan as having "exited" the store, creating a binary classification dataset where exited tags are assigned a value of 1 and tags still present are assigned a value of 0.

[0210] The method may then proceed to step 560.

[0211] At step 560, the system may train the machine learning model using the extracted features and exit labels. For example, the system may use the labeled dataset to train a machine learning model, such as an XGBoost model, that can predict the probability of a tag having exited the store based on the extracted features, improving metrics such as accuracy, precision, and recall.

[0212] The method may then end or repeat, as appropriate.

[0213] The training approach detailed above has been validated through experiments at multiple retail locations. As illustrated in FIG. 10, these experiments tested various parameters to determine improved thresholds for identifying misplaced or exited items. The tables show results from Location 1 and Location 2, demonstrating how varying time and distance thresholds affect model accuracy.Maintain Tag State Accuracy via Exit Probabilities

[0214] FIG. 6 illustrates an example method of maintaining RFID tag state accuracy in a retail environment via exit probability determined via machine learning, which may be implemented via a system as described herein (e.g., the system shown in FIG. 1). In some embodiments, the method may begin at step 610.

[0215] At step 610, the system may, using an RFID reader, detect signals from RFID tags in the retail environment. For example, as described herein, the system may continuously query tags and receive their signals throughout the retail environment during normalAttorney Docket No. RADR-026W001operations, collecting signal data that can be processed to determine tag locations and movements. The method may then proceed to step 620.

[0216] At step 620, the system may extract features from the signals, including at least location data and timing data. For example, as described herein, the system may calculate metrics such as time since last movement, distance from home zone, number of cashwrap visits, distances traveled in specific time periods, and correlations with point-of-sale events for each tag. Examples may include, more specifically: minutes between the last known movement for an EPC and an event for the SKU; time between last tag movement and last event for the SKU, minutes between the last scan and the last known movement >1 meter, and so on. The method may then proceed to step 630.

[0217] At step 630, the system may determine exit probabilities for RFID tags based on extracted features. For example, the system may use a machine learning model to determine these exit probabilities. The model may process the extracted features to generate a probability score between 0 and 1 for each tag, representing the likelihood that the tag has been removed from the retail environment. The method may then proceed to step 640.

[0218] At step 640, the system may set to lost the states of RFID tags with exit probabilities above a threshold. For example, the threshold may be 95%, which provides an improved balance between accuracy and coverage based on validation testing.

[0219] The method may then proceed to step 650.

[0220] At step 650, the system may identify RFID tags with exit probabilities below the threshold. For example, the system may compile a list of tags with moderate exit probabilities (e.g., between 0.5 and 0.95) that require further verification before being set to lost state.

[0221] The method may then proceed to step 660.

[0222] At step 660, the system may generate an audit list including the identified RFID tags. For example, the system may create a prioritized list of tags for manual verification, ranking items based on their exit probability, value, and sales velocity to streamline the audit process and maintain inventory accuracy.

[0223] The method may then end or repeat, as appropriate.

[0224] This methodology has been implemented and tested in real retail environments as shown in FIG. 12 A, which provides a breakdown of tag states across two different retailAttorney Docket No. RADR-026W001locations. The "Not in store" category represents tags that the system has determined have likely exited based on the exit probabilities as is calculated in step 630.

[0225] FIG. 11 provides a visual representation of how the distance threshold affects lost availability determinations, illustrating the sensitivity analysis that informs the threshold selection described in step 640. The chart shows a clear inflection point at 3 meters, although a 4-meter threshold was ultimately selected for implementation to account for the 95th percentile of error in location accuracy, as discussed in below.Preparing Filtered Data for Training Machine Learning Model

[0226] FIG. 7 illustrates an example method of preparing filtered data for training a machine learning model for predicting tag state, which may be implemented via a system as described herein (e.g., the system shown in FIG. 1). For example, these steps may be performed by the interrogator controller or any other suitable component of the system. In some embodiments, the method may begin at step 710.

[0227] At step 710, the system may receive RFID tag read data from multiple RFID tag readers. For example, as described herein, the system may collect tag read data from all ceiling-mounted sensors (e.g., RS510 sensors or any other suitable sensors) throughout the retail environment, including position information, signal strength, channel estimates, and tag identifiers. The method may then proceed to step 720.

[0228] At step 720, the system may filter the RFID tag read data to exclude RFID tag position changes below a threshold. For example, the threshold may be one meter. The method may then proceed to step 730.

[0229] At step 730, the system may send the filtered tag read data to a cloud service. For example, as described herein, the system may transmit only the significant tag movements to the cloud, substantially reducing the data volume while preserving information used for inventory analytics. This subsample approach ensures efficient use of network bandwidth and cloud storage resources. The method may then proceed to step 740.

[0230] At step 740, the system may use the filtered tag read data to generate features for training a machine learning model for predicting tag state (e.g., predicting a state of tags identified by the filtered tag read data).Attorney Docket No. RADR-026W001

[0231] For example, the system may generate features such as time since last movement, distance from home zone, visits to key areas like cashwrap, distance traveled in specified time periods, and correlations with POS events.

[0232] The method may then end or repeat, as appropriate.

[0233] The data filtering approach described above is helpful for the feature engineering pipeline illustrated in FIG. 18, which shows how raw tag data flows through the system from data collection to model training. By filtering position changes of less than one meter as described in step 720, the system significantly reduces data volume while preserving movement information.

[0234] This filtered data enables the creation of features described herein, such as distance_traveled_5_min_last_timestamp and distance traveled l O min last timestamp, which measure the distance a tag traveled in the minutes before its last position change. The inventors have recognized these features as advantageous to the model's performance.Cleaning Up Stale Tag Data

[0235] FIG. 8 illustrates an example method of cleaning up stale RFID tag data, which may be implemented via a system as described herein (e.g., the system shown in FIG.1). In some embodiments, the method may begin at step 810.

[0236] The inventors have appreciated various technical benefits of cleaning up stale tag data, including improved inventory accuracy by removing tags that have likely exited the store from inventory counts, reducing false inventory reports that could trigger unnecessary replenishment, and improving system performance by reducing data volume. The process helps differentiate between items that have actually exited the store versus those that are simply misplaced or in low-visibility areas, addressing a challenge in RFID inventory systems.

[0237] The filtered data can serve multiple purposes throughout the system, including training machine learning models for exit prediction, establishing home zones through clustering analysis, determining which side of shared walls tags are located on, and supporting inventory analytics. By filtering out minor position changes (e.g., less than one meter), the system may dramatically reduce data volume while preserving information that may be used for analytics and machine learning models. This filtered data enables feature extraction for the Machine Learning Exit Detection (MLED) model, which predicts tag states with high accuracy and maintains inventory integrity with minimal manual verification.Attorney Docket No. RADR-026W001

[0238] At step 810, the system may identify RFID tags that have been without recorded movement for longer than a threshold period. For example, the threshold may be 24 hours or another configurable time period based on store operations and customer traffic patterns. The method may then proceed to step 820.

[0239] At step 820, the system may determine whether identified RFID tags were last detected in a FOH or BOH location. For example, as described herein, the system may analyze the last known positions of stale tags and classify them based on the volume probabilities associated with those positions, distinguishing between sales floor areas and stockrooms. The method may then proceed to step 830.

[0240] At step 830, the system may apply an exit probability threshold to RFID tags based on the last detected RFID tag locations. For example, as described herein, the system may use different probability thresholds for tags last seen in different locations, with a higher threshold for tags last seen in areas where false positives would be more problematic, such as stockrooms or other low-read areas. The system may also apply different probability thresholds based on product value, sales velocity, or historical accuracy metrics for particular SKUs or product categories. For example, high-value items or fast-selling products might receive more conservative (higher) thresholds to reduce inventory errors that could impact sales, while slower-moving items might use lower thresholds. Additionally, the system may adjust thresholds dynamically based on store traffic patterns, time of day, or seasonal factors that affect tag readability. This approach to threshold configuration balances inventory accuracy with minimal false removals, while accounting for the business impact of potential errors. The method may then proceed to step 840.

[0241] At step 840, the system may set the states of RFID tags with exit probabilities above the threshold to "lost." For example, as described herein, the system may update the tag state in the database from "available" to "lost," which removes these items from available inventory counts and may trigger replenishment processes.

[0242] The method may then proceed to step 850.

[0243] At step 850, the system may monitor rates of state changes from lost back to available. For example, the system may track how often tags previously set to lost are subsequently detected again in the store, which serves as a key performance indicator for the accuracy of the exit prediction model and allows for continuous improvement of the threshold and model parameters.Attorney Docket No. RADR-026W001

[0244] The method may then end or repeat, as appropriate.

[0245] The process of cleaning up stale tag data can be observed in the examples presented in FIG. 16A through FIG. 16E, which show SKUs that were determined to be in the stockroom (BOH) rather than on the sales floor. In particular, FIG. 16D and FIG. 16E illustrate a case where all SKUs of a product are available except for a single size that remained in the stockroom throughout the entire day.

[0246] The thresholds applied in step 830 are informed by the experimental results shown in FIG. 10, which demonstrate how different time and distance parameters affect the identification of misplaced or stale items. As noted herein with reference to FIGS. 12A and 12B, the lost availability from items in the stockroom improved throughout the day in Location 2, showing the dynamic nature of the stale tag cleanup process as implemented through this method.Validating Tag Exit via Point-of-Sale (POS) Data

[0247] FIG. 9 illustrates an example method of validating RFID tag exits, which may be implemented via a system as described herein (e.g., the system shown in FIG. 1). In some embodiments, the method may begin at step 910.

[0248] At step 910, the system may collect POS data for items in a retail environment. For example, as described herein, the system may integrate with the retail POS system to receive transaction data including SKU information, timestamps, and transaction types (sales, returns, exchanges, etc.). The method may then proceed to step 920.

[0249] At step 920, the system may detect signals from RFID tags attached to the items. For example, as described herein, the system may, continuously query RFID tags and detect their response signals throughout the retail environment, and process the responses to track tag locations and movements over time. The method may then proceed to step 930.

[0250] At step 930, the system may compare timing between the last detected tag movements and POS events. For example, as described herein, the system may analyze the temporal relationship between when a tag was last observed to move and when POS events occurred for the corresponding SKU, identifying patterns that correlate with items exiting the store. This analysis may build a profile of how tag movements relate to actual sales events. The method may then proceed to step 940.Attorney Docket No. RADR-026W001

[0251] At step 940, the system may determine exit probabilities of the RFID tags based on timing comparisons. For example, as described herein, the system may calculate higher exit probabilities for tags whose last movements occurred close in time to POS events for their corresponding SKUs, especially sales events.

[0252] For example, the system may assign significantly higher exit probabilities to tags whose last movements were within 60 minutes of a sale event for their SKU, as historical data shows approximately 81% of exited tags fall into this category.

[0253] The method may then proceed to step 950.

[0254] At step 950, the system may update inventory data based on the determined exit probabilities. For example, the system may set tags with exit probabilities above a predetermined threshold to "lost" state, effectively removing them from available inventory counts and potentially triggering replenishment processes for the corresponding SKUs.

[0255] The method may then end or repeat, as appropriate.

[0256] The benefits of POS data integration in exit validation are demonstrated in FIG.19, which show the relationship between tag movements and POS events. The chart reveals that approximately 81.2% of exited tags had POS events occur within 60 minutes of their last movement, confirming the value of the timing comparison described in step 930.

[0257] The real -world impact of this validation approach is evident in FIG. 15A through FIG. 15C, which track the movement of two items over time. The timeline data illustrates how the system can correlate movement patterns with potential sales events to determine whether items have truly exited the store, as in step 940.Measuring Location Confidence

[0258] The inventors have recognized two immediate priorities for measuring uncertainty in state and location:1. EPC Removal Selection & Correction: Rank ordering EPCs based on their likelihood of having left the store.2. EPC Audit Shortlist: Providing ranked lists of items to audit, to confirm if they remain in the store and their locations.

[0259] This can provide a measure of confidence in location to end users.Attorney Docket No. RADR-026W001

[0260] In some embodiments, data on tag location history is not utilized, but instead the process can rely on existing data tag measurements or new one-time tag measurements captured over 1-2 hours.

[0261] The inventors have recognized that obtaining "ground truth" data is useful for understanding which items were in the store and when they left for testing and development. This may be based on prior or new hand scans.• In some embodiments, when implementing the system in a new store location:o Initial store visits are conducted to collect baseline data.o Multiple rounds of hand scans are performed over several weeks. o Machine learning models are trained based on data from approximately five stores to establish baseline performance.o Comparison focuses primarily on determining whether tags are present in or absent from the store.o System is then activated for normal operation.

[0262] The data captures and "ground truth" hand scans may be synchronized so that it is possible to link tag revisit rates, locations, and movements from the data captures to the times the items left the store (see discussion below on approach).

[0263] Some embodiments can run on the controller. For example, the granular data on tag reads may not be sent to the cloud service, which can make it impossible to calculate tag revisit rates on the cloud service.Approach

[0264] The inventors recognized the benefit of developing an approach that assumes a correlation among "tag and location staleness," location uncertainty, and exit probability.Generating Tag Staleness Feature

[0265] Since tag read rates can change by location in a store, the "tag staleness" metrics must take this into account. Additionally, stores can have very different average tag read rates and some embodiments consider this to ensure that it generalizes.

[0266] Outlined below are some different approaches:Attorney Docket No. RADR-026W0011. Comparisons across EPCs in a Given Style

[0267] This approach calculates the staleness of a given EPC relative to all EPCs in the same style. This assumes that all EPCs of the same style type should be in a similar location in the store and therefore have similar read rates. The table below illustrates how this method may rank order EPCs for a 1 hour capture:EPC Style Median Revisit Rate Time Since Last Read / Median Style # (seconds) Last Read Revisit Rate4945758949 38484 150 300 23485748390 38484 150 225 1.521283838384 38484 150 200 1.339

[0268] In this example, the median revisit rate from the style is used as the "expected" time between visits. By taking the ratio of the time since last read of the EPC to the expected time, based on all EPCs of the same style, it can be understood how much greater the actual elapsed time between reads is over an expectation as measure of staleness.2. Comparisons across EPCs in a Given Volume

[0269] This is as above, but the comparisons are over all EPCs that are assumed to be in the same volume based on the calculated volume probabilities from the location model. EPC EPC Volume (based Median Volume Time EPC Last Read / on volume Revisit Rate Since Median Volume probability) (seconds) Last Revisit Rate Read1283838384 Mens 1 120 180 1.593485748390 Mens 1 120 150 1.324945758949 Mens 1 120 75 0.63. Comparisons with EPCs Found in a Given Radius

[0270] In this approach, all EPCs are compared to all other EPCs within a defined radius (e.g., 5 m). This approach may directly account for the physical proximity of tags,Attorney Docket No. RADR-026W001regardless of predefined volumetric boundaries. Unlike volume-based comparisons which rely on organizational segmentation of the store, radius-based comparisons may capture the true spatial relationship between tags, creating neighborhoods based solely on physical distance. The inventors recognized this can be valuable for accurately assessing tags located near volume boundaries, where a tag might be physically closer to tags in an adjacent volume than to distant tags within its own assigned volume. By establishing comparison neighborhoods based on actual proximity measurements, this approach may more precisely reflect the local environmental conditions affecting read rates, such as physical barriers, sensor coverage patterns, and customer traffic flow that may not align perfectly with predefined volume boundaries. This approach may require calculating distances between tag positions, establishing a fixed radius threshold that appropriately captures local read rate variations while excluding unrelated distant tags.EPC Median EPC Revisit Rate Time Since EPC Last Read / Median Tags in Radius Last Read Radius Revisit Rate 1283838384 220 250 1.193485748390 220 220 124945758949 220 180 1.24. Comparisons with EPCs Found in a Derived Local Neighborhood

[0271] This approach is very similar to the above, except instead of using a specified radius, data driven clusters are derived utilizing a clustering process that groups EPCs together on their locations.EPC Median EPC Revisit Rate Time Since EPC Last Read / Median Tags in Neighborhood Last Read Neighborhood Revisit Rate 1283838384 180 250 1.493485748390 180 220 1.224945758949 180 180 1Attorney Docket No. RADR-026W001Pros and Cons of ApproachesApproach Pros ConsComparisons Easy to run on the controller. May be less robust than some across EPCs approaches, as the comparison in a given Simplest and fastest method (at least may be against a smaller number style within time constraints). of combined SKUs.If an EPC is not with its "peer styles," then the comparison may not be truly accurate, as it should be based upon how often a tag is read given its location relative to sensors and barriers.The approach may consider at least 2 locations (front of house and back of house).Comparisons Easy to run on the controller. May be less accurate than some across EPCs approaches, as averages are over in a given May be more robust than some more EPCs, and volumes can be volume approaches, as noise at the style level is quite large.averaged out.Comparisons are more based onlocation in the store and how easy tagsare to read in that location, rather thanagainst all EPCs in the style, whichassumes that they are all in the samelocation.Comparisons The average read rates are more Memory and time intensive to with EPCs accurately based on their location in the calculate the distance between all found in a store (i.e., an EPC out of home location EPC pairs.given radius is being compared to nearby SKUs, notto others of its style that may not beclose by).Comparisons Once the neighborhood centers are May be more complex than other with EPCs calculated in a store, EPCs can be approaches, as it may involve a found in a assigned a neighborhood by calculating process to be run per-store to derived the distance to the neighborhood center determine neighborhood centers neighborhoo not the distance between every EPC in order to generalize.d pair.May be regularly retrained if fixtures are moved around or clusters are moved relative to fixtures, changing neighborhoodcenters.Attorney Docket No. RADR-026W001

[0272] The inventors have considered how the approaches can be used with EPC correction and audit probability:

[0273] EPC Correction: If an EPC correction is appropriate, the EPC selected would be the one that is "most stale" relative to its SKU peers based on the method chosen.

[0274] Audit Priority: EPCs selected for audit would be those with the highest relative ratios to their peers.ValidationData Collection

[0275] The inventors have recognized that the process herein may be validated based upon some ground truth data collection. This would involve two sets of hand scans with a data capture in between: This set up allows identifying items that left the store in between the two hand scans and to analyze and model how well utilizing tag staleness in the proposed manner for EPC selection / correction / audit priority would work.

[0276] The data allows developing a simple confusion matrix and calculating precision / recall metrics, to get a sense of how much confidence the approach can provide.

[0277] In some embodiments, this data collection may run across multiple stores, making sure to collect data from "similar" stores and that there is wide coverage of store types.Existing Data

[0278] The inventors appreciated that existing data may also be used, especially when based on a similar set up.Store Simulations

[0279] The simulation process to test inventory processes can be deployed to help validate the processes herein.Explicitly Modelling "Exit Probability"

[0280] In some embodiments, a model (or models), including machine learning models (e.g., decision trees, gradient boosted models) as well as statistical or heuristic models, can be used to predict whether an EPC has left the store by generating more ground truth labels and building some machine learning (ML) models. Given current constraints of running predictionsAttorney Docket No. RADR-026W001on the controller, some embodiments use ML to generate a more data driven and multivariate heuristic.Data Collection

[0281] As outlined herein, more ground truth labels can be used to confirm that items have left the store. To ensure generalizability of the final heuristics, some embodiments collect and train models on a variety of store types. Some initial analysis would be conducted to determine clusters of stores with similarities on the attributes that matter (e.g., store layout, system performance, etc.).Target Generation

[0282] In order to model the probability that an item has exited the store, we can generate an "exit flag" that takes the value of 1 where the EPC was found in the first hand scan but not in the second hand scan.Feature Generation

[0283] Some embodiments include the following features that are likely related to location confidence and exit probability from our data capture:1. Revisit rate during the last 15 minutes, 30 minutes, and 1 hour2. Time since last read (limited to the length of the data capture, such as 1 hour) could take the value "not read"3. Number of locations in the last 15 minutes, 30 minutes, and 1 hour (new "locations" are defined as a movement > 2 meters)4. Item read in checkout zone in last hour?5. Item read in fitting room in last hour?6. Volume item last read (assuming there may be some pattern with "items on the way to the exit"7. Boolean identifying if the item was last read in its home location8. SKU(s) of the same type as the EPC found in POS feed

[0284] These features are deliberately general and not tied to specific locations (i.e., specific x, y, z co-ordinates) in a store or to specific SKU characteristics, as some embodiments generalize across stores.

[0285] The inventors have recognized that these features may use data that is currently only held on the controller (revisit rates). Based on these (current) limitations, processes may be run on the controller, as the inventors have appreciated that revisit rates may be an informative and predictive feature in the predictions (e.g., revisit rates may help predict whether a tag has actually exited the store).Attorney Docket No. RADR-026W001Generate a Heuristic to Rank the EPCs

[0286] The inventors have recognized that this process may be run on the controller, so the model can include rules in order to rank items. To do this, some embodiments use some simple models predicting exit probability based upon either (or a combination of):1. A decision tree - the branches of the tree lead to some rules that provide an "exit probability"2. A regression model - the coefficients lead to some rules to provide an "exit probability"Determining EPC Action

[0287] The output of the models is a probability that the item left the store. This probability may help determine which EPCs to remove from inventory as part of the EPC correction process and which EPCs to prioritize for audit and / or restocking.

[0288] EPC removal / correction can be done (solely) on the rank order of the EPC.

[0289] The number of EPCs that may be flagged for audit could be based on either (or a combination of) the following criteria: A target for store labor effort; Analysis of confusion matrices at different cut-offs (i.e., at X probability, Y will likely be found and Z will not).

[0290] Exit probability is discussed in more detail elsewhere herein.Modeling Tag States

[0291] The inventors have recognized the benefits of detecting EPCs that should be removed from inventory, as well as developing models that predict all tag states. This provides greater visibility into the confidence a tag is in a given state, given everything determined about it, and would significantly enhance understanding of "stale tags."

[0292] The inventors have recognized and appreciated certain pros and cons of running such a process on the controller vs. in the cloud:Attorney Docket No. RADR-026W001Pros ConsControlle 1. Data already exists on the 1. Controller cannot store long r controller to calculate histories of tag reads and locations revisit rates - it is planned that it will "at least 7 2. Costs would be days - up to 30"minimized, as compute is 2. The models are simpler and may on the device not compute complex feature 3. Predictions could be pipelinesserved with lower latencyor in real timeCloud 1. Allows for significantly 1. Raw tag reads might not be sent to more complex models as the cloud, so summary metrics may pipelines could be be defined and built on the constructed with many controller and sent to cloud instead features and longer history 2. Cost of compute of locations 3. Additional latency in serving predictions as data needs to be passed from controller to cloud andbackComparison of Approaches

[0293] Like described above, to develop an approach for modeling "staleness" to support EPC correction and location confidence use cases, the inventors have appreciated four different approaches to enable an understanding of tag staleness based on the time since a tag was last read and the "expected time between reads" from generated neighborhoods of "nearby tags." Neighborhoods were generated based on the "style and whether it was in the stock room or on the sales floor," "all other tags in the volume," "all tags in a 5 meter radius," and "all tags in a generated cluster. "

[0294] While significant differences in tag revisit rates were observed across the neighborhoods, high correlations were found between the ratios of neighborhood revisit rates and last tag reads and very high correlations with final SKU ranks.

[0295] Very strong correlations were also found between the "neighborhood normalized" ranks and a simple SKU ranking of time since last tag read.

[0296] Some results are based on three data captures (when one was at night): Retailer A Location 1 - 4-hour night capture; Retailer C 2-hour day capture; and Retailer B 1 hour day capture. The inventors have recognized the value of more data captures, hand scans to provide "ground truth," store simulations, and store "unit testing" (live simulations in a store of comer cases).Attorney Docket No. RADR-026W001Context

[0297] The inventors have recognized the outcomes of generating different location confidence metrics. For each of the methods, the approach is included below. There then follows a comparison of the methods to explain some similarities and differences between them.Overview of Approaches

[0298] Like described above, the inventors have recognized the value of ranking SKUs based on their "staleness." Methods of ranking SKU staleness include:1. Style & Volume: " Staleness" is based on the time since a tag was last read relative to the "expected" time between reads for all other tags in the same style and volume where "volume" is a simple split between sales floor (FOH) and stockroom (BOH) (e.g., median tag revisit rate for style number 1234 on sales floor)2. Volume: "Staleness" is based on the time since a tag was last read relative to the "expected" time between reads for all other tags in the same volume3. Fixed Radius: "Staleness" is based on the time since a tag was last read relative to the "expected" time between reads for all other tags within "X" meters ("X" = 5 meters for this analysis)4. Cluster: "Staleness" is based on the time since a tag was last read relative to the "expected" time between reads for all other tags within its location cluster. Location clusters are defined using 2 methods: 1. a k-means process, 2. a mean shift process Approach 1 - Style & Volume

[0299] This approach considers a tag "neighborhood" to be all tags that share the same style and volume where volume has been simplified to be simply "stockroom" or "sales floor" (e.g., style 1102 - sales floor). This approach may provide a distribution of median revisit rates across all style and volume combinations.Approach 2 - Volume

[0300] This approach considers a tag "neighborhood" to be all tags that share the same volume. There are significant differences between median revisit rates across volumes.Approach 3 - Fixed Radius

[0301] This approach considers a tag "neighborhood" to be all tags within 5 meters (this is configurable). This approach can provide median revisit rates of the "neighborhoods" such that there is some bi-modality in revisit rates across the "tag neighborhoods." Approach 4 - Clusters

[0302] In some embodiments, the RADAR RFID system uses the ceiling -mounted sensors described herein to interrogate passive RFID tags in a retail store or other environment.Attorney Docket No. RADR-026W001Each sensor has an antenna array that emits RF interrogation signals and detects responses to those RF interrogation signals from the RFID tags. Each response includes the electronic product code (EPC) of the corresponding RFID tag. The EPC encodes information about the item to which the RFID tag is attached.

[0303] In some embodiments, the RFID tracking system also includes a controller (e.g., a CA1010 controller) that is coupled to the sensors. The controller estimates the locations of the RFID tags from the responses detected by the sensors.

[0304] The controller also assigns each RFID tag to a corresponding location in the store called a home zone, which is where the RFID tag is "supposed to be" (discussed at length herein). These home zones may be not static predefined areas but dynamically created and updated based on empirical measurements. The controller creates these home zones and assigns RFID tags to the home zones based on a combination of estimates of the RFID tags' locations and user-specified parameters, including:1. the distance between home zones before a new home zone is created;2. the minimum number of RFID tags to form a new home zone; and3. the measurement period to use for home zone creation.

[0305] The controller may create these home zones and continuously adjust them through iterative clustering, responding to store merchandise movements and rearrangements. This empirical approach ensures that home zones reflect actual current merchandise layouts rather than outdated planograms or theoretical expectations.

[0306] Generally, the controller determines and assigns RFID tags to home zones by clustering RFID tags for the same product (e.g., style and color combination) based on their measured positions. It starts by determining the median estimated locations of the RFID tags for a given product over the user-specified measurement period, e.g., a 4-hour period over the preceding night (when the store was closed and the RFID tags were not moving). This can provide an empirical data foundation that automatically adapts to merchandise relocations. Next, the controller identifies one or more clusters of median estimated locations based on the user-specified minimum number of RFID tags per home zone and the user-specified distance between home zones. In some cases, the controller identifies multiple clusters, e.g., if the corresponding product is displayed at several different locations throughout the store. The system may then designate these clusters as home zones and assign RFID tags for the product to these home zones. As inventory is moved throughout the store (e.g., by staff), these clustersAttorney Docket No. RADR-026W001may shift during subsequent measurement periods, ensuring the home zones remain synchronized with actual store arrangements.

[0307] The controller can also estimate whether an item has been misplaced based on the RFID tag's estimated location, EPC, and home zone(s). To determine if a particular RFID tag has been misplaced, the controller determines the home zone(s) for that RFID tag, then calculates the distance between the RFID tag and the center of its (nearest) home zone. If the RFID tag is farther away from its nearest home zone than the threshold distance, then it may be flagged as misplaced.

[0308] This approach can be more advanced than others and generates "clusters of tags" utilizing either a two step k-means and agglomerative clustering approach or mean shift clustering. In other words, the store is split up into a number of clusters based on their position within the store and improved points for "centers of tag density."K-Means / Agglomerative ClusteringFor this analysis, a three-step clustering approach may be used:1. Generate an initial k-means with "large k" (100 initial clusters)2. Run agglomerative clustering and study the dendrogram to get a better understanding of the structure of the clusters and determine a cut-off and value for "k" (the number of clusters)3. Run a final k-means with the number of clusters determined from step 2

[0309] For Retailer A, choosing a cut-off of 8.5 generated 20 clusters. This means that a new cluster was formed when the cluster centers were more than 8.5 meters apart. This seems to be the most appropriate number based on the structure of how the clusters form.

[0310] The inventors also recognized that with 20 clusters, "within cluster sum of squares" is dropping slowly and silhouette scores are maximized (excluding a two-cluster solution which is not practically feasible).

[0311] For Retailer B and Retailer C, the same cut-off value was used, resulting in 27 and 26 clusters, respectively.Mean Shift Clustering

[0312] Mean shift clustering was also implemented for comparison purposes. It has the benefit of not explicitly requiring a number of clusters to be provided and may be easier to runAttorney Docket No. RADR-026W001across stores, as functions may not need to be written to automatically select an improved number of clusters.

[0313] With the hyper-parameter for "bandwidth," the mean shift clustering found fewer clusters than the k-means / agglomerative method.

[0314] As before, the inventors recognized differences in the median revisit rates across the clusters, regardless of method.Calculating Ratios and SKU Ranks

[0315] The final output of the process should be a ranking of SKUs based on their "staleness." In order to do this, the staleness of each EPC is calculated based on the ratio of the time since it was last read and the median revisit rate of the benchmark tags in its neighborhood (see example below):Median Revisit Rates of Ratio Median Revisit Rate to Time "neighborhood" Last ReadEPC Time Style Volum 5m Cluste Style & Volum 5m Cluste Since & e Radiu r Volum e Radiu r Last Volum s e sRead eXXXX 280 69 119 108 104 4.1 2.3 2.6 2.7 (EXAMPLE: 5m Radius Ratio = 2.6 = 280 seconds since last tag read / 108 seconds median time for all tags in 5m radius)

[0316] Using the ratios, the EPCs in a SKU can be ranked:Attorney Docket No. RADR-026W001Ratio Median Revisit Rate of SKU Rank "neighborhood" to Time Last ReadSKU EPC Style & Volum 5m Cluste Time Style Volum 5m Cluste Time Volum e Radin r Sine & e Radin r Sine e s e Volum s e Last e Last Read Read mi AAA 6.3 15.7 19.9 19.2 1,22 1 1 1 1 1 i A 4mi BBBB 3.0 7.5 9.4 9.2 586 2 2 2 2 2 inn CCCC 1.2 2.9 3.7 3.6 228 3 3 3 3 3iComparison of Methods

[0317] To understand the similarities in the ratios, the table below shows the correlations between the methods across the captures:Retailer A Retailer B Retailer C Night Day Day Fixed Radius Clustering - meanshift 0.97 0.94 0.88 Fixed Radius Time Since Last Read 0.93 0.39 0.81 Clustering - meanshift Time Since Last Read 0.9 0.39 0.93 Clustering - meanshift Volume 0.74 0.87 0.88 Clustering - K-Means Clustering - meanshift 0.72 0.98 0.97 Fixed Radius Clustering - K-Means 0.71 0.95 0.88 Fixed Radius Volume 0.71 0.91 0.99 Clustering - K-Means Time Since Last Read 0.68 0.39 0.91 vol ratio Time Since Last Read 0.59 0.31 0.81 Clustering - K-Means Volume 0.57 0.87 0.87 Fixed Radius Style & Volume 0.52 0.14 0.68 Clustering - meanshift Style & Volume 0.51 0.14 0.67 Style & Volume Time Since Last Read 0.49 0.07 0.64 Style & Volume Volume 0.39 0.17 0.68Clustering - K-Means Style & Volume 0.36 0.14 0.68

[0318] In general, there is a good level of alignment when comparing the ratios between the methods.

[0319] The ultimate outcome of the model is an SKU rank, so it is also beneficial to understand the strength of correlations between the ranks:Attorney Docket No. RADR-026W001Retailer A Retailer Retailer C Night B Day Day Clustering - meanshift Volume 0.99 0.99 1 Clustering - K-Means Clustering - meanshift 0.99 1 1 Volume Time Since Last Read 0.99 0.99 0.99 Fixed Radius Clustering - meanshift 0.99 1 0.99 Clustering - meanshift Time Since Last Read 0.99 1 1Fixed Radius Volume 0.99 1 1 Clustering - K-Means Volume 0.99 0.99 1Fixed Radius Time Since Last Read 0.98 0.99 0.99 Clustering - K-Means Time Since Last Read 0.98 1 0.99 Style & Volume Time Since Last Read 0.98 1 0.99 Fixed Radius Clustering - K-Means 0.98 1 1Style & Volume Volume 0.98 0.99 1 Clustering - meanshift Style & Volume 0.98 0.99 0.99 Clustering - K-Means Style & Volume 0.98 0.99 0.99Fixed Radius Style & Volume 0.97 0.99 0.99

[0320] Comparing the absolute rank of SKUs, the inventors recognized very significant correlation between all of the methods. The analysis suggests that a simple ranking of SKUs based on the time since last tag read could be sufficient for use cases such as EPC correction.Validating Performance

[0321] Three approaches may be taken for further validating the methods:1. Further data collection from hand scans and data captures2. Store simulation3. Store "unit tests"Further Data Collection

[0322] Further data may be collected, such as some ground truth data, by running two separate hand scans with a data capture in between (as illustrated below):

[0323] The order of events may be:1. Hand scan in the morning before the store opens2. 2 hour data capture before the store closes3. Hand scan in the morning before the store opens

[0324] This set up may allow collecting some "ground truth" on what was sold throughout the day and utilizing the data capture during store operating hours to build a SKUAttorney Docket No. RADR-026W001ranking from a snapshot in time and generating some performance metrics on how well the ranking can detect actual sold items (e.g., "precision and recall").

[0325] Some embodiments may extend the data capture to the full day of store operating without significant performance degradation.Store Simulation

[0326] The inventors test processes with store simulation, using the following scenarios:Attorney Docket No. RADR-026W001Store "Unit Tests"

[0327] For this validation, the inventors built unit tests to physically recreate within a store in order to simulate some examples that are likely to be rare but could pose challenges (and demonstrate the differences) in the processes herein. A full list of these "unit tests" may be developed, with examples below:Scenari Description What this may show01 Move a tag from high revisit rate That tag revisit rates may change their "neighborhoods" to low revisit rate revisit rates based on their neighborhood "neighborhoods" location even if it is being held by a customer and not in a stack2 "Purchase" items from different Like SKUs will be next to each other, for home locations at different rates the neighborhood correction to be useful it eg. may be needed to pick up reads from tags • Simulate different outside of their home neighborhood. These "journeys" - pick up, walk tests may help determine the degree to around etc. - a 15 minute which "moving tags" can be read outside of shopping trip their home neighborhood• Picking up an item,immediately "checkingout" and leaving• Picking up an item andleaving (without checkingout)3 Multiple scenarios where two of This may test the ability of the process to the same SKU are moved from detect the purchased item against another low / medium / high revisit rate item not in its home neighborhood but left neighborhoods at similar times - under different scenarios with different for one a purchase is simulated, for expected read ratesthe other...• is left in a visible area (notin a stack) of a low revisitrate neighborhood• is left in a non-visible area(in a stack) of a low revisitrate neighborhood• is left in a visible area of ahigh revisit rateneighborhood• is left in a non-visible area(in a stack) of a high revisitrate neighborhoodAttorney Docket No. RADR-026W001Deploying and Measuring Performance:

[0328] Based on the validation exercises, the inventors have recognized the processes to utilize in specific situations. The inventors recognized the value of measuring performance of these processes, which may be achieved via some system performance metrics. Such metrics may include:1. % of times an EPC initially selected for removal is not "corrected" (i.e., we got it right)2. % of times an EPC selected for removal from EPC correction is not "corrected" again (i.e., we got it right)3. # of EPC corrections4. Time to reconcile inventory (i.e., wall clock time until EPC converged / for how long was the inventory incorrect)5. Process runtime / complexity (because there may be a trade-off between accuracy and processor load)

[0329] Automating these system performance metrics will also enable A / B testing different processes. Due to the challenges of measuring accuracy through "ground truth," the inventors have recognized that small A / B tests to gain confidence in process updates may provide continual improvement.Clustering in Home Zones Process

[0330] The inventors have recognized and appreciated a methodology and approach to assigning EPCs to "home zones," and herein present analysis that demonstrates:1. How processes are working on Retailer A data from Location 12. A comparison to "ground truth" home zones collected via Retailer C hand scans and simultaneous data capturesUtility of a "Home Zone" Process

[0331] The inventors recognized that the location where an EPC is "supposed to be" can be very useful for availability analytics. Availability reports can include insights on "misplaced items," which make it useful to understand if the EPC is in the FOH or BOH and whether the item was in its home location with like products.Overview of Process Approach and Methodology

[0332] In some embodiments, a general principle of the process is to cluster tags of the same product (style & color combination) based on the distance between them. The current process may be based on agglomerative clustering where tags are iteratively joined intoAttorney Docket No. RADR-026W001"clusters." The tags would be combined by iteratively joining them based on distance. In this example, there are two "groups" of tags ~0.9 meters apart.

[0333] An advantage of this clustering approach is that it can rely on empirical measurement data rather than predefined zones or manual configuration. By processing the actual measured positions of tags, the system creates an accurate representation of the current store layout that dynamically updates as store associates rearrange merchandise. When products are relocated to new fixtures or display areas, subsequent measurement periods will capture these changes, and the home zones will automatically migrate to reflect the new arrangement. This dynamic adaptation to the constantly changing retail environment ensures that inventory analytics remain accurate without requiring manual reconfiguration when merchandise is moved.

[0334] To form home zones, the following rules may be used (these can be configured):1. A new home zone is generated if there is a distance of >3 meters between the centers of formed clusters.2. A "home zone" may, in some embodiments, only be declared if three or more tags are in the cluster (this is to differentiate a "home zone" from a misplaced product). This rule does mean that a product with <3 tags will not be assigned a "home zone." The analysis shows that applying this rule -65% of products (style & color combined) could be given a home zone but -95% of tags would be covered. This was true across both Retailer A and Retailer C example captures.

[0335] In some examples, the process was "trained" on the median location over a 4-hour night capture in Location 1. The inventors recognized the value of generating "home zones" over a similar period and when tags are not moving.

[0336] Consider a unisex item that has two home zones - one in menswear and one in womenswear. The home zones that the process generated in Location 1 are shown below. This demonstrates how the process is working, all tags in a home zone cluster are within 3 meters of each other and two home zones are formed as the centers of the two clusters are >3 meters apart. There are two clear clusters with -45 meters between the centers of each.Retailer C "Ground Truth " Analysis

[0337] The following section describes a validation analysis performed at Retailer C to assess the accuracy of the system's exit detection by comparing system predictions against manually-collected ground truth data.Attorney Docket No. RADR-026W001Overview of Validation

[0338] To further validate the process, a study from Retailer C may be used. In this study, the fixture of all tags are identified and the X / Y location of the center of the fixture are provided. For this analysis, the provided X / Y location was considered to be the ground truth for the tag "home zone."

[0339] A simultaneous data capture may be run at the same time as the identification of the fixtures (and home zones). Using this data capture, each EPC is assigned a home zone utilizing the process (assuming tags should be <3 meters apart and home zones should have >2 tags).

[0340] To understand the "accuracy" of the process, the assigned home zones for each EPC are compared to the ground truth "home zone."Results of Validation

[0341] At the overall level, the average distance from the assigned "home zone" from the process is found to be -l.lrn from the "ground truth." This is a strong result given that the ground truth is actually the center of the fixture and not the "true home zone."Attorney Docket No. RADR-026W001

[0342] To identify outliers, the distance is summarized at the product (style & color) level. The top 20 products with the largest discrepancies are shown below:Product (Style & Color) Avg. Distance to Ground Truth Number Tags BLACK COL09-452331 15.5 3OFF WHITE COL01-450310 13.5 14BLACK COL09-450311 12.9 24GREEN COL55-450310 12.9 16BLACK COL09-450310 7.1 7DARK BROWN COL38-451523 6.7 3BEIGE COL32-449632 6.3 8BLUE COL66-449882 5.9 3BLACK COL09-449864 5.8 3BEIGE COL32-450509 5.6 5BEIGE COL32-457209 5.6 4NAVY COL69-452679 5.5 7BLACK COL09-449722 4.9 9GRAY COL05-450310 4.4 6GRAY COL06-455472 4.4 6NAVY COL69-450529 4.0 6LIGHT PURPLE COL70-452852 3.9 5OFF WHITE COL01-450701 3.0 9WHITE COL00-450756 2.9 4OFF WHITE COL01-450771 2.9 3Attorney Docket No. RADR-026W001

[0343] The table shows that despite the overall strong results, there are (as expected) some outliers. Taking a deeper dive into one of the products illustrates the common theme that causes large distances. The table below shows all tags for BLACK COL09-452331 :EPC Product Assigned Assigned Ground Ground Fixture Distance (Style & Heme Home Truth Truth From Color) ZoneX ZcneY Home Home Ground Position Position ZoneX ZoneY Truth Position Position000027B BLACK 24.78 13.92 44.53 4.37 32e 21.94 COL09-4523312D0000 BLACK 24.78 13.92 44.53 2.76 32d 22.69 B COL09-4523310000013 BLACK 24.78 13.92 25.93 12.32 12c 1.97COL09-452331

[0344] The table shows that there are three ground truth fixtures. In this example, the three tags are all found in the vicinity of a particular fixture and assigned to that fixture. The cause of this may be two-fold:1. There was a timing discrepancy between the data capture and when the tag was assigned to the fixture2. There was an error in data collection

[0345] In both cases, this points to a data collection issue over a process issue. Analyzing multiple examples from the table of outliers, this was consistently found to be the case.

[0346] To solve for the above, the validation was changed to measure the distance between the assigned home zone from the process to the "nearest ground truth home zone." When this change was implemented the average distance between an assigned home zone and its nearest ground truth home zone fell to 0.8 meters.

[0347] There are still a handful of outliers even after this change. The table below shows the numbers of tags by distance to ground truth:Distance Number Tags % TotalAttorney Docket No. RADR-026W001All Tags 5,245 100% Assigned home zone >2 meters from ground truth 115 2%Assigned home zone >3 meters from ground truth 9 0.18%

[0348] This further illustrates that the majority of ground truth home zones were found within 2 meters of their "ground truth" with only a small fraction more than 3 meters away from their ground truths.How this Process May Be Used

[0349] As discussed above, a primary use case is finding misplaced items. When identifying if an item is misplaced, one may compare a tag's location to its nearest home zone. A tag more than a given distance (e.g., 2 meters or 3 meters, but any suitable distance may be used) from its nearest home zone (e.g., to an edge or to the center of the home zone) is classified as misplaced. This affects availability if there are no SKUs of the same style, color, and size within a given distance of the home zone (as above, most likely 2 meters or 3 meters).Setting Home Zone Parameters

[0350] To generate Home Zones, three parameters may be set that dictate how the Home Zones are created:1. The distance between two potential Home Zones before a new Home Zone is created.For example, if there are two groups of tags where the average positions are 1.5m apart, should these be grouped into a single Home Zone? Where this threshold is set has a strong influence on the number of Home Zones created.2. The minimum number of EPCs to form a Home Zone (e.g., are two tags close together enough to form a Home Zone). Where this is set changes the "coverage" of Home Zones; for example, if there are only two EPCs for a product observed in FOH and the threshold for the number of EPCs to form a Home Zone is three, then that product cannot be included in downstream analytics.3. The number of days of data to use for Home Zone creation (e.g., include all tags with a movement or status change over 1 day, 7 days, 14 days). This choice affects both the coverage (as discussed above) and Cloud costs, as increasing the look back period increases the amount of data to be processed.

[0351] The sections below discuss analysis undertaken to set these parameters (Note that these metrics are for Location 1 for Retailer A, however, the patterns are very consistent across ~20 Retailer A stores).Setting Distance between Home Zones

[0352] Some examples show the sensitivity of the number of Home Zones generated for a 3 meters distance vs. 6 meters.Attorney Docket No. RADR-026W001

[0353] When a 3 meter distance cut-off is used, 64% of products (style & color combination) have a single Home Zone, with 21% having 2 Home Zones. With a 6 meter cutoff, this changes to 79% of products with a single Home Zone and 15% with 2.

[0354] The inventors appreciated that 6 meters can be preferable distance setting. With a 3 meter setting, products that are clearly a single display get split into two Home Zones.

[0355] The trade-off of having fewer Home Zones is that the average distance to the Home Zone increases, however, this is never >3 meters, and -95% of the time this is < 1.5 meters.Minimum Number of EPCs for a Home Zone

[0356] The minimum number of EPCs required to form a Home Zone affects the number of SKUs where a Home Zone can be generated. For example, if only two EPCs for a given product were observed in FOH over 7 days, it may not be possible to generate a Home Zone if the threshold was set to 3.

[0357] The table below shows the % of total SKUs and EPCs where a Home Zone could be generated taking all SKUs / EPCs observed in FOH over 7 days (assumes a 6 meter distance):SKU Coverage EPC Coverage3 EPCs 84% 93%4 EPCs 79% 90%

[0358] Based on the above, the inventors recognized the threshold can be set to 3 EPCs in order to capture 84% of SKUs and 93% of EPCs. This means that 16% of SKUs and 7% of EPCs had < 3 EPCs in FOH for the entire product over 7 days, and they will be excluded from downstream analytics.Number of Days of Data

[0359] The number of days of location data also affects the coverage. With more days of data, it becomes more likely that an EPC will have a significant movement or a status change. The table below shows the coverage when different periods are used (assumes minimum of 3 EPCs):SKU Coverage EPC Coverage1 Day 72% 85%Attorney Docket No. RADR-026W0017 Days 82% 92%14 Days 82% 93%

[0360] The table shows a large jump in coverage from 1 day to 7 days, but no significant increase when changing to 14 days. Based on this, the inventors recognized that 7 days may be used.Examples of Single Home Zones

[0361] 0 - 0.5m Average Distance Examples (40% of products)

[0362] 0.5 - Im Average Distance Examples (47% of products)

[0363] 1 - 1.5m Average Distance Examples (8% of products)

[0364] 1.5 - 2m Average Distance Examples (3% of products)

[0365] > 2m Average Distance Examples (3% of products)Examples of Products with >1 Home Zone

[0366] The following are examples where multiple Home Zones are generated for the same product (-21% of products). It is unclear in these instances if the product is actually merchandised in multiple spaces or if the process is detecting errors in reads or locations where the items are typically misplaced.

[0367] Downstream processes may consider EPCs to be misplaced if they are not near any home zone (e.g., 3 meters), so the effect is to increase the amount of space considered to be the Home Zone and provide conservative estimates of misplaced items.Operationalizing Home Zone Creation

[0368] Some embodiments update Home Zones every evening. This may capture any changes to floor sets or planograms that happened that day and ensure the Home Zones used are the most up to date possible.Shared Wall Determination Process

[0369] Based on raw positioning, given current error range, the inventors have recognized that it is difficult to determine which side a tag is on of a shared wall between two volumes. For functionality such as restock, getting better accuracy contributes to the usefulness of the application. In practice, many tags end up changing volumes from one side of the wall to the other, due to the jitter of the system. When one takes a higher level approach, and looksAttorney Docket No. RADR-026W001at the positions of multiple reads, and groups together nearby tags of the same product id, one can use additional data to make a better educated guess.

[0370] The inventors have appreciated that this process may be particularly suited for scenarios where items sharing the same type, such as 'product id' or a combination of 'style' and 'style number', are grouped together; for example, a stack of grey t-shirts of type A. Such grouping may allow the process to outperform the baseline solution, which selects volumes based on individual tag reads.

[0371] However, in some embodiments the process may not be as suited for environments with chaotic stacking. In stores with such disorganization, product clusters may not form, and the process may default to performance comparable to the baseline (e.g., relying solely on live tag reads without clustering benefits).

[0372] In cases where a store exhibits no discernible order, the clustering feature (which is configurable) generally should be disabled. This modification may ensure that the process's performance remains unaffected by the lack of clustering in such settings.Steps Overview

[0373] In the below steps product id where is mentioned, the product id may be the style and / or style_number.Generate EPC to product id Mapping Table

[0374] In some embodiments, the interrogator controller may not store information from the product catalogue, but rather this may be stored in the retail backend. To communicate with the retail backend to inform for which epcs product id is needed for (e.g., style and style number), the field 'product id' can be added to tag reads which is already being sent to the retail backend. If the product id field is blank, the controller does not have the product id in the epc_product_id_mapping table. When the retail backend receives a tag read with no product id it checks if the epc has been bounded. If it has been bounded it will send back the epc with product id (e.g., style and style number). If it had not been bounded, it will send nothing back. At time that any epc gets bounded or rebounded, the retail backend should send to the controller the epc and product id.epc product id mapping

[0375] This table may be stored in the controller.Attorney Docket No. RADR-026W001Field DefinitionEpc Electronic Product Code for tags read.product id product id from the product catalogue for the given epc. Can be set as the style and style number.date last rea the date a tag was last read at updated in the clustering storage step. Used tod remove tags that are no longer being read.Store Reads

[0376] For each epc, store the last 5 tag reads that were read in the last 24 hour period (refresh table every 24 hours). The store reads data includes the fields epc, position x m, position_y_m, and position z m.Perform Clustering

[0377] In some embodiments, every 24 hours, the system may perform clustering on the stored tags reads. After each clustering, the previous cluster table is erased and updated with the new clusters. The clustering process can use a modified version of Density -Based Spatial Clustering of Applications with Noise (DBSCAN). DBSCAN can allow for large clusters as long as the samples are all densely packed. Some embodiments use a modified version of DBSCAN that incorporates a maximum radius that all points within a cluster need to be within from the centroid of the cluster.

[0378] There are two clustering steps that occur, the first takes in reads from only a single epc at a time. Clustering is performed at the epc level so that if a tag happened to move during the time the data was being stored, rather than storing the mean position, which would not represent a position where the tag was actually present, clusters would form in positions where a tag spent some time (e.g., at least 3 reads).

[0379] Now from the centroids of cluster position for each epc, the epc values can be looked up in the epc_product_id_mapping table to get their associated product id. Then clustering is performed at the product id level. All epcs of the same product id are clustered based on position. All clusters are stored (there could be multiple areas in the store where a certain type of product are displayed). Based on the centroid position of each cluster the cluster_containing_name_probability_pairs and cluster_containing_set_probability_pairs (limited to boh and fitting room) are calculated in the same manner the volume probabilities are calculated for a tag read. The following information may be stored in the cluster table.Attorney Docket No. RADR-026W001Field Definitionproduct id the product id of all epcs that are part of the clustercentroid x m the mean value of position x m for all epcs in the clustercentroid_y_m the means value of position_y_m for all epcs in the clustercentroid z m the mean value of position z m for all epcs in the clusterradius m max distance set in the clustering process to constrains the size of a cluster. Should represent physical space, ie size of fixture. cluster_containing_name_probability_pai the volume probability based on the centroid of rs a cluster cluster_containing_set_probability_pairs the volume probability limited to foh and fittingroom based on the centroid of a cluster Utilize the Clustering Table to Generate containing name probability pairs

[0380] For all tag reads, map the EPC to product id using epc_product_id_mapping, see if for that product id, there exists a value in the clustering table where the following conditions are met:• abs(position_x_m[tag_read] - centroid x mfcluster]) <= radius mfcluster] + error x m [tag read]• abs(position_y_m[tag_read] - centroid_y_m[cluster]) <= radius mfcluster] + error_y_m [tag read]• abs(position_z_m[tag_read] - centroid_z_m[cluster]) <= radius m [cluster] + error z m [tag read]

[0381] If so, set containing_name_probability_pairs of the tag read to cluster_containing_name_probability_pairs of the cluster and set the BOH and fittingroom fields of containing_set_probability_pairs to cluster_containing_set_probability_pairs of the cluster. If there are multiple clusters that meet this criteria, pick the cluster whose centroid is closest to the tag read. If there are no clusters within the set range, set containing_name_probability_pairs and containing_set_probability_pairs to the values of containing_name_probability_pairs_unclustered and containing_set_probability_pairs_unclustered, respectively.Attorney Docket No. RADR-026W001Availability Process

[0382] The inventors have appreciated ways to generate availability KPIs and calculate metrics and herein provide an example of preliminary results from Location 1 and Location 2, as well as a number of deep dive examples of SKUs from Location 1.Overview of Process Approach and Methodology

[0383] The process may rely on first generating Home Zones for products (style & color combinations). Once these Home Zones are generated, it may be possible to determine the expected location(s) of products (and SKUs) within the store.

[0384] The process generates four "states" with regards to its availability:1. Available: The SKU has >0 EPCs "at home" (within a given distance of the Home Zone)2. Misplaced: The SKU has 0 EPCs "at home" (within a given distance of the Home Zone) but >1 EPCs elsewhere in FOH3. Stockroom: The SKU has 0 EPCs in FOH but >0 EPCs in BOH4. Not in store: The SKU is on the floor set but 0 EPCs had unit state "available" Setting "Misplaced SKU" Parameters

[0385] To identify misplaced SKUs, two parameters may be determined:1. The distance from the Home Zone and item for the item to be determined "misplaced" 2. The time an item is away from the Home Zone for the item to be determined "misplaced"

[0386] To obtain a better understanding of the sensitivity of these parameters, the process may be executed with varying values, and the proportion of total minutes that SKUs were "misplaced" may be recorded. The tables in FIG. 10 show the results of this experiment from Location 1 and Location 2 for one day. The inventors recognized that the misplaced item lost availability is very sensitive to the distance threshold but much less so to the time threshold:• In both stores the variance from using a 5 -minute threshold to a 1-hour threshold was ~1%• The variance from using a 2-meter threshold to a 6-meter threshold was ~6% and ~5% in Location 2 and Location 1 respectively

[0387] FIG. 11 shows a chart of the lost availability for both stores by the distance threshold, assuming a 5-minute time threshold. There is a clear inflection point at 3 meters, however, since that is considered to be the 95th percentile of error in location accuracy, theAttorney Docket No. RADR-026W001inventors have determined to use 4 meters for further analysis. Because there are only minimal differences between the time bands, a 5-minute threshold is used for time misplaced.Experimental Results

[0388] FIG. 12A shows a table of the KPI metrics for Location 2 and Location 1. The inventors recognized the following:1. Data is restricted to store business hours only2. "Not in store" here is based upon a floor set consisting of any SKU observed in FOH in the last 7 days for each store3. As discussed above, 4 meters and 5 minutes are used for the distance and time away from Home Zone thresholds when designating "misplaced items"

[0389] The run of the process shows on this specific day that Location 2 had slightly worse availability than Location 1. The drivers of availability were, however, very different between the two stores:• Location 1 had a much larger problem with items not being in store than Location 2 with ~3 times more lost availability than Location 2 (based on the floor set definition of SKUs observed in the FOH in the prior 7 days)• Location 2 had a much larger problem with misplaced items and items in the stockroom with ~2X more lost availability than Location 1 for these reasons

[0390] FIG. 12B shows a table of the hourly breakdown. In both cases, availability increased throughout the day mostly driven by improvements in "items not in store." The likely driver of this is items being received and tags gradually being read by the system throughout the day.

[0391] In Location 2, the lost availability from items in the stockroom decreased throughout the day.

[0392] In Location 1, the lost availability from misplaced items increased throughout the day.Misplaced Items

[0393] In this section, examples of misplaced items will be shown for Location 1.

[0394] FIG. 13A shows one example: a straightforward example where a single SKU of the "Flex Lived-In Khaki Short" was misplaced the entire day.Attorney Docket No. RADR-026W001

[0395] FIG. 13B shows analysis of locations for all EPCs of the SKU found: a single EPC in the same location throughout the day.

[0396] FIG. 13C shows an example of a single SKU of "Flex Slim Lived-In Khaki Pant" misplaced throughout a part of the day.

[0397] FIG. 14 shows (on the left) the time-weighted positions of EPCs for the product that were used to generate the Home Zone and (on the right) tag locations from the misplaced SKU.

[0398] FIG. 14 shows an example illustrating the trade-offs that may be made with regards to the distance threshold. Looking at this example, it is likely that the SKU was misplaced throughout the day, but we can only be "certain" about a proportion of the locations that were >4 meters from the Home Zone. This illustrates how thresholds used in some embodiments are providing more conservative estimates of misplaced items.

[0399] FIG. 15A shows an example of another case where the SKU was misplaced for part of the day, but the data is more clearcut. The example is for the " Stretch Mom Jean" where two SKUs were apparently misplaced for part of the day as shown in the table in FIG. A9.

[0400] FIG. 15B shows a map indicating the locations of the two SKUs highlighted throughout the day.

[0401] FIG. 15C shows timestamps that also help to illustrate what has happened with these SKUs. In this example, at -18:58 (UTC) one EPC from each SKU was moved from BOH to FOH but placed in an incorrect location ~10m from the Home Zone. At -21 : 15 (UTC) both EPCs were moved to be -1 meter from the Home Zone and began to be considered "available." Stockroom Items

[0402] In this section, examples of items in the stockroom will be shown for Location 1.

[0403] FIG. 16A shows a first example, where the tag locations identify an SKU of "Flex Slim Lived-In Khaki Pant" that was in the "stockroom" for a portion of the day.

[0404] FIG. 16B shows a map indicating the tag locations identified throughout the day. The inventors appreciated that locations throughout the day are in the BOH.

[0405] FIG. 16C shows analysis of the timestamps (in UTC), which provides more information. In this example there were two EPCs of the SKU in the BOH with none in the FOH at the start of the day. At -19:48 (UTC) two EPCs were moved from the BOH to the FOHAttorney Docket No. RADR-026W001with both being located ~0.4 meters from the Home Zone marking the item as available again. The timestamps for the EPC ending "E3C" show the item in motion moving toward its FOH Home Zone.

[0406] FIG. 16D shows an example where all SKUs of a product are available except a single size which is in the stockroom for the entire day, as shown in FIG. 16E.Shared Wall Problem

[0407] The inventors appreciated what is called herein the "shared wall problem" impacts these metrics, and the inventors have appreciated the techniques herein may help ensure accuracy in light of the shared wall problem. As explained above, the inventors recognized the difficulty of determining which side a tag is on when there is a shared wall between two volumes. In practice, many tags end up changing volumes from one side of the wall to the other due to system jitter. The inventors appreciated that analyzing positions of multiple reads and grouping together nearby tags of the same product id may provide a better educated guess.

[0408] This process is particularly well-suited for scenarios where items sharing the same type, such as product id or a combination of style and style number, are grouped together (for example, a stack of grey t-shirts of type A). Such grouping is useful for the process to outperform a baseline solution that selects volumes based on individual tag reads. In stores with chaotic stacking where product clusters do not form, the clustering feature can be disabled to ensure performance remains unaffected.

[0409] FIG. 17A demonstrates this example via SKUs of "Knit Ruched-Front Midi Dress."

[0410] Most SKUs are considered "in the stockroom" with one SKU "misplaced." FIG.17B shows a map indicating (on the left) the Home Zone generated for the item based on the FOH locations that were read. On the right is all reads from these SKUs.

[0411] In this case, it is difficult to determine if the SKUs were intended to be in the stockroom only, but it could be argued from this map that this is the case. The solution for the shared wall would most likely identify all of these locations as BOH based on the density.

[0412] The inventors have recognized that having information on the real floor set also helps address this. If the item was not intended for FOH, it would not have been on the floor set and would not have been considered when generating the metrics.Attorney Docket No. RADR-026W001 MLED: Exit Probability Modeling

[0413] The inventors have recognized several priorities for measuring uncertainty in state and location via exit probability, including:1. EPC Removal Selection & Correction: Rank ordering EPCs based on their likelihood of having left the store2. EPC Audit Shortlist: Providing ranked lists of items to audit to confirm if they remain in the store and their locations3. Providing a measure of confidence in location to end app users in Lookup

[0414] In some embodiments, this may be achieved through a simple ranking of EPCs by time since last read. An alternative methodology correcting for the "neighborhood read rates" can also be used.

[0415] The inventors determined to implement simple time-based ranking of EPCs by their likelihood of having exited the store, used to prioritize manual audit verification. The inventors recognized the differences between the read rates within neighborhoods was not sufficient to drive a difference in the ranking of EPCs.

[0416] In some embodiments, an "exit probability" model may be used to determine, for every EPC at specified time points, the probability that the EPC remains in the store. This probability can then be used to determine the ranking of EPCs by SKU for priorities #1 and #2 described above and as a measure of location confidence for priority #3.

[0417] The output of this analysis may be an understanding of:1. "How much better" than the simple time based approach (measured by % of EPCs correctly identified as having "exited") utilizing more data from a variety of sources with more sophisticated modeling approaches can achieve.2. The importance of features and their source to prioritize feature sets to send to cloud.3. The value of POS in addition to 1st party data (e.g., if we add POS, exit detection is X% more accurate).4. The stability of models (over the short and medium term) to determine how often they should be retrained. This will involves "validation data" over a number of time periods (see validation section).5. The difference in accuracy of developing one model per store, one model per "store cluster," or one model for all.Ground Truth Approaches

[0418] In order to model "exit probability," the inventors recognized the value of a "ground truth" label that identifies EPCs that left the store. Some embodiments include twoAttorney Docket No. RADR-026W001main approaches to modeling, which are described below. These are utilizing audit data and developing a "proxy-label."Audits

[0419] In this approach, the inventors appreciated the utility of the audit data obtained from stores as the "ground truth" label. Since regular audits occur in stores, the audit data provide an ongoing feed of labels to train and re-train models.Developing a Proxy-label

[0420] Some embodiments send location and state changes to the cloud and build up a history of this information for analytics and modeling. This approach may utilize this data to generate a "proxy-label" with some simple metrics. Two such metrics would be:1. Time since last location change2. Time since last status change

[0421] Utilizing features such as these, a label could be generated that would indicate the item is believed to no longer be in the store. The approach is analogous to a customer chum model where a definitive label cannot be defined; that is, it is common to define a label such as "churned from loyalty program" on the basis of a feature such as "time since shopped," on which a more sophisticated model is developed to proactively predict that label. In this case, "tag churn" (exit probability) would be predicted.

[0422] To define the "label," hand scan data may be used to define a one-off "ground truth" on samples across a number of stores. Using this hand scan data, historical data may be analyzed to determine how a "proxy-label" could be defined. The charts below are illustrative of this.

[0423] If the data were real, it may be possible to be "confident" generating a label at 15 days since location change. If needed, two features could be combined such as status change and days since last location change. However, the inventors recognized it may be desirable not to combine too many features at this point, as a goal is not to predict the label but to feel confident in definition of a tag having left the store.Attorney Docket No. RADR-026W001Pros & Cons of ApproachesPros ConsAudit • The data is known to be "true" • The data has selection bias, as there Data (with some human error) are rules that determine which EPCs • Does not require any extra are selected for auditseffort for data collection, asaudits occur as part of theprocess• Will provide an ongoingsample for model retraining• There is minimal (potentiallyno) lag between the label andthe observation data (seediagram in model set up)Prox • Once confident in the rules for • Requires one-off data collection y label generation, can generate (though could perhaps utilize Label samples from any store on planned and existing hand scan any time period data)• Not a "true label," and there will be some error in the labels• Concerns about rules being different by store (e.g., 10 days is good in one store but need 15 days in another)• There will be a lag between when the label is defined and the observation data for model development (see section on modelset up)Data

[0424] In order to gain an understanding of the importance and value of data in predicting tag exit, data from multiple sources are utilized for model development and training. Some of this data will be very different by store, and there may be need to assess which features to use for "general models" vs. store specific. In some cases, some embodiments could "normalize" certain features within stores and attempt to utilize them as features in a general model.Examples of Features Developed from Cloud• Last known location (x,y,z)• Last known volume• Last known stateAttorney Docket No. RADR-026W001• Time since last state change• Time since last location change• Number of locations with >2 meters movement in the last 1 minute, 5 minutes, 10 minutes etc.• Number of locations with >3 meters movement in the last 1 minute, 5 minutes, 10 minutes etc.Captures• Read rate in the last 1 minute, 5 minutes ... for EPC• Read rate in the last 1 minute, 5 minutes ... for "neighborhood"• Channel estimatesPOS• Average rate of sale of SKU in the last 1 day, 7 days, 1 month• SKU priceProduct Catalog• Style• Category• Store (embedding)Model Set Up

[0425] For the audit label, models with a very small "fallow period" (lag between the observation period and the label) may be used. With the "proxy label," the inventors recognized the value of waiting until the label can be confidently applied.

[0426] Analysis may determine fallow periods and observation periods.Validation

[0427] In order to validate the models, the following may be done:1. Utilize available audit data2. Utilize new or existing hand scans3. "Simulate exits" with store test casesAttorney Docket No. RADR-026W001 Exit Probability Example Embodiments

[0428] Some embodiments of the exit probability model include:• Feature Engineering: This pipeline generates features that are used for training / serving. The same pipeline and code generates features for both to ensure consistency.• Training: Automated training pipeline (e.g., with Kubeflow pipelines in Vertex Al) that:o Uploads hand scanso Generates ground truth labels from the hand scans, identifying if EPCs were found in the scan (label 1) or not (label 0)o Pre-processes the data by splitting into training / testing / validation sets, scales numeric features and one-hot-encodes categorical featureso Runs Exploratory Data Analysis (EDA) on the featureso Trains an ML model - e.g., an XG Boost model; hyper parameter optimization is optionalo Runs model validation providing metrics on inventory accuracy, ROC, Fl and model accuracy; allows for batch evaluation for out of sample validation o Executes an automated champion / challenger model evaluation that updates models to point to model versions that are the most performant • Prediction: Scheduled prediction pipeline to generate exit predictions using the trained models from the model registry (e.g., via Vertex Al) and the prediction service (e.g., via Vertex Al).

[0429] The following sections will contain more details on each module.Feature Engineering

[0430] This section outlines the feature engineering module, which generates the features required for training and serving the ML exit predictions from Tag Events, as illustrated in FIG. 18, etc., at top left to Feature Store at top right.Data Inputs

[0431] The features are generated from the following Big Query sources:• tag-event_v2_unpacked: Full history of tag eventsAttorney Docket No. RADR-026W001• sku-ean-mapping: Contains product catalog data (i.e., SKU and product attributes) • space-mapping: Maps space id to store id• home zones: History of product home zones• inventory normalized event: History of SKU inventory eventsFeature Entities

[0432] The features generated fall into four entities:1. Last Known Tag Positions and States

[0433] Generated from a "last known state volume" process that creates the following features:• Last known tag read states• Last known tag unit states• Last known volume• Minutes since last position change (using execution datetime as the anchor)• Minutes stale• Distance in meters from home zone at last movement position• Binary flag identifying if last movement was during store hours• Hour of last position change (in local store hours)2. Tag Movements / Distance Traveled

[0434] Generated from a "tag movements" process that creates the following features:• The distance traveled in meters in the 5, 10, 15, 20, 30, 45, and 60 minutes since the last movement• The distance traveled in meters in the 5, 10, 15, 20, 30, 45, and 60 minutes since the pipeline execution datetime• The number of movements in the 5, 10, 15, 20, 30, 45, and 60 minutes since the last movement (including the last movement)• The number of movements in the 5, 10, 15, 20, 30, 45, and 60 minutes since the execution datetimeAttorney Docket No. RADR-026W0013. Time Since Visited Volumes and Number of Visits

[0435] Generated from a "volumes" process that creates the following features:• mins_last_[volume]_mvmt: The number of minutes since last observed movement in the specified volume.• mins_btwn_[volume]_mvmt_last_mvmt: The number of minutes between last observed movment in the specified volume and the last position timestamp (movement).• [volume]_visits_ll: The number of times in the last day where a tag was observed in the volume. The definition of a "visit" is a read in the volume >90 minutes from the prior read.• [volume]_visits_17 : The number of times in the last 7 days where a tag was observed in the volume. The definition of a "visit" is a read in the volume >90 minutes from the prior read.• [volume]_visits_ll_last_mvmt: The number of times in the last day (from the last position timestamp) where a tag moved in the volume. The definition of a "visit" is a movement in the volume >90 minutes from the prior movement.• [volume]_visits_17_last_mvmt: The number of times in the last 7 days (from the last position timestamp) where a tag moved in the volume. The definition of a "visit" is a movement in the volume >90 minutes from the prior movement.4. Inventory Event Features

[0436] The Big Query table is populated from the "inventory normalized events" table. From the normalized events the following features are generated:• Minutes between the last read (movement) and the closest event• Minutes between the last read (movement) and the closest sale, return, transfer ... • Number of SKU events in the 1, 5, 10, 15, 30, 60, 1 day, 3 days, and 7 days since last movement• Number of sale / return / transfer... events in the 1, 5, 10, 15, 30, 60, 1 day, 3 days, and 7 days since last movementsAttorney Docket No. RADR-026W001

[0437] Some embodiments may use a separate directed acyclic graph (DAG) (e.g., "feature engineering training dag"), which is a simplified DAG that can be triggered with a specific date that is needed that matches a hand scan date and generates a separate Big Query table that allows WRITE APPEND. This is because training happens over multiple snapshots that often occur on the same day. This DAG is also simplified to remove partition sensing steps and the use of Airflow macros that were failing as they were set up to work for scheduled runs not manually triggered specific dates.Feature Store Output

[0438] The feature entities are merged together and inserted into a Big Query table that contains the history of features generated for every execution of the pipeline, in some embodiments.Model Training

[0439] FIG. 18 illustrates a model training module from Orchestrated Model Training Pipeline to ML Metadata Store & Model Registry. This module is used to train the exit probability models utilizing features from the Feature Engineering pipeline. Models may be trained using, for example, Kubeflow pipelines within Vertex Al. Vertex Al may be used for artifact management, storing trained models, pipeline runs, and experiment tracking.Pipeline StepsGenerate Ground Truth Labels

[0440] Includes the following three steps:1. Ingest Hand Scans: Ingests the raw hand scan data from a cloud service bucket, normalizes and writes to a comma-separated variable (csv) file or other suitable file 2. Upload Hand Scans: Uploads the normalized hand scan data from the csv in to Big Query3. Generate Labels: Generates binary labels where tags not found in the hand scan (exited) are the positive classMerge Ground Truth & Features

[0441] Merges the features generated from the Feature Engineering pipeline with the hand scan labels for the appropriate snapshots to generate data for model training and validation.Attorney Docket No. RADR-026W001 Pre-Processing

[0442] Splits data into training, testing and validation datasets, pre-processes data by standardizing numeric features and one-hot-encoding categorical features.Feature EDA

[0443] From the training data, generates the following EDA outputs:• Profiles of features (histograms and relationships with the exit label)• Correlations between features with simple Pearson Correlation and utilizing Feature Agglomeration to generate feature clusters• Feature ranking and variable selection using multiple linear and non-linear methods (e.g., absolute regression coefficients, simple squared correlation, Random Forest Feature Importance, Recursive Feature Elimination and Best Subsets)Model Training

[0444] The pipeline currently utilizes, for example, an XGBoost model for training. Configuration options allow for hyper-parameter turning, for example, utilizing Vizor (a Bayesian parameter optimization) in Vertex Al.Model Validation

[0445] Validates the trained model on test set validations and writes the results into the model artifact registry. Surfaces the validation results into the UI (e.g., a Vertex Al UI). The following metrics are calculated:• Inventory accuracy• AUC• ROC• Model accuracyChampion / Challenger

[0446] Updates the default model version to the trained model based on a specified performance metric if it improves upon a parent champion model.Attorney Docket No. RADR-026W001 Prediction

[0447] This module utilizes the trained ML models stored in the Model Registry to generate and serve predictions, as illustrated in FIG. 18 from Orchestrated Batch Prediction Pipeline downward.

[0448] In some embodiments, the system includes monitoring and testing capabilities including:• Unit testing of the Batch Prediction pipeline• Integration testing of the combined Feature Engineering and Batch Prediction pipelines • Dashboards showing trending and distributions for:o All key input featureso Predicted probabilitieso Tags being set from LOST to AVAILABLE over timeo Accuracy utilizing spot hand scansPipeline StepsExtract Features

[0449] Extracts the features from the Big Query Feature Store for the required date / time.Batch Prediction Service

[0450] Utilizes the trained model and pre-processor with the Vertex Al Batch Prediction Service to generate the exit probabilities. The prediction service may use a custom predict class (e.g., in a Docker container stored on Artifact Registry).Model Prediction Processing

[0451] Normalizes the prediction output into the format for storing prediction history and the format needed to utilize the NIEF table for tag state changes.Prediction History

[0452] Stores the history of EPC exit probabilities for all EPCs over all pipeline runs to allow for full debugging when combined with the feature history from the Feature Store and tag state changes.Attorney Docket No. RADR-026W001Tag State Change

[0453] Utilizes Dataflow and Pub / Sub to send tag state changes to the controller via the Normalized Inventory Event Feed (NIEF).Model Monitoring and Alerting

[0454] Monitoring and alerting within the MLED pipeline includes:• Alerts on exit volumes (e.g., when volumes for a store are 0 or above a specified threshold)• Data skew monitoring via Vertex Al Model Monitoring service: Alerts when the distributions of features used for prediction differ more than a specified threshold from the original training dataCI / CD

[0455] Some embodiments include CI / CD pipelines (e.g., with Jenkins) that execute the following:1. Runs unit tests: unit tests are written for all Feature Engineering methods2. Deploys models: copies and deploys models into the required environment3. Builds prediction container: Builds the prediction container and uploads to the Artifact Registry in the required environment4. Deploys batch prediction pipeline: Copies Vertex Al pipeline for prediction into the Cloud Composer bucket in the cloud service to be accessible by Airflow5. Deploys Airflow DAGs: Copies Airflow DAGs from the repo into the Cloud Composer bucket in the cloud service to deploy into the required environment Model Performance

[0456] The inventors have been able to provide a model able to achieve a maximum performance of 96.1% on a hand scan on a given evening. This is based on scoring all tags that did not have the "lost" unit state before the hand scan. This step may be necessary to remove tags that exited during a sensor outage.

[0457] The 96.1% accuracy is based on a probability cut-off of 95% (e.g., only tags with probability >= 95% were flagged as exited). The table below shows how these change depending on the cut-point chosen:Attorney Docket No. RADR-026W001Units In Units In In Radar But Not in Radar Probability Units In InventoryRFID Ground Not In Hand but In Hand Cut-Off Both AccuracySystem Truth Scan Scan 0.5 38,522 39,971 38,151 94.5% 371 1,820 0.6 38,767 39,971 38,363 94.9% 399 1,603 0.7 38,983 39,971 38,554 95.4% 429 1,417 0.8 39,188 39,971 38,724 95.7% 464 1,247 0.9 39,383 39,971 38,866 95.9% 517 1,105 0.95 39,583 39,971 38,988 96.1% 595 983 0.99 40,120 39,971 39,096 95.2% 1,024 875

[0458] To collect training data, scans may be performed in pairs: for example, a starting scan at the end of day and an ending scan at end of day, conducted across many days. This process may be repeated for multiple stores over multiple weeks to establish ground truth about what items are in stores versus what items have left. This paired scanning approach helps provide certainty about inventory status and movement.Importance of Model Features

[0459] The inventors have recognized the cumulative impact that a feature has had on determining whether an EPC has exited or not.

[0460] The inventors have recognized the POS events signals are very beneficial to the model prediction with the number of minutes between the last movement for an EPC and the closest event (of any type) for the SKU being number 2, and the number of minutes between the last movement for an EPC and the closest sale event for the SKU being number 3.

[0461] The inventors have recognized ten useful features for training and executing the model:Feature Definition Impact on exit As the time between The number of minutes the last position change between the hand scan and the hand scan mins_since_last_positionand the last known increases the movement >lm probability of exitincreasesAttorney Docket No. RADR-026W001The number of minutes When an EPC has an between the last known event close to its last mins btwn event last read movement for an EPC known movement it and an event for the has higher exit SKU probabilityThe number of minutes When an EPC has a between the last known sale event close to its mins btwn sale event last read movement for an EPC last known movement and a sale event for the it has higher exit SKU probabilityAs the distance between the product The distance in metersHome Zone and the between the last known last distance to home zone last known position position and the Homeincreases the Zone for the productprobability of exit increasesIf the last known The tag was lastposition was in the positioned in the last known volume cashwrap cashwrap volume the cashwrap (withprobability of exit probability >0.5)increasesThe number of "visits"to the cashwrap in the24 hours since the lastknown tag movement>lm. The definition of As the number of visits a "visit" is a position to cashwrap in the last cashwrap vi sits l 1 last timestamp read in the volume >90 day increases the minutes from the prior probability of exit position read (i.e., increases consecutive readswithin 90 minutes areclassified as a singlevisit).Tags that move during store hours (especially The hour of the lasthour last move in the evening) have a position change >lmhigher probability of exitAttorney Docket No. RADR-026W001As the distance The distance in meters traveled in the 5 that a tag traveled in the minutes before the last distance_traveled_5_min_last_timestam5 minutes before the position changeP last known position increases, the change probability of exit increasesAs the distance The distance in meters traveled in the 10 that a tag traveled in the minutes before the last distance_traveled_10_min_last_timesta10 minutes before the position change mplast known position increases, the change probability of exit increasesIf the last known The tag was lastposition was in the positioned in ecomm last_known_volume_ecomm_packing ecomm packing packing (withvolume the probability probability >0.5)of exit increasesAdditional Considerations

[0462] The inventors have recognized it is desirable to ensure tags that are still in a store are not falsely read as having exited. Such false reading may cause premature inventory replenishment based on false "absolute zeros" (e.g., the system reports that restock is needed when it is not needed).

[0463] The inventors have recognized potential difficulties with full fleet rollout, such as the following:• Training MLED models in different ways while considering generalizability.o Stores are often different sizes and different characteristics (traffic, assortments, sensor positioning and layouts, etc.).o Outlier stores will likely perform poorly.• Robust methods for testing and monitoring fleet-wide rollout.

[0464] To address such difficulties, some embodiments identify clusters of stores with similar characteristics, and work with customer success to identify stores in each cluster to collect ground truth data. The inventors have appreciated value in choosing at least two storesAttorney Docket No. RADR-026W001from each cluster (for training and validation) and scanning both in morning and in evening across multiple days.

[0465] In some embodiments, the system includes capabilities to update tag states from lost to exited specifically for MLED-predicted exits, to avoid confusion with lost states set through other means, such as audits.

[0466] The inventors have appreciated that features or metrics regarding a tag that drive exit probability are:• Distance the tag travels before last known movement• Last known position• Last known distance from home zone• Time, type, and number of IEF events around when an EPC was last observed to move

[0467] In some embodiments, the model handles different scales of implementation:• For backlog cleanup: The system can process large volumes of historical stale tag data to establish initial inventory accuracy.• For ongoing maintenance: The system applies exit probability detection continuously to maintain accurate inventory levels.• For new deployments: The system can be customized based on store characteristics including size, layout, inventory patterns, customer traffic, and sensor positioning. Time From Event Profiles

[0468] FIG. 19 shows a chart of a plot of exemplary time between an EPC's last movement and its closest event. This chart is based on a sample of tags in the database with the ground truth taken from the training data that was based on hand scans taken on the final day in multiple locations. Only tags that last moved from the first day onwards were considered.

[0469] The chart shows that the maj ority of tags that exited received an event very close to their last movement. However, there are very long tails on either side and a number of tags with no event at all, as shown in the table below:Minutes Between Closest SKU Event & Last Read Exited Tags % Exited Tags<-60 Minutes 3,811 7.3%>60 Minutes 3,434 6.6%Attorney Docket No. RADR-026W001>-60 , <60 42,439 81.2%No Event 2,581 4.9%All Exits 52,265 100.0%

[0470] -5% of exited tags had no events for the SKU, with a further -13% more than 60 minutes on either side of the last movement time.

[0471] Analyzing the same data for sale events shows a similar picture.Fitting Room Analytics and Product PlacementFitting Room Analytics

[0472] The inventors have recognized and appreciated the value of comprehensive fitting room analytics as a standalone capability derived from RFID tag movement data in retail environments. In some embodiments, the system may detect when an RFID tag enters a fitting room zone and records a fitting room session. A session may begin when an RFID tag attached to a product is detected within the fitting room zone and may end when the tag is no longer detected within that zone.

[0473] In some embodiments, the system may track and display any of the following fitting room metrics, as illustrated in FIG. 20A: (a) number of sessions, which may be the count of fitting room visits for a given product over a specified time period; (b) dwell time, which may be the duration each tagged item spends in the fitting room zone during a session; (c) recovery time, which may be the elapsed time from when a tagged item exits the fitting room zone to when it returns to the sales floor; (d) conversion rate, which may be the percentage of fitting room sessions resulting in a purchase of the tried item, as correlated with point-of-sale data; (e) percentage of units sold from the fitting room, which may be the proportion of total unit sales for a product that were preceded by a fitting room session; and (f) percentage of sales dollars from the fitting room, which may be the revenue proportion attributable to purchases preceded by a fitting room session.

[0474] FIG. 20A shows an example fitting room analytics dashboard user interface. The dashboard includes navigation tabs for viewing metrics at a fleet level, a district level, a store level, and a product details level. A filter bar allows selection of department, category, and product identifier. A metrics toggle bar enables users to select which metrics to display, including sessions (number of fitting room visits), SKUs tried (number of distinct products), conversions (number of purchases following fitting room trials), conversion rate (percentageAttorney Docket No. RADR-026W001of trials resulting in purchase), dwell time (duration in fitting room), recovery time (time from fitting room exit to sales floor return), percentage of units sold from fitting room (proportion of total sales preceded by fitting room trial), and percentage of sales dollars from fitting room (revenue proportion from fitting room conversions). A hierarchical table displays key performance indicators (KPIs) by product hierarchy, including sessions today, sessions yesterday, sessions over the last seven days, conversions today, conversions over the last seven days, conversion rate today, and conversion rate over the last seven days. The table supports hierarchical drill-down from department level to category level to individual product level, allowing users to identify fitting room performance patterns at various levels of granularity.

[0475] In some embodiments, the dashboard also displays a product ensembles table, as shown in the lower portion of FIG. 20A. A product ensemble may include two products from different categories that are tried together within the same fitting room session (e.g., jeans and a shirt). The product ensembles table may display, for each product pair: the product identifiers (e.g., P_0437, P_3376_A), the product names (e.g., "Low-Rise Jean," "Cozy Crew Neck Shirt"), the respective categories (e.g., "Fashion," "Core Fashion"), the number of cooccurrence sessions over a specified period (e.g., 1,666 sessions), the individual conversion rates for each product in the pair (e.g., 16.7%, 35.0%), the conversion rate when either product in the pair is purchased, and a product ensemble similarity score that quantifies how frequently the products are tried together.Fitting Room KPI Trends

[0476] In some embodiments, the fitting room analytics dashboard further includes a KPI trends view, as illustrated in FIG. 20B. The KPI trends view displays a time-series line chart 2080 plotting a selected fitting room metric over a date range (e.g., thirty days). Selectable metrics include sessions, SKUs tried, conversions, conversion rate, dwell time, recovery time, percentage of units sold from the fitting room, and percentage of sales dollars from the fitting room. The horizontal axis represents dates (e.g., daily granularity, with date labels along the axis) and the vertical axis represents the magnitude of the selected metric (e.g., session count ranging from 0 to 1,000,000). Data points at each date are connected by a trend line 2082, and each data point may be indicated by a marker (e.g., a circle). This time-series visualization enables users to identify trends in fitting room activity, such as periodic weekly cycles with higher session counts on certain days of the week, seasonal spikes associated with promotional events, or anomalous drops warranting investigation. The KPI trends view may support theAttorney Docket No. RADR-026W001same fleet, district, store, and product details navigation tabs as the main fitting room analytics dashboard.Dwell Time and Purchase Correlation

[0477] In some embodiments, the system analyzes the correspondence between dwell time in the fitting room zone and whether the tagged item was subsequently purchased. By correlating fitting room dwell time with point-of-sale data for the same tagged product, the system computes a conversion probability as a function of dwell time. This analysis enables identification of dwell time ranges that correlate with higher or lower purchase likelihood, such as a session with a dwell time of three to five minutes may correlate with a higher conversion rate than a session with a dwell time of less than one minute (which may indicate that the customer quickly decided against the product) or a session with a dwell time exceeding ten minutes (which may indicate indecision or difficulty with fit). Retailers may use this dwell-time-to-purchase correlation to refine staffing decisions (e.g., deploying fitting room attendants when extended dwell times are detected) or to evaluate product design issues (e.g., identifying products with unusually long dwell times but low conversion rates, which may indicate sizing or fit problems).Trial and Conversion Segmentation

[0478] In some embodiments, the system classifies products into trial and conversion segments based on fitting room session frequency and conversion rate. FIG. 21 illustrates an example user interface for trial and conversion segment analysis. The system computes, for each product over a specified time period (e.g., sixty days): the number of fitting room sessions, the number of sessions per product, and the conversion rate.

[0479] Based on these computed metrics, the system may classify each product into one of the following segments: (a) "Champions," which may be products with high trial frequency and high conversion rate; (b) "Untapped Opportunity," which may be products with high trial frequency and lower conversion rate; (c) "Secret Stars," which may be products with lower trial frequency and high conversion rate; (d) "Low Trial and Conversion," which may be products with both lower trial frequency and lower conversion rate; and (e) "Very Low Trial," which may be products with minimal fitting room engagement. The segment boundaries may be configured based on percentile thresholds or absolute thresholds set by a user.

[0480] FIG. 21 shows an example table 2110 of trial and conversion metrics aggregated by segment. For each segment, the table 2110 may show the number of products in thatAttorney Docket No. RADR-026W001segment, the total number of sessions (fitting room visits) for all products in the segment, sessions per product (average trial frequency), conversion rate (percentage of trials resulting in purchase), percentage of total sessions (proportion of all fitting room activity), and percentage of total conversions (proportion of all purchases). A segment definition diagram 2120 illustrates the two-axis classification, with trial frequency categories (Very Low Trial, Low Trial, Medium Trial, High Trial) on the horizontal axis and conversion rate categories (Low Conversion, High Conversion) on the vertical axis, creating a matrix of segments. This segmentation enables a retailer to identify products that may benefit from increased exposure (Secret Stars: high conversion but low trial), products that may need merchandising changes to improve conversion (Untapped Opportunity: high trial but lower conversion), and products that may warrant strategic review (Low Trial and Conversion, and Very Low Trial segments).Consideration Sets and Product Ensembles

[0481] In some embodiments, the system identifies consideration sets by analyzing products tried together within the same fitting room session. A consideration set comprises a group of products within the same category (e.g., multiple jeans) that a customer tries during a single fitting room visit. In contrast, a product ensemble comprises products from different categories tried together during a single fitting room session (e.g., a jean and a shirt). The system computes a consideration set similarity score between product pairs based on cooccurrence frequency within fitting room sessions.Floor Plan Improvement via Product Embeddings

[0482] The inventors have recognized and appreciated the value of improving product placement within a retail environment based on product embeddings derived from customer interaction data collected via RFID tag tracking. In some embodiments, the system generates product embeddings that encode the similarity between products as determined by fitting room co-occurrence (e.g., the probability that two products will appear in the fitting room together during the same customer visit). Products with higher co-occurrence frequency receive higher similarity scores in the embedding space.

[0483] FIG. 22 illustrates three embedding generation approaches used in some embodiments. A first approach 2202 generates Prod2Vec embeddings by passing product identifiers through a projection layer and training with a contrastive loss function on fitting room co-occurrence labels. The Prod2Vec approach 2202 uses embeddings that represent the similarities between products derived from customer interactions, where the embeddings are learned representations that capture which products are considered together by customersAttorney Docket No. RADR-026W001during fitting room sessions. A second approach 2204 generates product description and attribute embeddings by encoding product descriptions and attributes through a pre-trained language model encoder (e.g., a DistilBERT encoder), followed by a fully connected layer and a dropout layer, and training with a contrastive loss function on fitting room co-occurrence labels. A third approach 2206 generates product image embeddings by encoding product images through a convolutional neural network encoder (e.g., a ResNet-50 encoder) and training with a contrastive loss function on fitting room co-occurrence labels. In each approach, the contrastive loss function trains the model to produce embeddings where products that frequently co-occur in fitting room sessions have higher cosine similarity than products that do not co-occur.

[0484] In some embodiments, the system constructs a customer decision tree from the generated product embeddings using a three-step process, as illustrated in the lower portion of FIG. 22. In a first step, the system generates embeddings for all products using one or more of the three approaches described above, producing embedding vectors 2240 that represent each product as a point in a multi-dimensional embedding space. An example embedding vector representation 2240 is shown with numerical values (e.g., "0.124 0.345 0.982 ... 0.387"). In a second step 2220, the system calculates cosine similarity between the embedding vectors of each pair of products, producing a similarity matrix 2250 that quantifies the relationship between all product pairs. An example similarity matrix 2250 is shown with numerical similarity values between 0 and 1 (e.g., "1.000.85 0.72 0.41" for one product's similarities to four products). In a third step 2230, the system builds a hierarchical decision tree 2260 (e.g., a dendrogram) based on the similarity scores, grouping products that customers interact with most similarly and identifying consideration sets at various similarity thresholds. A small dendrogram visualization 2260 illustrates the hierarchical tree structure with branch points connecting products at different similarity levels.

[0485] FIG. 23 illustrates an example customer decision tree user interface. The dendrogram displays products as leaf nodes along the right side, with each product shown with a small product image or icon (not shown in this figure) and product name (e.g., "Product A: High-Waisted Stovepipe Jean," "Product B: Rigid High-Waisted Stovepipe Jean"). Internal branch points in the dendrogram indicate similarity levels, with horizontal branch lines showing where products or groups merge based on similarity scores. A user-adjustable similarity threshold control is shown at the top as a slider with a scale (e.g., 0% to 100%) and as a dashed vertical line through the dendrogram at the selected threshold position (e.g., 50%).Attorney Docket No. RADR-026W001Products that share a common branch point to the right of the threshold line (i.e., merge at a similarity level above the threshold) are grouped into the same consideration set. The interface may display a count of the resulting consideration sets in a prominent box 2310 (e.g., "Consideration Sets: 9"). Each consideration set is visually indicated by a dashed bounding region or bracket around the grouped products. The interface may also display key insights, such as that products branching close to leaf nodes (like Product A) are more unique with fewer substitutes and higher lost-sales risk when out of stock, while products branching further from leaf nodes are frequently tried together and / or considered interchangeable. In some embodiments, products like Products F and G may be interacted with together frequently, and customers commonly consider both when making purchase decisions.

[0486] In some embodiments, the customer decision tree enables identification of product substitution patterns and consideration behavior. Products that branch together at a high similarity level (i.e., close to the leaf nodes in the dendrogram) are highly substitutable, as customers commonly consider both when making purchase decisions and frequently try them together in fitting rooms. Products that branch at a low similarity level (i.e., further from leaf nodes, closer to the root of the dendrogram) are more unique, with fewer substitutes. If such a unique product is out of stock, lost sales are more likely because customers have fewer alternative products to purchase. This information is used to prioritize inventory replenishment and product placement decisions.Non-Fitting-Room Interaction Embeddings

[0487] In some embodiments, the system extends the product embedding generation beyond fitting room co-occurrence data to include non-fitting-room interaction data derived from RFID tag tracking on the sales floor. For example, the system may identify products that are observed to move together in the retail environment — that is, products whose RFID tags exhibit correlated movement patterns (e.g., being picked up and carried together, being placed in the same shopping basket, or being relocated to the same area by customers or staff) even without entering the fitting room zone. These non-fitting-room co-movement signals are used as additional training labels for the embedding models described above, supplementing the fitting room co-occurrence labels. By incorporating both fitting room co-occurrence data and sales-floor co-movement data, the system generates product embeddings that more broadly capture customer interaction patterns, including interactions by customers who do not use the fitting room.Attorney Docket No. RADR-026W001 Floor Plan Improvement Process

[0488] In some embodiments, the system improves product placement within a retail environment based on the consideration sets derived from the customer decision tree. FIG. 24 illustrates the floor plan improvement process in two primary steps and provides a visual comparison between a current layout and an improved layout.

[0489] Step 1 (2410): Projection of Improved Positions. The system projects improved positions for each consideration set onto a store map based on customer interaction patterns, such as high-traffic areas, conversion hot spots, or proximity to fitting rooms. Improved positions are shown in FIG. 24 as colored dots on the store map, with each color representing a different consideration set and a larger center dot marking the centroid of the improved positions for that set. The system calculates the center of improved positions for each consideration set.

[0490] Step 2 (2420): Reservation of Locations. The system reserves available fixture locations for each consideration set, selecting positions closest to the calculated center of improved positions. Reserved positions in FIG. 24 are shown as organized clusters, with each consideration set assigned to a contiguous region of the store map near its improved center.

[0491] FIG. 24 further illustrates a comparison of a current layout 2430 and an improved layout 2435 for the same store floor plan. In the current layout 2430, products from different consideration sets (shown with different marker shapes, like circles, triangles, diamonds, and rotated squares) are interspersed throughout the store. 2440 indicates products from different consideration sets being interspersed in the current layout. In the improved layout 2435, the most similar products (i.e., those within the same consideration set) are located together on the same or adjacent fixtures, and the most similar consideration sets are located in proximity to each other, creating a layout that reflects actual customer shopping behavior. 2442 indicates products from the same consideration set being co-located in the improved layout. Both layouts include store features such as walls, a cashier area, and a fitting room for context.

[0492] In some embodiments, the system incorporates traffic hot spot data when performing floor plan improvement. Products identified as "Champions" in the trial and conversion segmentation (i.e., products with both high trial frequency and high conversion rate, as described above) may be placed at traffic hot spots identified via RFID movement data analysis. Traffic hot spots are locations within the retail environment where RFID tagAttorney Docket No. RADR-026W001movement data indicates higher customer traffic density or higher conversion rates. This placement strategy increases the visibility and accessibility of high-performing products.

[0493] In some embodiments, the system supports cross-merchandising improvement by identifying product ensembles (cross-category product pairs with high co-occurrence in fitting rooms, as shown in the product ensembles table of FIG. 20A) and placing them in proximity on the store floor plan, even if they belong to different departments. For example, jeans and shirts identified as a popular ensemble may be placed on adjacent fixtures or in nearby sections, encouraging customers to try complementary products together. Similarly, products from categories that exhibit high cross-category substitution (e.g., women's shorts and women's jeans, which may be close substitutes for certain customer segments) may be placed in proximity to facilitate customer comparison and reduce lost sales from out-of-stock conditions in either category.

[0494] In some embodiments, when generating the floor plan, the system may compute inter-set similarity scores between pairs of consideration sets in addition to the pairwise product-level similarity scores used to construct the hierarchical grouping structure. An interset similarity score between two consideration sets may be computed as the average pairwise cosine similarity between the embedding vectors of all products in one consideration set and all products in the other consideration set (i.e., the mean of all cross-set product pair similarities). Alternatively, the inter-set similarity may be computed as the cosine similarity between the centroid embedding vectors of the two sets, where the centroid of a consideration set is the mean of the embedding vectors of all products in that set. The system may then use these inter-set similarity scores to assign floor regions: consideration set pairs with higher interset similarity scores are assigned to adjacent or proximate regions of the store floor plan, while consideration set pairs with lower inter-set similarity scores may be assigned to more distant regions. This may ensure that the physical layout of the store reflects not only within-set product similarity (products in the same consideration set placed on the same or adjacent fixtures) but also between-set similarity (consideration sets that are more similar to each other placed in neighboring floor areas).Multi-Department and Multi-Category Improvement

[0495] In some embodiments, the floor plan improvement process operates across multiple departments and multiple categories simultaneously. Rather than improving product placement within a single department in isolation, the system considers interactions across department boundaries — for example, placing a popular jeans-and-shirts ensemble (spanningAttorney Docket No. RADR-026W001two departments) in adjacent areas even though those departments might otherwise be located in separate regions of the store. The system may combine consideration sets from multiple departments into a unified improvement, weighting cross-department ensemble co-occurrence alongside within-department consideration set similarity, to produce a holistic floor plan that reflects actual multi-category shopping behavior.Product Movement and Adjacency Monitoring

[0496] In some embodiments, the system monitors product movement patterns and changes in product adjacency over time following a floor plan change. The system may compare product adjacency relationships before and after a floor plan improvement is applied, detecting whether the improved placement is being maintained or has drifted. For example, if staff restocking or seasonal changes have caused products to migrate from their assigned improved locations, the system may generate alerts or reports indicating that the improved layout is no longer being followed.Example Flowchart of Product Substitution Analysis for Lost Sales Minimization

[0497] In some embodiments, the system performs product substitution analysis to minimize lost sales, as illustrated in FIG. 25. The method begins at step 2510, where the system collects fitting room co-occurrence data for product pairs — that is, data indicating which products were tried together during fitting room sessions over a specified period (e.g., the last 30, 60, or 90 days).

[0498] At step 2520, the system computes a substitution probability for each product pair based on fitting room co-occurrence frequency. Substitution probability may be calculated as the co-occurrence frequency divided by the total number of trials for the products, or using other statistical measures of co-occurrence. A high substitution probability indicates that customers frequently try both products during the same session and consider them interchangeable. For example, the system may determine that two styles of women's jeans have a high substitution probability because customers frequently try both during the same fitting room session before purchasing only one.

[0499] In some embodiments, the system computes two distinct co-occurrence metrics for each product pair: (a) a co-trial-without-co-purchase probability, which may be the probability that both products are brought to the fitting room during the same session but only one (or neither) is purchased, which indicates substitutive behavior; and (b) a co-trial-and-co-purchase probability, which may be the probability that both products are brought to the fittingAttorney Docket No. RADR-026W001room during the same session and both are purchased, which indicates complementary behavior. Products with high co-trial-without-co-purchase probability are classified as substitutes, while products with high co-trial-and-co-purchase probability are classified as complements.

[0500] At step 2530, the system compares the substitution probability for each product pair against a configurable threshold. If the substitution probability is above the threshold, the system classifies the product pair as high substitution (or frequently tried together) at step 2535. If the substitution probability is below the threshold, the system classifies the product as unique or having few substitutes at step 2537.

[0501] At step 2540, the system assesses the lost sales impact for unique (low-substitution) products. The assessment considers factors such as sales velocity, current inventory level, and substitutability score. Products with few substitutes present a higher lost-sales risk when out of stock because customers have fewer alternative products to purchase, resulting in higher likelihood of abandoned transactions. For example, a product that appears at the top of a customer decision tree with no nearby branches (as shown in FIG. 23) may have few or no substitutes, so an out-of-stock condition for that product leads to more lost sales than an out-of-stock condition for a product that has several close substitutes.

[0502] In some embodiments, whether a product is classified as having few substitutes may be determined from the product's position in the hierarchical grouping structure (e.g., the dendrogram illustrated in FIG. 23). Specifically, a product may be classified as having few substitutes when it branches at a low similarity level in the dendrogram (e.g., when the product's nearest branch point in the hierarchical structure merges with another product or group only at a similarity level well below the consideration set threshold, indicating that no other product in the assortment is a close substitute). Conversely, a product may be classified as having many substitutes when it branches at a high similarity level, merging with one or more other products near or above the consideration set threshold. This structural classification based on the hierarchical grouping structure may be complementary to the substitution probability threshold applied at step 2530: a product with a low co-trial-without-co-purchase probability with every other product simultaneously may branch near the leaf nodes of the dendrogram, confirming few substitutes per the hierarchical structure. The replenishment alert at step 2555 may be prioritized for products identified as having few substitutes based on the hierarchical grouping structure.Attorney Docket No. RADR-026W001

[0503] At step 2550, the system determines whether a unique product is currently out of stock or near out-of-stock status. If yes, at step 2555, the system generates a high-priority replenishment alert. At step 2560, the system prioritizes the replenishment queue based on substitutability scores and lost-sales risk scores, with unique products and high-velocity unique products ranked higher for replenishment than high-substitution products. At step 2570, the system monitors and updates substitution patterns over time based on ongoing fitting room cooccurrence data, adapting to seasonal changes and product assortment changes.

[0504] In some embodiments, the system supports test-and-learn experiment workflows for floor plan improvement. A test-and-learn workflow comprises the following phases: (a) a test setup phase, in which the system generates an improved floor plan for a test group of stores (or a test region within a store) while maintaining the current floor plan for a control group; (b) an implementation monitoring phase, in which the system monitors RFID tag data to verify that the improved product placement is being adhered to in the test group and flags deviations; and (c) a measurement phase, in which the system compares key performance indicators (e.g., conversion rate, sales velocity, fitting room sessions, lost sales) between the test group and the control group over a measurement period to assess the impact of the improved floor plan. In some embodiments, KPIs are compared between groups using statistical significance testing (e.g., a two-sample t-test or Mann-Whitney U test on per-store or per-day metric values) to determine whether differences between the test group and the control group are attributable to the floor plan change rather than to random variation. The system may provide dashboards and reports summarizing the results of the test-and-learn experiment, including the computed KPI differences and statistical significance indicators (e.g., p-values or confidence intervals).Neighborhood-Based Exit DetectionExample Flowchart of Enhanced Neighborhood-Based Blind Spot Detection

[0505] The inventors have recognized and appreciated that neighborhood-based analysis of RFID tags can significantly improve the accuracy of exit detection (for example, improving accuracy from approximately 80% to approximately 95-97%). This improvement is achieved by distinguishing between read failures caused by physical blind spots (areas with poor reader coverage) and read failures caused by actual product exit. FIG. 26 illustrates a method of neighborhood-based blind spot detection for RFID tags.Attorney Docket No. RADR-026W001

[0506] At step 2610, the system identifies a neighborhood of RFID tags based on spatial proximity. In some embodiments, the neighborhood is defined using a nearest-neighbors approach — that is, for a given target tag, the system identifies the tags whose most recent detected positions are closest to the target tag's most recent detected position, regardless of absolute distance. In this approach, even if a tag is relatively isolated, the system still identifies whichever tags are nearest as its neighborhood. In alternative embodiments, the neighborhood may be defined as all RFID tags located within a predetermined distance of a given tag (e.g., 2 meters, 3 meters, or 5 meters), or all RFID tags assigned to the same derived neighborhood center (e.g., a cluster centroid computed from tag positions, or tags on the same shelf or fixture). In some embodiments, the system also considers tag density during periods when tags are expected to be stationary (e.g., overnight when the store is closed) to calibrate expected neighborhood density for each region of the store.

[0507] At step 2620, the system determines the time-since-last-read for each tag in the identified neighborhood. This metric represents the elapsed time since an RFID reader last successfully received a response from each tag.

[0508] At step 2630, the system determines whether all tags in the neighborhood have not been read within a threshold period (e.g., the past 7, 14, or 30 days). If all tags in the neighborhood are unread within that threshold, the system proceeds to step 2635, where it classifies the neighborhood as a blind spot (e.g., an area where the RFID reader has poor or no coverage) rather than concluding that the individual tags have exited the retail environment. This distinction is important because a blind spot (e.g., a low shelf that is obstructed from reader antennas) affects all tags in a region simultaneously, whereas tag exit typically affects individual tags independently. In some embodiments, the blind spot classification is stored and associated with that physical region for future tag reads. The blind spot determination may be nearly binary in some cases (i.e., some physical locations are consistently obstructed), but in other cases may exhibit intermittent behavior where the reader occasionally detects tags in the blind spot region at certain times. In some embodiments, stored blind spot classifications are periodically re-evaluated (for example, after reader maintenance, after store fixture changes, or at a configurable interval) so that a region is reclassified as readable if tags within it begin to be detected again, and so that new blind spots caused by changes in the store environment are identified and recorded.

[0509] In some embodiments, the threshold condition at step 2630 may be configurable and not require that all tags in the neighborhood have been unread. Rather, the system mayAttorney Docket No. RADR-026W001compare the number of unread tags against a configurable fraction or count; for example, classifying the neighborhood as a blind spot when a majority of neighboring tags (e.g., more than 50%) have not been read within the threshold period, or when a configurable percentage (e.g., 75% or 90%) exceeds the threshold. A higher required fraction is more conservative, resulting in fewer blind spot classifications and more individual exit determinations; a lower required fraction is more aggressive, resulting in more blind spot classifications and fewer individual exit determinations. In some embodiments, the configurable fraction is set based on the historical reader coverage quality of different regions of the store, such that regions with known intermittent coverage use a lower fraction threshold and regions with reliable coverage use a higher fraction threshold.

[0510] If not all tags in the neighborhood are unread (i.e., at least some tags have been recently read), the system proceeds to step 2640 (the "No" branch). At step 2640, the system analyzes the target tag individually. The system selects the target tag (e.g., the particular tag for which exit probability is being determined).

[0511] At step 2650, the system compares the time-since-last-read for the target tag against the time-since-last-read values for the neighboring tags. This comparison yields a relative read recency measure. For example, if the target tag has not been read for 30 days but the neighboring tags were most recently read 7 days ago, the target tag exhibits a relative read gap that suggests it may have exited the retail environment while its neighbors remain present. In another example, if the target tag was first read one month ago and the surrounding tags were most recently read one day ago, that discrepancy signifies an anomalous pattern warranting closer assessment.

[0512] At step 2655, the system also considers the trajectory of the target tag (e.g., the sequence of detected positions of the target tag over time prior to the last read). In some embodiments, the system maps this trajectory to determine whether the tag's last known movement was toward a known blind spot area. In some embodiments, the trajectory direction is computed as the vector from the median of the tag's earlier detected positions to its most recently detected position; this vector is then compared against the direction toward the centroid of each known blind spot region to determine whether the tag was moving toward that region. A tag whose trajectory leads into a known blind spot is more likely to still be present but unreadable, and therefore receives a downward adjustment to its exit probability compared to a tag whose trajectory leads toward a store exit.Attorney Docket No. RADR-026W001

[0513] At step 2660, the system applies a location-specific trained machine learning model to the combined features. In some embodiments, the system maintains separate trained machine learning models for front-of-house (FOH) and back-of-house (BOH) environments. The FOH model may incorporate cashwrap proximity weighting (e.g., tags detected near a cashwrap or checkout area receive higher exit probability weighting, because proximity to checkout is correlated with purchase or removal from the store). The BOH model may incorporate movement pattern considerations specific to stockroom operations, such as patterns of restocking or receiving. Differential accuracy expectations may apply, with the FOH model generally achieving higher accuracy (e.g., 95-97%) than the BOH model (e.g., 85-90%) due to more consistent reader coverage and clearer signal patterns in sales floor environments compared to stockroom environments.

[0514] At step 2670, the system determines the exit probability for the target tag based on the combined features (time-since-last-read, neighborhood analysis, location-specific features) and the location-specific model. The exit probability is output as a score between 0 and 1, where 1 indicates highest confidence that the tag has exited. At step 2680, the system updates the tag state based on the determined exit probability. For example, tags with exit probability above a threshold (e.g., 0.95) may be set to "lost" state, indicating that the associated product is likely no longer in the retail environment.Decision Tree Classification for Exit Determination

[0515] In some embodiments, the machine learning model used for exit determination at step 2660 comprises a decision tree classifier or an ensemble of decision tree classifiers (e.g., a random forest or gradient-boosted decision tree ensemble). The decision tree classifier receives as input features including time-since-last-read for the target tag, time-since-last-read for each of the neighboring tags, whether the neighborhood is classified as a blind spot, the target tag's distance from its home zone, the target tag's last detected location relative to known blind spot regions, and whether the target tag was last detected near a cashwrap or store exit. The decision tree may be hierarchical, recursive, and converging — that is, the tree evaluates features at each node, branches based on feature values, and converges toward a classification (exited or present) with increasing confidence at each level. Hyperparameters such as the number of trees in an ensemble and the number of features considered at each split may be tuned to balance accuracy and computational cost.

[0516] In some embodiments, the FOH model and the BOH model may be trained separately on distinct datasets and may employ different feature sets, producing differentAttorney Docket No. RADR-026W001learned decision boundaries. The FOH model may be trained on labeled RFID data collected from the sales floor (front-of-house) environment, where features such as cashwrap proximity, proximity to store exits, and sales-floor dwell time are informative. The BOH model may be trained on labeled RFID data collected from the stockroom (back-of-house) environment, where features such as receiving dock proximity, restocking cycle timing, and stockroom zone transitions carry greater predictive weight. Because the two models are trained on different data drawn from operationally distinct environments, they produce different learned thresholds and decision boundaries for classifying a tag as exited versus present.Exit Detection Near Recent Error Zones

[0517] In some embodiments, the system maintains a history of locations in the retail environment where detection errors have recently occurred (referred to herein as "error zones"). When the system determines that a target tag is proximate to a location in the error zone history, the system may reduce the exit probability for that tag as a computed numerical adjustment. The exit probability score may be decreased by a configurable amount (e.g., subtracted by a fixed decrement, or multiplied by a scaling factor less than 1.0) to reflect the increased likelihood that the tag is present but unreadable due to a local detection error rather than having exited the retail environment. For example, if a tag's exit probability as computed from time-since-last-read and neighborhood analysis is 0.80, and that tag is proximate to an error zone, the system may reduce the exit probability to 0.55 before comparing against the exit threshold. In some embodiments, after reducing the exit probability, the system may additionally flag the tag for supplemental review by a user or a downstream process, but the reduction of the exit probability is the primary operative adjustment.Enhanced Availability Analytics

[0518] In some embodiments, the system provides an enhanced availability analytics dashboard, as illustrated in FIG. 27A. The dashboard displays the following metrics, which may be displayed in real time: (a) sales floor availability percentage, which may be the percentage of expected SKUs that have at least one unit available on the sales floor; (b) not-in-store percentage, which may be the percentage of expected SKUs that are not detected anywhere in the store (indicating potential customer purchase, loss, or theft); (c) misplaced percentage, which may be the percentage of SKUs detected in the store but not in their designated home zone (indicating merchandising or restocking activity); and (d) in-stock-room percentage, which may be the percentage of SKUs detected in the back-of-house but not on the sales floor (indicating merchandise awaiting restocking or in receiving).Attorney Docket No. RADR-026W001

[0519] Each metric in FIG. 27A is displayed as a large card with the current value prominently shown (e.g., 84.90% for sales floor availability) along with hour-over-hour, day-over-day, and week-over-week change indicators (with directional arrows ▲ ▼ indicating increases or decreases), enabling a user to identify trends and anomalies at a glance. The dashboard supports viewing at fleet, district, and store levels via navigation tabs, with the selected level indicated by underlining or bold formatting. For fleet and district views, users may select specific stores via filter controls that display the current store identifier (e.g., "Store ID: [002402]"). The large, prominent display of each metric facilitates quick assessment of current inventory health.

[0520] The dashboard further displays a KPIs-by-day-and-hour heat map, as shown in the lower portion of FIG. 27A. The heat map displays availability metrics (sales floor availability, not-in-store percentage, misplaced percentage, or in-stock-room percentage, selectable via tabs with the selected tab underlined or highlighted) by day of week (rows: Monday through Sunday) and hour of day (columns: typically store operating hours, e.g., 10:00 to 22:00), with each cell showing the metric value for that time period (e.g., "90.3%" for Monday at 10:00). This enables identification of temporal patterns, such as a consistent decline in sales floor availability on weekends (e.g., Saturday and Sunday showing lower percentages) due to higher customer traffic and purchasing, or a consistent increase in misplaced percentage during certain hours when restocking or fitting room activity is higher.

[0521] In some embodiments, the dashboard may also include a widespread out-of-stocks view. This view identifies SKUs that are out of stock across a high percentage of stores in the fleet, indicating potential systemic supply chain or demand planning issues. The widespread out-of-stocks table displays columns for SKU identifier (e.g., "0043783408"), product name (e.g., "Everyday Luxe Hoodie"), department (e.g., "Women's Sweatshirts"), category (e.g., "Womens Hoodies"), style identifier (e.g., "493"), size (e.g., "M", "L"), number of stores where the SKU is on the floor set (e.g., 489 stores), number of stores where the SKU is currently out of stock (e.g., 316 stores), and the percentage of stores where the SKU is out of stock (e.g., 65%). Products with high out-of-stock percentages across the fleet (e.g., >50%) are highlighted for immediate attention. This fleet-wide view enables store managers and supply chain teams to identify systemic issues affecting multiple locations and coordinate replenishment or product discontinuation decisions across the fleet.Attorney Docket No. RADR-026W001 Availability KPI Trends

[0522] In some embodiments, the availability analytics dashboard further includes a KPI trends view, as illustrated in FIG. 27B. The KPI trends view displays a time-series line chart 2782 plotting a selected availability metric over a date range (e.g., several months). Selectable metrics include sales floor availability percentage, not-in-store percentage, misplaced percentage, and in-stock-room percentage. The horizontal axis represents dates (e.g., daily granularity) and the vertical axis represents the selected metric value (e.g., percentage ranging from 80.00% to 92.50%). Data points at each date are connected by a trend line 2784, and each data point may be indicated by a marker (e.g., a circle). The KPI trends view is available at fleet, district, and store levels. This time-series visualization enables users to identify long-term availability trends, such as a gradual improvement in sales floor availability over a multi-month period as operational processes improve, or a sudden drop in availability indicating a disruption (such as a system outage or a large-scale inventory event). Such trends may not be apparent from the day-and-hour heat map alone, which captures intra- week patterns rather than longitudinal patterns.Conclusion

[0523] While various inventive embodiments have been described and illustrated herein, those of ordinary skill in the art will readily envision a variety of other means and / or structures for performing the function and / or obtaining the results and / or one or more of the advantages described herein, and each of such variations and / or modifications is deemed to be within the scope of the inventive embodiments described herein. More generally, those skilled in the art will readily appreciate that all parameters, dimensions, materials, and configurations described herein are meant to be exemplary and that the actual parameters, dimensions, materials, and / or configurations will depend upon the specific application or applications for which the inventive teachings is / are used. Those skilled in the art will recognize, or be able to ascertain using no more than routine experimentation, many equivalents to the specific inventive embodiments described herein. It is, therefore, to be understood that the foregoing embodiments are presented by way of example only and that, within the scope of the appended claims and equivalents thereto, inventive embodiments may be practiced otherwise than as specifically described and claimed. Inventive embodiments of the present disclosure are directed to each individual feature, system, article, material, kit, and / or method described herein. In addition, any combination of two or more such features, systems, articles, materials,Attorney Docket No. RADR-026W001kits, and / or methods, if such features, systems, articles, materials, kits, and / or methods are not mutually inconsistent, is included within the inventive scope of the present disclosure.

[0524] Also, various inventive concepts may be embodied as one or more methods, of which an example has been provided. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.

[0525] All definitions, as defined and used herein, should be understood to control over dictionary definitions, definitions in documents incorporated by reference, and / or ordinary meanings of the defined terms.

[0526] The indefinite articles "a" and "an," as used herein in the specification and in the claims, unless clearly indicated to the contrary, should be understood to mean "at least one. "

[0527] The phrase "and / or," as used herein in the specification and in the claims, should be understood to mean "either or both" of the elements so conjoined, i.e., elements that are conjunctively present in some cases and disjunctively present in other cases. Multiple elements listed with "and / or" should be construed in the same fashion, i.e., "one or more" of the elements so conjoined. Other elements may optionally be present other than the elements specifically identified by the "and / or" clause, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, a reference to "A and / or B," when used in conjunction with open-ended language such as "comprising" can refer, in one embodiment, to A only (optionally including elements other than B); in another embodiment, to B only (optionally including elements other than A); in yet another embodiment, to both A and B (optionally including other elements); etc.

[0528] As used herein in the specification and in the claims, "or" should be understood to have the same meaning as "and / or" as defined above. For example, when separating items in a list, "or" or "and / or" shall be interpreted as being inclusive, i.e., the inclusion of at least one, but also including more than one, of a number or list of elements, and, optionally, additional unlisted items. Only terms clearly indicated to the contrary, such as "only one of or "exactly one of," or, when used in the claims, "consisting of," will refer to the inclusion of exactly one element of a number or list of elements. In general, the term "or" as used herein shall only be interpreted as indicating exclusive alternatives (i.e., "one or the other but not both") when preceded by terms of exclusivity, such as "either," "one of," "only one of," orAttorney Docket No. RADR-026W001"exactly one of." "Consisting essentially of," when used in the claims, shall have its ordinary meaning as used in the field of patent law.

[0529] As used herein in the specification and in the claims, the phrase "at least one," in reference to a list of one or more elements, should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but not necessarily including at least one of each and every element specifically listed within the list of elements and not excluding any combinations of elements in the list of elements. This definition also allows that elements may optionally be present other than the elements specifically identified within the list of elements to which the phrase "at least one" refers, whether related or unrelated to those elements specifically identified. Thus, as a non -limiting example, "at least one of A and B" (or, equivalently, "at least one of A or B," or, equivalently "at least one of A and / or B") can refer, in one embodiment, to at least one, optionally including more than one, A, with no B present (and optionally including elements other than B); in another embodiment, to at least one, optionally including more than one, B, with no A present (and optionally including elements other than A); in yet another embodiment, to at least one, optionally including more than one, A, and at least one, optionally including more than one, B (and optionally including other elements); etc.

[0530] In the claims, as well as in the specification above, all transitional phrases such as "comprising," "including," "carrying," "having," "containing," "involving," "holding," "composed of," and the like are to be understood to be open-ended, i.e., to mean including but not limited to. Only the transitional phrases "consisting of' and "consisting essentially of' shall be closed or semi-closed transitional phrases, respectively, as set forth in the United States Patent Office Manual of Patent Examining Procedures, Section 2111.03.

Claims

Attorney Docket No. RADR-026W001CLAIMS1. A method of tracking radio-frequency identification (RFID) tags, each RFID tag storing a unique electronic product code (EPC) encoding attributes of an object to which that RFID tag is attached, the method comprising:transmitting, by an RFID tag reader, a series of interrogation signals to the RFID tags over a measurement period;receiving, by the RFID tag reader, a series of replies to the series of interrogation signals, the series of replies including the EPCs of the RFID tags;estimating, based on the series of replies, respective median locations of the RFID tags over the measurement period;grouping the RFID tags into clusters based on the respective median locations and the EPCs of the RFID tags, each cluster including at least a predetermined number of RFID tags for a corresponding type of object, and each cluster being separated from each other cluster for the corresponding type of object by at least a predetermined distance; and designating each cluster as a home zone for the corresponding type of object.

2. The method of claim 1, wherein grouping the RFID tags into clusters further comprises forming clusters for a given type of object on opposite sides of a wall.

3. The method of claim 2, wherein the wall separates a front-of-house portion of a retail environment from a back-of-house portion of the retail environment, a first one of the clusters for the given type of object is in the front-of-house portion of the retail environment, and a second one of the clusters is in the back-of-house portion of the retail environment.

4. The method of claim 1, further comprising, after the measurement period:transmitting, by the RFID tag reader, an interrogation signal to an RFID tag attached to an object;receiving, by the RFID tag reader, a reply to the interrogation signal from the RFID tag;estimating, based on the reply, a location of the RFID tag;determining, based on a distance between the location of the RFID tag and the home zone for the object, that the object is misplaced; andproviding, to a user, an indication that the object is misplaced.Attorney Docket No. RADR-026W0015. The method of claim 1, wherein at least one type of object has more than one cluster designated as a home zone.

6. A method of tracking an object with a radio-frequency identification (RFID) tag attached thereto in a retail environment, the method comprising:transmitting, by an RFID tag reader, a series of queries to the RFID tag; receiving, by the RFID tag reader, at least one response from the RFID tag to the series of queries;extracting, with a machine learning model from the at least one response, features associated with movement and position of the RFID tag, the features including: (i) a time since a last changed position of the RFID tag, (ii) a distance from the RFID tag to a home zone for the RFID tag, (iii) a number of times the RFID tag was within a first predetermined distance of a checkout of the retail environment, (iv) a distance traveled by the RFID tag within a first predetermined period since a last known position change of the RFID tag, and (v) a number of movements of more than a second predetermined distance within a second predetermined period before the last known position change of the RFID tag; and determining, with the machine learning model based upon at least one of the features, a stock status of the object, the stock status indicating whether the object is available on a sales floor, in a stockroom, or out of stock.

7. The method of claim 6, wherein the features further comprise a time of day of the last known position change of the RFID tag.

8. The method of claim 6, wherein the features further comprise a last known position of the RFID tag.

9. The method of claim 6, wherein the object is an article of clothing and the features further comprise a period since the last known position change of the RFID tag and a last visit of the RFID tag to a fitting room.

10. The method of claim 6, wherein the first predetermined distance is one meter.

11. The method of claim 6, wherein the second predetermined distance is one meter.Attorney Docket No. RADR-026W00112. The method of claim 6, wherein the first predetermined period is five minutes.

13. The method of claim 6, wherein the second predetermined period is five minutes.

14. The method of claim 6, wherein the features further comprise a period between the last known movement of the RFID tag and a most recent event associated with the object.

15. The method of claim 14, wherein the most recent event comprises a sale of the object.

16. The method of claim 6, wherein the features further comprise a number of visits by the RFID tag to the checkout within twenty-four hours of the last known position change of the RFID tag, wherein a visit comprises detection of the RFID tag within the first predetermined distance of the checkout and separated from other detections by at least ninety minutes.

17. The method of claim 6, wherein the features further comprise a distance traveled by the RFID tag within a third predetermined period since the last known position change of the RFID tag.

18. The method of claim 17, wherein the third predetermined period is ten minutes.

19. The method of claim 6, wherein the features further comprise whether the RFID tag was last detected within a predetermined distance of an e-commerce packing area with greater than a threshold probability.

20. The method of claim 19, wherein the threshold probability is 0.5.

21. The method of claim 6, wherein:determining the stock status includes determining a probability that the RFID tag has been removed from the retail environment since the RFID tag reader received the at least one response from the RFID tag.Attorney Docket No. RADR-026W00122. The method of claim 21, further comprising:notifying a user to replace the object attached to the RFID tag in response to the probability exceeding a threshold probability.

23. The method of claim 6, further comprising:extracting, by the RFID tag reader, information from the at least one response; transmitting the information from the RFID tag reader to an interrogator controller; andexecuting the machine learning model on the interrogator controller.

24. The method of claim 23, further comprising:receiving, by the interrogator controller, an indication of an event associated with the RFID tag,wherein determining a probability is based on a time and / or type of the event.

25. The method of claim 24, wherein the event comprises a sale of the object, a theft of the object, and / or a loss of the object.

26. The method of claim 6, further comprising:determining whether to restock the retail environment based on the stock status.

27. A method of training a machine learning model for detecting exits of items tagged with respective radio-frequency identification (RFID) tags from a retail environment, the method comprising:identifying clusters of retail environments having similar characteristics; collecting ground truth data from at least two retail environments from each cluster, including performing morning and evening hand scans to confirm tag presence by verifying whether each tag is physically present without determining precise tag coordinates of the RFID tags in the at least two retail environments across multiple days;collecting RFID tag data from the RFID tags with fixed RFID tag readers between the hand scans;extracting features from the RFID tag data collected between the hand scans; generating exit labels based on the ground truth data; andAttorney Docket No. RADR-026W001training the machine learning model using the extracted features and exit labels.

28. The method of claim 27, wherein the similar characteristics comprise at least one of:store size;store layout;customer traffic pattern;product assortment;sensor positioning; andsensor layout.

29. A method of maintaining RFID tag state accuracy in a retail environment, comprising:detecting, with an RFID tag reader, signals from RFID tags in the retail environment; extracting features from the signals including at least location data representing estimated positions of the RFID tags and timing data representing timestamps of tag detections;determining, using a machine learning model, exit probabilities for the RFID tags based on the extracted features;setting states of RFID tags with exit probabilities above a threshold to lost; identifying RFID tags with exit probabilities below the threshold; andgenerating an audit list comprising the identified RFID tags.

30. A method of processing RFID tag data in a retail environment by an interrogator, comprising:receiving RFID tag read data including tag identifiers, signal characteristics, and read timestamps from multiple RFID tag readers;filtering the RFID tag read data to exclude RFID tag position changes of less than one meter; andusing the filtered tag read data to generate features for training a machine learning model for predicting a state of tags identified by the filtered tag read data.

31. A method of cleaning up stale RFID tag data, comprising:identifying RFID tags without recorded movements for longer than a threshold period;Attomey Docket No. RADR-026W001determining whether identified RFID tags were last detected in a front-of-house location or a back-of-house location;applying an exit probability threshold to the RFID tags based on last detected locations of the RFID tags, wherein a different exit probability threshold applied is for the front-of-house location than the back-of-house location;setting states of tags with exit probabilities above the exit probability threshold to lost; andmonitoring rates of state changes from lost back to available, wherein a state change from lost to available is triggered when a previously lost tag is detected again.

32. A method of validating RFID tag exits, comprising:collecting point-of-sale (POS) events for items in a retail environment;detecting signals from the RFID tags attached to the items;comparing timing between last detected movements of the RFID tags and the POS events, wherein the movements are inferred from changes in position data derived from RFID tag read data;determining exit probabilities of the RFID tags based at least in part on the timing comparisons; andupdating inventory data based on the determined exit probabilities.

33. A method of determining exit status of radio-frequency identification (RFID) tags in a retail environment, the method comprising:identifying, by a processor, a neighborhood of RFID tags based on spatial proximity of most recently detected positions of the RFID tags, the neighborhood comprising a target tag and a plurality of neighboring tags;determining, by the processor, a corresponding time-since-last-read for the target tag and for one or more neighboring tags of the plurality of neighboring tags;determining, by the processor, whether a number of tags in the neighborhood have not been read within a threshold period;in response to determining that the number of tags in the neighborhood have not been read within the threshold period, classifying the neighborhood as a blind spot; andin response to determining that at least one neighboring tag has been read within the threshold period, comparing the time-since-last-read of the target tag to the respective time-Attorney Docket No. RADR-026W001since-last-read values of the one or more neighboring tags and determining an exit probability for the target tag based at least in part on the comparison.

34. The method of claim 33, wherein identifying the neighborhood of RFID tags comprises identifying, for the target tag, whichever RFID tags have most recently detected positions closest to a most recently detected position of the target tag, regardless of absolute distance between the target tag and the neighboring tags.

35. The method of claim 33, further comprising:determining a trajectory of the target tag based on a sequence of detected positions of the target tag prior to a last read of the target tag; andadjusting the exit probability for the target tag based on whether the trajectory leads toward a region previously classified as a blind spot.

36. The method of claim 33, further comprising:applying a first trained machine learning model to determine the exit probability when the target tag is located in a front-of-house portion of the retail environment; and applying a second trained machine learning model, different from the first trained machine learning model, to determine the exit probability when the target tag is located in a back-of-house portion of the retail environment.

37. The method of claim 36, wherein the first trained machine learning model incorporates a cashwrap proximity weighting such that tags detected near a cashwrap receive a higher exit probability weighting.

38. The method of claim 33, further comprising:maintaining a history of locations in the retail environment where detection errors have occurred; andin response to determining that the target tag is proximate to a location in the history, reducing the exit probability for the target tag.

39. A method of improving product placement in a retail environment using radio-frequency identification (RFID) tag data, the method comprising:Attorney Docket No. RADR-026W001collecting, by a processor, co-occurrence data indicating which products tagged with respective RFID tags were detected together within a fitting room zone of the retail environment during respective fitting room sessions over a specified period;generating, by the processor, product embeddings for a plurality of products based on the co-occurrence data, wherein products with higher co-occurrence frequencies receive higher similarity scores in an embedding space;computing, by the processor, pairwise similarity scores between the product embeddings;constructing, by the processor, a hierarchical grouping structure based on the pairwise similarity scores, the hierarchical grouping structure defining consideration sets of products at a selected similarity threshold; andgenerating, by the processor, a floor plan for the retail environment by assigning products within each consideration set to proximate fixture locations and assigning consideration sets with higher inter-set similarity to adjacent regions.

40. The method of claim 39, wherein generating product embeddings comprises at least one ofpassing product identifiers through a projection layer and training with a contrastive loss function using the co-occurrence data as training labels;encoding product descriptions and attributes through a language model encoder and training with a contrastive loss function using the co-occurrence data as training labels; and encoding product images through a convolutional neural network encoder and training with a contrastive loss function using the co-occurrence data as training labels.

41. The method of claim 39, further comprising:identifying traffic hot spots within the retail environment based on RFID tag movement data indicating locations of higher customer traffic density; andassigning products classified as high trial frequency and high conversion rate to fixture locations at the identified traffic hot spots.

42. The method of claim 39, further comprising:computing, for each product pair, a co-trial-without-co-purchase probability indicating a probability that both products are brought to the fitting room zone during a same session and at most one is purchased; andAttorney Docket No. RADR-026W001classifying product pairs with the co-trial-without-co-purchase probability above a substitution threshold as substitutes.

43. The method of claim 42, further comprising:determining that a product classified as having few substitutes based on the hierarchical grouping structure is out of stock or near an out-of-stock status; and generating a replenishment alert for the product.

44. The method of claim 39, further comprising:generating the floor plan for a test group of stores while maintaining a current floor plan for a control group of stores;monitoring RFID tag data in the test group to verify adherence to the improved floor plan; andcomparing key performance indicators between the test group and the control group over a measurement period.