Paper deep dive
AirLog: Store-Level Indoor Life Logging Made Easy
Zihui Yun, Jiaying Du, Yue Yu, Zhewei Liu, Zhen Xiang, Longfei Shangguan, Zhenlin An
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/29/2026, 12:02:04 PM
Summary
AirLog is a smartphone-based system that automatically reconstructs store-level indoor trajectories in shopping malls and generates human-readable life journals. It utilizes ambient Wi-Fi SSIDs as semantic anchors and inertial dead reckoning (PDR) for motion tracking, fused with spatial maps derived from indoor directory images via a geometry-first parser. The system avoids traditional radio-map construction, enabling scalable, survey-free indoor localization and journaling.
Entities (8)
Relation Signals (7)
AirLog → converts → GeoJSON
confidence 95% · AirLog converts directory images into spatial maps... serialized as GeoJSON
AirLog → generates → LLM
confidence 95% · summarized into journals by an LLM
AirLog → uses → Indoor Directory Image
confidence 95% · AirLog repurposes... indoor directory images
AirLog → uses → Wi-Fi SSID
confidence 95% · AirLog repurposes semantic information exposed by ambient Wi-Fi SSIDs
AirLog → evaluatedon → Microsoft Public Indoor Traces
confidence 90% · evaluate it on both a large-scale public dataset... Microsoft public indoor traces
AirLog → fuses → Pedestrian Dead Reckoning (PDR)
confidence 90% · fuses Wi-Fi semantic anchors with inertial dead reckoning
AirLog → usesmodel → GPT-4o
confidence 85% · GPT-4o labels one polygon-masked crop at a time
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:This paper presents AirLog, a smartphone-based life journaling system that automatically reconstructs users' store visits in shopping malls and summarizes them into human-readable journals. Unlike conventional indoor localization systems, AirLog avoids labor-intensive radio-map construction and dedicated wireless localization infrastructure and algorithm calibrations. Instead, it repurposes two cues already available in commercial spaces: semantic information exposed by ambient Wi-Fi SSIDs and indoor directory images. AirLog converts directory images into spatial maps and fuses Wi-Fi semantic anchors with inertial dead reckoning to recover store-level trajectories, which are then summarized into journals by an LLM. Such store-level life logs can support applications such as personal memory recall, activity reflection, and automated diary generation without requiring users to manually record where they have been. We implement AirLog on commodity smartphones and evaluate it on both a large-scale public dataset and a self-collected dataset. The results demonstrate that AirLog substantially improves store-level region recovery, semantic matching, trajectory reconstruction, and journal quality over existing baselines. A human evaluation further shows that the generated journals are coherent and faithful to users' visits.
Tags
Links
- Source: https://arxiv.org/abs/2609.31864v1
- Canonical: https://arxiv.org/abs/2609.31864v1
Trouble viewing inline? Open PDF directly →
Full Text
98,487 characters extracted from source content.
Expand or collapse full text
AirLog: Store-Level Indoor Life Logging Made Easy Zihui Yun University of Georgia Athens, Georgia, USA ayucharon666@gmail.com Jiaying Du University of Georgia Athens, Georgia, USA msz@uga.edu Yue Yu University College London London, United Kingdom ucfnyul@ucl.ac.uk Zhewei Liu University of Toronto Mississauga Mississauga, Ontario, Canada zwei.liu@utoronto.ca Zhen Xiang University of Georgia Athens, Georgia, USA zxiangaa@uga.edu Longfei Shangguan University of Pittsburgh Pittsburgh, Pennsylvania, USA longfei@pitt.edu Zhenlin An ∗ University of Georgia Athens, Georgia, USA zhenlin.an@uga.edu Abstract This paper presentsAirLog, a smartphone-based life jour- naling system that automatically reconstructs users’ store visits in shopping malls and summarizes them into human- readable journals. Unlike conventional indoor localization systems,AirLogavoids labor-intensive radio-map construc- tion and dedicated wireless localization infrastructure and algorithm calibrations. Instead, it repurposes two cues al- ready available in commercial spaces: semantic information exposed by ambient Wi-Fi SSIDs and indoor directory im- ages.AirLogconverts directory images into spatial maps and fuses Wi-Fi semantic anchors with inertial dead reckoning to recover store-level trajectories, which are then summarized into journals by an LLM. Such store-level life logs can sup- port applications such as personal memory recall, activity reflection, and automated diary generation without requiring users to manually record where they have been. We imple- mentAirLogon commodity smartphones and evaluate it on both a large-scale public dataset and a self-collected dataset. The results demonstrate thatAirLogsubstantially improves store-level region recovery, semantic matching, trajectory reconstruction, and journal quality over existing baselines. A human evaluation further shows that the generated journals are coherent and faithful to users’ visits. 1 Introduction Automatically recording daily experiences has long been a goal of mobile sensing and context-aware computing [17,24]. In this paper, we focus on store-level indoor life logging: au- tomatically reconstructing which stores or venues a user visited in a shopping mall, in what order, and for how long, ∗ Corresponding author. and summarizing these visits into a human-readable jour- nal. Such logs can support personal memory recall, visit retrospection, and context-aware assistants. As illustrated in Figure 1, instead of merely recording that a user spent two hours at a mall, a useful log could report that the user visited Macy’s, then Uniqlo, had lunch at the Cheesecake Factory, and later stopped by the Apple Store, together with the temporal order and dwell time of these visits. When users cannot remember the details of a past visit, they can also directly query the life log, e.g., “Where did I go for lunch after visiting Uniqlo?” At first glance, this task appears to be a straightforward ex- tension of existing mobile lifelogging systems where coarse- grained mobility context is often obtained directly from GPS traces, digital maps, and nearby points of interest (POIs), after which an LLM can interpret or summarize the resulting context [45,47,62,63,69]. Indoors, however, the store-level trajectory itself is not readily available, largely because GPS is unreliable while conventional Wi-Fi localization typically requires site-specific surveys, fingerprint databases, or dedi- cated localization pipelines [4,5,25,37,68,71]. Such require- ments make store-level life logging difficult to scale across large and frequently changing commercial spaces. Our key observation is that commercial indoor environ- ments already expose two useful spatial cues: (i): ambient Wi-Fi SSIDs, which often reveal semantic information about nearby stores [30,52,72]; and (i): indoor directory images, which expose region locations and readable labels. This leads to our central question: Can these readily available but weak semantic cues be combined with smartphone motion sensing to recover store-level trajectories without relying on traditional Wi-Fi radio map or localization infrastructures? Answering this question introduces two challenges: arXiv:2609.31864v1 [cs.NI] 25 Sep 2026 Yun et al. Yesterday (Nov 12, 2024) Ross Park Mall What did I do after Macy's? You went to The Cheesecake Factory for lunch after visiting Uniqlo. After Macy's, you visited Uniqlo, then had lunch at The Cheesecake Factory. After I went to Uniqlo yesterday, where did I go for lunch? Wi-Fi IMU Floor-plan A day in the mall (Passive sensing) User carries a smartphone while visiting stores in a shopping mall. Automatically generated life journal Airlog reconstructs store-level visits and summarizes them into a journal. Querying the life log Users can ask retrospective questions about their past visits. Entrance Macy's Browsed clothing and accessories. Uniqlo Looked at jackets and tried. Apple Store Explored products and left the mall. 10:02 - 10:20 (18 min) 10:24~ 10:48 (24 min) 11:15~ 11:58 (43 min) 12:05~ 12:25 (20 min) Cheesecake Factory Had lunch. — More examples — 10:02 AM 10:24 AM 11:15 AM 12:05 PM Total time in mall: ~2 h 23 min Visited 4 stores Figure 1:AirLogturns passive smartphone sensing into a store-level indoor life log. (A) While the user visits a shopping mall, the smartphone passively collects Wi-Fi scans and inertial measurements, and a publicly available floor-plan image is used as spatial context. (B)AirLogreconstructs the store-level trajec- tory—what stores were visited, in what order, and for how long—and summarizes it into a human-readable journal. (C) The life log is queryable: users can ask retrospective questions, such as where they went for lunch after visiting a specific store. (1)public indoor maps are rarely available as machine-readable region layers. While outdoor map services provide struc- tured roads, venue boundaries, and POIs, indoor floor- level information often exists only as directory images rather than labeled region geometry that downstream modules can query [21,22,38,39]. Direct reasoning over these images is unreliable because current vision- language models (VLMs) still struggle with dense map text and fine-grained spatial relationships [61]. (2)Ambient Wi-Fi provides semantic but inherently ambigu- ous location evidence. An SSID may reveal that a partic- ular store is nearby, but it does not uniquely determine whether the user is inside that store, walking past it, or located in an adjacent venue. RSSI further varies with multipath, crowd blockage, device orientation, and het- erogeneous AP placement. Thus, Wi-Fi should be treated as a probabilistic semantic anchor rather than a direct location estimate. In this paper, we proposeAirLog, a smartphone-based system for store-level indoor life journaling. As shown in Fig. 3, given passive smartphone Wi-Fi scans, inertial motion recordings, and a crowd-accessible floor-plan image,AirLog infers which stores a user likely visited, in what order, and for how long, and then summarizes these visits into a life journal. Rather than asking an LLM to infer the entire jour- nal from raw sensor traces and a floor-plan image,AirLog decomposes the task into three structured tasks. •First, the map layer converts a human-facing indoor direc- tory image into a labeled region map. The parser serializes AutoLife (Prior Work) AirLog (Ours) Wild Indoor Commercial Life Context-level Store-level Semantic Trajectory LLM/ VLM Context Fusion General Daily Life Multi-Agent + Map Constraints Location Context Wi-Fi Context Motion Context Wi-Fi IMU Map/POI LLM/VLM Fusion Life Journal Store A (Visit) Corridor (Pass) Store B (Visit) Waiting Area (Wait) Store C (Visit) Wi-Fi Semantic Agent Motion Agent Map Context Agent Review Agent Planning Agent Store-level Trajectory GRANULARITY SCOPE TECHNICAL DESIGN Agentic System GPS / GIS + Online Maps Fine Grained Floor Plan Image Figure 2: From general daily life logging to store-level in- door life journaling. AutoLife captures coarse multimodal context, whileAirLoguses structured indoor maps and map-constrained multi-agent reasoning to reconstruct store-level trajectories. this map as a GeoJSONFeatureCollection 1 , allowing later modules to query region geometry and labels instead of reasoning over raw pixels. The parser itself does not emit corridor polygons, walkable areas, or connectivity. A sepa- rate deterministic map-preparation stage derives the corridor mask and trajectory graph used by the fusion layer. However, indoor directory images are designed for hu- man viewing rather than machine parsing. VLMs can miss dense labels or distort region boundaries, while generic seg- mentation and traditional CV methods are brittle to text, icons, reflections, and perspective distortion [8,55].AirLog therefore uses a geometry-first map parser: GPT Image 2 canonicalizes the visual layout into a clean region mask, de- terministic CV aligns and vectorizes the regions, and GPT-4o labels one polygon-masked crop at a time before writing the returned label set directly to the source GeoJSON feature. •Second, the trajectory layer grounds Wi-Fi and motion ob- servations onto the map to recover store visits. Wi-Fi provides useful semantic hints: an SSID such as Uniqlo_Guest may suggest a nearby Uniqlo. However, the SSID names can be ambiguous, abbreviated, and do not reveal whether the user actually entered the store, walked past it, or visited a neigh- boring store. AirLogtherefore treats SSIDs as probabilistic semantic anchors rather than direct location estimates. It filters and aggregates Wi-Fi scans, derives pedestrian dead reckoning (PDR) motion from inertial measurements, and jointly eval- uates candidate store sequences against map connectivity, walking distance, and motion continuity. The decoded se- mantic trajectory is then reviewed against these tool outputs; 1 GeoJSON is an industry-standard JSON-based format for encoding map features and their geometry [7]. InAirLog, each parser-produced feature represents one region usingPolygonorMultiPolygongeometry and stores its readable strings in properties.labels. AirLog : Store-Level Indoor Life Logging Made Easy detected semantic or physical inconsistencies are fed back into the loop to improve the final trajectory. • Third, the journal layer converts the validated trajectory and motion sensor data into a human-readable account of the visit. A store trajectory alone does not distinguish, for example, spending 25 minutes shopping inside a store from briefly passing its entrance.AirLogtherefore combines dwell time, POI type, motion-derived activity cues, and trajectory confidence to derive a sequence of supported behavior events. An LLM then summarizes this validated sequence into a journal, rather than generating entries directly from raw Wi-Fi or sensor observations. We evaluateAirLogwith two main data sources: large- scale Microsoft public indoor traces containing over 19.7k labeled Wi-Fi scan samples [35], and a self-collected six-city corpus containing 374,567 Wi-Fi observations over 603.0 ac- tive hours. From the latter, we construct 185,140 candidate SSID–POI pairs for semantic analysis. The public traces sup- port quantitative semantic trajectory recovery, while the self-collected data support SSID semantic validation and end-to-end journal generation validation. We testAirLog’s local deployment on edge devices, where sensing data can be collected during the day and processed locally afterward. On-device execution takes longer than cloud, but it preserves privacy by keeping raw data and journals on the user’s de- vice, while smaller models still achieve comparable accuracy. In summary, this paper makes the following contributions: •We proposeAirLog, a novel agentic platform for survey- free, site services-free indoor life journaling in wild envi- ronments using smartphone sensing and crowd-accessible map artifacts. • We design a geometry-first map parser that converts human- facing indoor directory images into labeled region maps serialized as GeoJSON, making region polygons and their readable labels directly queryable by downstream mod- ules. • We introduce a grounded semantic trajectory and jour- naling layer that treats ambient Wi-Fi SSIDs as semantic landmarks, fuses them with PDR and floor-plan topol- ogy through graph-constrained inference, and generates third-person journals from validated visits rather than raw sensor context. 2 Related Works 2.1 Life Journaling Lifelogging systems transform daily observations into records for behavior understanding [6,13,17,24]. Existing approaches broadly use either vision-based or non-visual sensing. Wear- able cameras and smart glasses capture rich egocentric con- text [23,28,32,58,65], but incur high processing costs and privacy concerns, while smartphone-based approaches ex- ploit motion, wireless, audio, and location signals [31,54, 56,70]. Recent LLM-based systems further translate hetero- geneous sensor data into open-vocabulary event descrip- tions [2,11,29,36,45,51,62,63], but largely rely on prompt- level context and target outdoor mobility or coarse activities. As illustrated in Fig. 2, AirLog instead grounds Wi-Fi and mo- tion evidence on indoor map structure to recover store-level semantic trajectories before journal generation. 2.2 Indoor Localization Existing indoor localization methods [1,12,25,48,71] broadly follow fingerprinting- or model-based paradigms. Finger- printing methods require site-specific radio measurements and machine learning [4,5,26,68], while ranging- or AoA- based methods depend on calibrated infrastructure and as- sumptions specific to each deployment [60,71], making both costly to deploy across changing commercial environments. Prior survey-free localization systems avoid manual site sur- veys through crowdsourcing, PDR, map constraints, and opportunistic landmarks, but still primarily target metric coordinates and typically benefit from observations accumu- lated across users or visits [50,53,57,66].AirLoginstead targets site-free semantic localization, inferring visited stores from a single visit without site-specific fingerprints or accu- mulated crowdsourced observations. Wi-Fi SSIDs provide a promising survey-free semantic signal by encoding busi- ness names and location-related tokens [30,52,72];AirLog revalidates this signal at scale and fuses it with PDR and map topology to recover store-level trajectories rather than isolated place labels. 2.3 Map Understanding with VLMs Recent work uses VLMs for map reading, landmark reason- ing, and navigation from floor-plan images [9,33,74], yet current LVLMs struggle to jointly handle OCR, topology, and route planning [61]. Meanwhile, public indoor-map services rarely expose floor-level geometry and topology as queryable layers [21,22,38,39], and recovering queryable regions and labels from human-facing maps remains challenging.AirLog addresses these gaps by converting directory images into labeled GeoJSON region maps for reasoning over explicit polygons and labels. Existing map-parsing pipelines com- bine OCR and classical segmentation [34,55,64], but remain brittle on photographed directories with text, icons, legends, reflections, and perspective distortion;AirLoginstead canon- icalizes region geometry with image generation, vectorizes polygons using deterministic CV, and labels each region with a per-region semantic subagent [16]. 3 A Glimpse of AirLog Imagine Bob visiting Rose Park Mall on a Saturday afternoon. Before entering, he photographs the mall directory. He first enters Uniqlo and browses, then walks past KFC, and later Yun et al. Map Parsing Trajectory Construction Journal Generation Floor Plan (image) Inthelatemorning, theuserentersa shoppingmall... Lifejournal ... ✓Dining (Eating) POI:McDonald's Duration:1,200s ✓Retail (Shopping)... BehaviorInference Floor Plan Generation Semantic Trajectory Journal Generation GeoJson Sensors Readable Map Trajectory Figure 3:AirLogoverview.AirLogdecomposes life logging into three structured layers: (1) Labeled Region Map Construction: a geometry-first parser converts an indoor directory image into a labeled region map serialized as GeoJSON. (2) Trajectory Reconstruc- tion: a trajectory agent grounds sensor data and Wi-Fi semantics using downstream map constraints to infer store-level movement. (3) Journal Generation: a journal agent summarizes the validated trajectory into a third-person life journal. spends about 25 minutes at Starbucks. Bob never checks in or labels these visits; during the trip, his phone records only passive Wi-Fi scans and IMU readings. To reconstruct Bob’s path,AirLogmust first convert this directory photo into a labeled region map (§4). Although readable to Bob, the raw image cannot directly serve as map context for the trajectory agents, which need explicit store geometry and POI labels. The resulting GeoJSON stores la- beled region polygons for Uniqlo, KFC, and Starbucks so later modules can combine the map with Bob’s Wi-Fi and motion observations. As Bob walks, his phone periodically observes SSIDs, in- cludingUniqlo_WiFi,KFC_FREE_WiFi, andStarbucks_WiFi. However, these SSID observations are sometimes ambigu- ous: seeingKFC_FREE_WiFi, for example, does not imply that Bob entered KFC: he may simply be walking past it.AirLog therefore treats Wi-Fi observations as probabilistic seman- tic anchors, uses PDR to estimate how Bob moves between observations, and jointly decodes them under the floor-plan topology. Said differently, the Wi-Fi SSIDs indicate where Bob may be, motion sensing indicates how Bob moved, and the map constrains where Bob can move. These cues together can help distinguish a brief pass near KFC from sustained visits to Uniqlo and Starbucks (§5). Finally,AirLogaggregates the reconstructed trajectory into store-level visits using dwell time and motion cues, and summarizes the validated sequence into a human-readable journal, e.g., “In the afternoon, Bob visited Uniqlo, walked through the dining area, and later spent about 25 minutes at Starbucks.” (§6). 4 Constructing Labeled Region Maps from Floor-Plan Images We begin with the map-representation bottleneck in Bob’s example. His floor-plan image contains the store-level in- formation needed to reconstruct his visit—where Uniqlo, KFC, Starbucks, and other POIs are located—but exposes this information only as pixels. As shown in Figure 4(a)– (b), a public map may identify the mall while providing no queryable floor-level POI geometry, whereas the floor-plan image visually encodes both store regions and their names. For the trajectory reconstruction in the next section, how- ever,AirLogneeds more than an image: it needs an explicit mapping from each POI label to its polygon so that down- stream modules can query where stores are and reason about their spatial relationships. Challenge: recovering both region geometry and readable labels. Turning a human-facing floor-plan im- age into such a representation requires solving two coupled problems. First, region geometry must be recovered from a visually cluttered image. Store boundaries are mixed with text, logos, icons, legends, decorative backgrounds, and weak separators. Second, each recovered region must be associated with the correct readable label set. Recognizing a store name alone is insufficient: the system must know exactly which region that name describes. A mistake in either geometry or semantic association produces an incorrect map and can subsequently mislead trajectory reconstruction. Why conventional and direct approaches fall short. The closest prior work on this problem is the shopping-mall- plan parser of Su et al. [55]. Their pipeline first recognizes a mall directory, applies threshold/edge preprocessing and two-stage region growing to segment individual rooms, and then uses OCR-recognized room identifiers to retrieve the corresponding room names. This work establishes that ex- plicitly coupling segmentation and recognition can recover both geometry and semantics from structured shopping-mall plans. AirLog advances this line of work to more heterogeneous human-facing floor-plan images, where text, logos, icons, leg- ends, reflections, weak separators, and perspective distortion make classical region growing and OCR-based matching brit- tle. Rather than requiring an OCR-readable room identifier that can be matched back to a directory,AirLogcanonical- izes the visual layout into explicit region geometry and then assigns semantics directly to each recovered region with a per-region semantic subagent. Other direct alternatives expose complementary limita- tions. Classical color-filling methods can merge neighboring stores or retain non-POI structures when colors and back- grounds vary, while edge detection can mistake text strokes AirLog : Store-Level Indoor Life Logging Made Easy (a) Public Map(b) Floor-plan regions(GT) 14 12 11 15 8-9 16 17 (c) SAM3 14 (d) Color Filling 14 12 11 15 17 (e) BoundaryDetection 12 11 15 16 14 (c) AirLog(Ours) 14 12 11 15 16 8-9 17 Background wrongly recognized Multiple Stores Merged Icons Wrong Recognized Sensitive to noise Stores missing Sensitive to noise Figure 4: From public maps to labeled region maps. (a) The public map provides venue-level context but lacks floor-level region geometry. (b) The floor-plan image contains the annotated regions and readable labels. (c) SAM 3 absorbs background structures and merges adjacent regions. (d) Color Filling is distracted by icons and other visual structures. (e) Boundary Detection extracts text and noise as candidate boundaries. (f )AirLogmore closely recovers the annotated region geometry and labels. This example is qualitative; aggregate results appear in Table 1. and decorations for region boundaries. Generic segmenta- tion models such as SAM 3 [8] face a different mismatch: they are designed primarily for object-like regions in natural images rather than tenant-scale graphic regions, and can therefore absorb background structures or merge adjacent POIs. Semantic extraction alone does not solve the problem either. Standalone OCR can miss small, stylized, rotated, or low-resolution labels, and even a correctly recognized string does not specify which region polygon it belongs to. A full- map VLM faces a related coupling problem: it must recognize every POI, predict each label’s position, and bind that posi- tion to a region. A correctly recognized name can therefore still be omitted, misplaced, or attached to the wrong polygon. Figure 4 illustrates representative geometry failures. Appen- dix A.4 pairs two exact-region OCR failures in Figure 18 and places the coordinate-binding and geometry-to-semantics mechanisms together in Figure 19, beside the claims that they support. These observations motivate a geometry-first, sequential factorization. Instead of asking one model to jointly recover region boundaries, recognize readable labels, predict label coordinates, and bind labels to regions,AirLogfirst recov- ers explicit region polygons and then uses each polygon to define one constrained semantic task. As shown in Figure 5, the geometry stages in (a)–(d) produce a geometry-only re- gion map serialized as GeoJSON. Each recovered polygon is then projected onto the original floor-plan image to produce the polygon-masked crop in (e). The per-region semantic subagent processes one crop at a time in (f ), and its returned label set is written directly to the same GeoJSON feature (a)Input floor-plan crop (A) Geometry pass (b) Clean region mask (c) CV based Mask Alignment (d) CV based Region polygons (B) Semantic pass (g) Labeled GeoJSON output FeatureCollection "properties": "labels": ["Nordstrom"] "geometry": "type":"Polygon" , ... Apply each polygon as a mask pandora Phone Mania LUNA (f) Semantic subagent VLM (e) Polygon-masked crop Figure 5: Geometry-first map-parsing pipeline. From (a) an input floor-plan crop,AirLoggenerates (b) a clean region mask, performs (c) mask alignment, and extracts (d) region polygons. Each polygon produces (e) a polygon-masked crop for (f ) the per-region semantic subagent. The returned label set is written directly to its source feature, producing (g) a labeled region map in GeoJSON. in (g). The current pipeline therefore predicts no label co- ordinate and requires no separate polygon–label binding stage. 4.1 Geometry Pass: Recovering Regions The geometry pass addresses the first question: where are the regions? Directly extracting boundaries from Figure 5(a) is unreliable because region geometry is entangled with text, logos, icons, and background variation.AirLogtherefore first canonicalizes the map’s visual appearance before ex- tracting precise geometry. Starting from the input crop in (a), AirLoguses GPT Image 2 [43] to generate the clean region mask in (b). The model redraws floor-plan regions as solid regions over a uniform background while suppressing text, icons, legends, and other visual clutter. The generated colors carry no POI semantics; they only make individual spatial regions explicit. This use of image generation is intentional: the generative model handles heterogeneous appearance, while deterministic processing remains responsible for spa- tial precision. The generated mask cannot be used directly because image generation may introduce a small global translation or scale change relative to the source floor-plan image.AirLogthere- fore aligns structural gradients between the generated mask and source image under a restricted transformation consist- ing of uniform scaling and horizontal and vertical transla- tion, producing (c). Rotation and non-rigid deformation are excluded, and the system falls back to the identity trans- formation when reliable alignment cannot be established. Deterministic CV then separates the uniform background from foreground blocks, extracts 4-connected components, traces exterior contours and holes, and converts them into validPolygonorMultiPolygonfeatures. Candidates with insufficient support in the original image are removed. The output in (d) is a geometry-only map that explicitly repre- sents the floor-plan regions but does not yet label them. Yun et al. 4.2 Per-Region Semantic Subagent Once the region polygons are fixed,AirLoguses them to constrain semantic labeling. Here, per-region describes the invocation unit: the subagent processes one recovered region at a time, while the region polygon is used only to construct its input crop. For each polygon, the system rasterizes its exact geometry over the original floor-plan image, retains source pixels inside the polygon, suppresses pixels outside the polygon and inside any holes, and crops the result to the polygon’s bounding box. This produces one polygon-masked crop per GeoJSON feature while preserving the printed POI name and excluding neighboring labels, legends, and deco- rations. The current system invokes this GPT-4o semantic sub- agent once for each polygon-masked crop. Under the shared visible-label policy, the subagent returns"labels":[...] containing all readable strings in the target region. Because each crop is generated from a known polygon, the returned labels are written directly to that same feature. In contrast to the earlier full-map baseline, the current method predicts no label coordinates and requires no point-in-polygon, nearest- neighbor, or learned cross-region binding stage. 4.3 Generate GeoJSON After all recovered regions are labeled,AirLogexports a Geo- JSONFeatureCollection. Each feature containsPolygon orMultiPolygongeometry, along with a canonical label field,properties.labels, containing the recognized strings. The parser does not emitname,category,confidence, cor- ridor polygons, walkable areas, traversability, or connectivity fields. The result is a machine-queryable labeled region map. For Bob’s example, Uniqlo, KFC, and Starbucks become la- beled region polygons that later modules can query directly. A separate deterministic map-preparation stage derives the corridor mask and navigability graph used in the next sec- tion; these downstream artifacts are not part of the parser’s GeoJSON output. 5 Semantic Trajectory Reconstruction Given the labeled region map obtained from §4,AirLognext reconstructs the user’s store-level trajectory from Wi-Fi and inertial sensing. Rather than asking an agent to directly pre- dict locations, inAirLogwe propose a multi-agent pipeline (§5.1) that coordinates semantic Wi-Fi anchoring (§5.2), PDR- based motion estimation (§5.3), and map-constrained infer- ence (§5.4) to construct a trajectory that is both semantically plausible and physically feasible. Observation / Inference Result Feedback for Re-planning Motion Context Map Constraints Location Context Planning Agent (Re-planning & Active Sensing) 1. Wi-Fi Semantic Agent (Location Context) 2. Motion Agent (Motion Context) 3. Map Context Agent (Spatial Constraints) 4. Context Fusion & Review Agent (Review, Reasoning & Fusion) Wi-Fi Scans LLM Reasoning Semantic Location Context IMU Data Motion Estimation Motion Context GeoJSON Map Navigability Graph Review & Understand Viterbi Inference Tool Trajectory Estimate Set Next Goal / Path Collect Data Supporting Data / Tools POI / Store Database Venue Information Time & Calendar Figure 6: Trajectory Reconstruction inAirLog. The framework organizes three complementary contexts for indoor trajectory re- covery: Wi-Fi Semantic Agent derives semantic location context, Motion Agent derives motion context, and Map Context Agent derives GeoJSON map constraints. The Context Fusion & Review Agent integrates these inputs and invokes structured inference to recover a geometrically consistent, store-level trajectory. 5.1 Agent-Orchestrated Trajectory Reconstruction Figure 6 illustrates the agent-orchestrated context-fusion pipeline. Before processing each episode, a Planning Agent selects and schedules the required modules in a ReAct-style loop [67], revising the plan when evidence is missing or con- flicting.AirLogorganizes the resulting evidence into three complementary context streams: Wi-Fi semantic context for location candidates, motion context for relative displacement and heading changes, and map context for physically admis- sible transitions. A structured fusion decoder combines these contexts to recover a globally consistent trajectory, while a review stage detects failures and triggers targeted replanning when necessary. 5.2 Wi-Fi Semantic Agent Validating SSID semantics in the wild. Prior work has shown that Wi-Fi SSIDs can encode location and service information [52]. We re-evaluate their usefulness in cur- rent commercial environments using Wi-Fi observations col- lected across six major cities (Table 6). Associating business- class SSIDs with nearby OpenStreetMap POIs yields 185,140 candidate SSID–POI pairs. As summarized in Table 7, SSID semantics remain observable at scale, but lexical matching alone is too sparse and noisy to serve as a direct signal (repre- sentative failure cases in Table 8). We therefore treat SSIDs as weak semantic evidence and jointly reason over SSID strings, RSSI rank, POI names, and map context. We further verify that SSID distributions are venue- and floor-specific using cross-session observations from a public benchmark [35], spanning over 5.5 hours of data across 2 venues and 14 floors. Simple SSID-set Jaccard matching achieves 87.88% accuracy in identifying the correct site–floor pair, suggesting that AirLog : Store-Level Indoor Life Logging Made Easy SSIDs provide a useful cue for automatic floor identification without site-specific surveys. 5.2.1 SSID Semantic-noise filtering. Although SSID strings provide useful place-level cues, raw Wi-Fi scans contain substantial semantic noise that can mislead the anchoring agent and inflate the LLM context. We make two observations from our traces. • Generic infrastructure SSIDs (e.g., T-mobile, China-mobile) are prevalent but they provide less location information. Many visible SSIDs correspond to carrier hotspots, venue- wide guest networks, or router-default names. These net- works provide little store-specific evidence and would increase LLM input length, so we remove them with a keyword blocklist. •The strongest RSSI is not necessarily the nearest or most useful POI cue. Indoor RSSI is unstable under multipath propagation, wall attenuation, crowd blockage, AP place- ment, and transmit-power variation [25, 71]. Therefore,AirLogdoes not assign location using the strongest SSID alone. After filtering, we aggregate scans into sliding windows, represent each SSID by its maximum RSSI within the window, and pass the top-ranked SSID-RSSI pairs to the LLM as soft semantic evidence. This allows the model to reason over multiple possible store cues instead of relying on a single nearest-signal rule. Empty windows are dropped, and gaps longer than 600 s are flagged to avoid incorrect PDR integration across disconnected sessions. 5.2.2 LLM-based semantic anchoring. After semantic-noise filtering, each Wi-Fi window is represented by a compact set of top-ranked SSID–RSSI pairs.AirLogthen converts this evidence into store-level semantic anchors. We formu- late SSID-to-POI grounding as a language understanding task [30]: the model must interpret noisy SSID strings, abbre- viations, guest-network names, and RSSI ranks in the con- text of the venue’s POI list and floor-plan topology. For each window,AirLogconstructs a structured prompt containing three inputs: the venue’s POI names, map-derived spatial descriptions, and the filtered top-푁SSID–RSSI pairs. The prompt instructs the model to treat RSSI as relative proximity evidence rather than a deterministic distance measurement; the full prompt is shown in Appendix D.2. To provide spatial grounding, we augment each POI with a natural-language descriptor derived from the navigability graph. The descriptor summarizes the POI’s relative position and nearby stores, e.g., “Uniqlo is north of KFC and adjacent to the main corridor entrance,” as illustrated in Figure 7. These relational cues help the model disambiguate SSIDs that are lexically similar or spatially overlapping, and prevent it from treating all name matches as equally plausible. Rather than forcing a single location decision, the model outputs all Filtered Wi-Fi Window System Role: You are a precise positioning assistant. Output ONLY valid JSON. SYSTEM TASK: Identify ALL possible store locations for each timestamp... MAP NODES: [Uniqlo, KFC, Zara, Starbucks, ...] INSTRUCTIONS: 1. List ALL SSIDs that match the Map Nodes. 2. Do NOT limit to one store per timestamp. Prompt --- WIFI WINDOW (T=25.0s) --- 1 KFC_FREE_WiFi : -48 dBm 2 Uniqlo_Store : -54 dBm 3 MALL_FREE_WIFI : -65 dBm 4 TP-Link_AP_98 : -78 dBm 5 ChinaNet_Auto : -82 dBm *... (5 more signals omitted)* Window size: Δw=1s LLM "results": [ "time": 0.5, "locs": [ "name": "Uniqlo", "confidence": 0.9, "name": "KFC", "confidence": 0.4 ] , ... ] Structured JSON Output Spatial Context Description: Uniqlo: North of KFC, adjacent to the main corridor entrance. Zara: Next to Starbucks, across from the information desk. H&M: Between Starbucks and Sephora ... Map Nodes: [Uniqlo, KFC, Zara, Starbucks, H&M, ...] Spatial Context Description & POI Context Figure 7: LLM-based semantic anchoring pipeline. Wi-Fi RSSI observations, spatial context, and POI context are combined into a structured prompt, producing multi-label POI hypotheses with confidence scores. plausible nearby stores with confidence scores between zero and one. The resulting predictions are post-processed into semantic anchors. Each predicted POI name is matched to the corre- sponding GeoJSON feature and mapped to its polygon cen- troid or store region in the navigability graph. Generic out- puts such as floor identifiers, restroom tokens, and corridor labels are discarded. Each surviving tuple(푡,푛,푐), consisting of timestamp푡, store name푛, and confidence푐, becomes a probabilistic claim that the user was near store푛at time푡. These sparse anchors provide semantic location evidence, which complements the continuous but drift-prone PDR es- timate used next. 5.3 Motion Context via PDR Wi-Fi semantic anchors provide place-level cues but are sparse and intermittent, leaving gaps between valid obser- vations.AirLogbridges these gaps with continuous relative motion from the phone’s inertial sensors using a standard PDR pipeline [25,31,59]: step detections and rotation-vector- derived headings are integrated into relative displacement and heading change between consecutive sensing windows. Wi-Fi anchors provide semantic positions, while PDR pro- vides continuous motion between them. The earliest high- confidence anchor can initialize the PDR path when no sur- veyed starting point is available. Residual heading errors are corrected during fusion using well-separated anchors. PDR remains a motion context rather than a store-level trajectory because its errors accumulate over time [37]. Yun et al. 5.4 Context Fusion & Review Agent Prior Wi-Fi IMU floor-plan fusion systems can learn dense location histories by training fusion models over collected traces [27].AirLogtargets a different setting: wild indoor venues where no site-specific fingerprint survey or training trajectory is available. We therefore use a semantic-priority Viterbi decoder that performs explicit probabilistic optimiza- tion over the GeoJSON-derived navigability graphG[15,49], as shown in Figure 8. The decoder combines three scores. Wi- Fi semantic anchors provide the emission term: nodes near a high-confidence candidate store receive higher likelihood. PDR provides the transition term: candidate graph edges are favored when their distance and heading agree with the observed displacement and turn angle. The map provides the hard topological constraint: Viterbi only considers adjacent nodes inG, which restricts the search to physically walkable routes. We summarize the decoded trajectory as ˆ 푆 1:푇 = arg max 푆 1:푇 ∈G ∑︁ 푡 훽퐸 푡 (푆 푡 )+ 훼퐴 푡 (푆 푡−1 ,푆 푡 ) +훾퐷 푡 (푆 푡−1 ,푆 푡 )+퐶 푡 (푆 푡−1 ,푆 푡 ) ,(1) where퐸 푡 is semantic consistency with Wi-Fi anchors,퐴 푡 and퐷 푡 are IMU heading and distance consistency, and퐶 푡 is an anchor-based heading-calibration bonus when con- secutive high-confidence semantic anchors imply a reliable direction. Concretely, for a candidate transition from node 푢to node푣, with edge direction휃(푢,푣), edge lengthℓ(푢,푣), IMU motion primitive(푑 푡 ,Δ휙 푡 ), and active semantic anchors ˆ 퐿 푡 =(푛 푗 ,푐 푗 ), these terms are 퐸 푡 (푣)= ∑︁ (푛 푗 ,푐 푗 )∈ ˆ 퐿 푡 푐 푗 − dist(푣,P 푛 푗 ) 2 2휎 2 ! ,(2) 퐴 푡 (푢,푣)= log max ( cos(휃(푢,푣)−Δ휙 푡 ), 휖 ) ,(3) 퐷 푡 (푢,푣)=− ( 푑 푡 − ℓ(푢,푣) ) 2 .(4) HereP 푛 푗 denotes the polygon or centroid of the POI named by anchor푛 푗 , and휎controls the spatial bandwidth of se- mantic evidence. In short, Wi-Fi estimates where the user is likely to be, IMU estimates how the user moved, and the map defines where the user can move. Review and Feedback Loop: This optimization produces an initial trajectory estimate, whichAirLogthen evaluates using diagnostic signals such as position jumps, stuck seg- ments, and PDR-to-map scale drift. If these indicate a low- quality decode, a deterministic grid search re-runs Viterbi over preset(훼,훽,훾, calib_weight)configurations and keeps the lowest-variance result; if issues persist, the Planning Agent inspects the flagged windows using read-only Wi-Fi, anchor, spatial, and PDR context and selects one bounded cor- rective action, such as re-running semantic anchoring with adjusted parameters or known SSID mappings, reweighting WiFi Semantic (Emission) How well each node matches semantic anchors IMU Motion (Transition)Topological Map (Constraints) Only allow transitions along walkable edges Viterbi finds the most likely path that best explains both WiFi semantics and IMU motion under map constraints Semantic-Priority Viterbi Fusion Most Likely Path How well movement matches distance & heading on map Figure 8: Semantic-priority Viterbi fusion. Wi-Fi semantic an- chors score where the user is likely to be, PDR motion primitives score how the user could move, and the GeoJSON graph restricts the search to physically walkable transitions. The decoder first estimates a globally consistent trajectory, which is then reviewed by the multimodal feedback agent. 14:12-14:57 15:07-15:42 16:00-16:51 17:15-17:41 17:41-18:05 Behavior Inference Life journal "A productive afternoon! After grabbing a latte at Starbucks, The user spent some time at Zara and then headed to Apple to check out the new gadgets..." Reconstructed Trajectory & Map Information Motion Calibration "browsing","passing",... (low linearity, few steps) Candidate Motion POI: Food Court Behavior: Browsing Dwell Time: 3 min (18:30-18:33) Motion Disambiguation LLM PDR/ IMU signals LLM LLM Figure 9: Workflow of Diary Generation. The reconstructed trajectory is aggregated into store-level visits by a behavior infer- ence module (A). A motion calibration module (B) then refines each activity label using PDR geometry and location context. Finally, a diary agent (C) synthesizes the calibrated behavior sequence into a third-person Life Journal. the fusion decoder within a fixed valid range, falling back to substring-based Wi-Fi matching, patching a single behav- ior label, or accepting the result. Only Wi-Fi anchoring and fusion may be revisited, each within a small retry budget, while map parsing and PDR are fixed; the investigation itself is capped by a round limit and a wall-clock budget, after which the best result found is used. The Planning Agent’s de- cision calls use a low sampling temperature, so repeated runs can select different parameter values within these bounded ranges rather than always reaching an identical trajectory. 6 Journal Generation Spatial fusion produces a corrected trajectory, but the re- sult remains a localization sequence rather than a human- readable account. As shown in Fig. 9,AirLogtransforms this sequence into a diary through three stages: behavior AirLog : Store-Level Indoor Life Logging Made Easy inference (A), motion calibration (B), and journal generation (C). 6.1 Behavior Inference As shown in Fig. 9(A), behavior inference converts the fused trajectory into store-level visits. For each POI,AirLog merges consecutive contacts inside or near the store region, producing an enter time, exit time, dwell duration, semantic confidence, and local motion summary. Consecutive con- tacts are merged using a trajectory-adaptive temporal toler- ance, allowing segmentation to adapt to different sampling rates. Short contacts are treated as passing behavior, whereas longer dwell periods form candidate visits. Structural POIs such as corridors, elevators, and restrooms are excluded re- gardless of dwell time. For each candidate visit, the inference agent uses the store name and category, dwell duration, neighboring POIs, and an IMU-derived activity cue to infer the most likely behavior. The motion cue can come from either a conven- tional smartphone activity recognizer or a zero-shot activity model [29,31,54]. The output is a compact behavior se- quence containing the store, time interval, dwell duration, inferred action, and confidence. 6.2 Journal Generation Motion Calibration. As illustrated in Fig. 9(B), motion cali- bration refines candidate activity labels using PDR-derived motion features and location context. Step count, displace- ment, path linearity, and mean speed first produce a candi- date label fromstationary, walking, browsing, transit, which an LLM then disambiguates using the store category, dwell time, and neighboring POIs. Before narration, an evidence- filtering step removes unreliable WiFi anchors. Specifically, a POI’s WiFi evidence is discarded when its active window exceeds a predefined fraction of the visit while scan den- sity remains low, indicating repeated background detections rather than a concentrated visit. This rule overrides even high-confidence WiFi anchors. Journal Generation. As shown in Fig. 9(C), journal gener- ation transforms the validated visits and calibrated actions into a third-person Life Journal. The diary agent first con- structs a structured outline containing the chronological flow, dominant activity, inferred purpose, and salient visit characteristics, and then rewrites the outline as a fluent nar- rative. This two-stage process keeps the journal grounded in validated behaviors while allowing natural transitions and varied phrasing. 7 Implementation Prototype.AirLogis implemented as a Python prototype. A preprocessing module normalizes timestamps, converts Table 1: Map-parsing results on the complete benchmark. Higher is better. The twoAirLogrows share identical region ge- ometry and differ only in semantic-labeling configuration. MethodArea F1↑ Region Purity ↑ Matched Label F1 ↑ Color Filling0.2770.5100.523 Su et al.-style Boundary Detection0.1440.4680.506 SAM 30.0370.3060.155 AirLog w/o semantic subagent0.7050.7670.734 AirLog w/ semantic subagent0.7050.7670.814 raw Wi-Fi records into SSID–RSSI snapshots, resamples IMU streams for PDR, and serializes each session into a struc- tured JSON file. The JSON representation contains Wi-Fi windows, PDR motion primitives, and map context includ- ing POI names, polygon centroids, and graph-based spatial descriptors. To facilitate reproducibility, we will publicly re- lease the implementation and experimental configurations upon publication. Model backends. We evaluateAirLogwith both closed- source and open-weight LLMs. GPT-4o is the default back- bone, while GPT-5.4, Claude Sonnet 4.6, gpt-oss-120b, Gemma 4, Qwen 3, and DeepSeek-R1 are used for cross-model eval- uation [3,40,41,44]. The map-semantic comparison addi- tionally uses GPT-5.5 [42]. Across backbones, preprocessing, graph construction, Viterbi parameters, prompts, and post- processing remain fixed. Prompting and tools.AirLoguses a lightweight ReAct- style agent loop [67], with LLMs handling semantic reason- ing and deterministic Python modules performing validation and spatially constrained trajectory fusion. Agent outputs follow structured JSON schemas for semantic anchoring, behavior inference, and diary generation. Public dataset. We use the sample data from Indoor Loca- tion Competition 2.0 [35], which provides smartphone sensor traces, Wi-Fi scans, floor-plan images, and GeoJSON maps from two large shopping malls in Hangzhou. Site 1 covers five floors with 642 traces totaling 563.8 minutes, while Site 2 covers nine floors with 429 traces totaling 290.8 minutes. Overall, the dataset contains 1,071 traces, 19,712 Wi-Fi scans, and 8,704 surveyor-labeled waypoints. Each trace provides time-synchronized sensor streams and surveyed waypoints with absolute coordinates, enabling quantitative evaluation of zero-shot trajectory reconstruction in multi-floor indoor environments. Self-collected dataset. We collect smartphone traces us- ing Sensor Logger and export the recorded data for offline processing [10]. A total of 603.0 hour data was collected (Table 6). For each experiment, a volunteer starts record- ing before entering the venue and carries the phone natu- rally while performing ordinary activities. Wi-Fi scans are recorded at the application default interval 10 s, and IMU Yun et al. streams are resampled to 50 Hz before PDR processing. Each session is paired with a venue floor plan and a manual refer- ence record of visited stores and dwell intervals. The floor plan is captured by the volunteer as a photo of the mall di- rectory before the session.AirLoguses readable floor-plan images obtained from public venue websites, online map directories, or crowdsourced photos when available. We use the self-collected dataset primarily for Wi-Fi semantics anal- ysis and journaling validation. For journaling evaluation, we recruited 11 volunteers to annotate a subset of the traces in Table 6 with reference journals, yielding a total of 33 h 37 min of recorded time, of which 20 h 55 min correspond to indoor activity. 8 Evaluation 8.1 Map Parsing Evaluation Metrics. We report three complementary scores. (1) Area F1 measures tolerant one-to-one region recovery, capturing whether annotated regions are successfully recovered with- out requiring exact boundaries. (2) Region Purity measures the fraction of predicted area overlapping annotated fore- ground, reflecting how well predictions avoid background and other clutter. (3) Matched Label F1 evaluates semantic label accuracy after geometry-based matching between pre- dicted and ground-truth regions. Detailed matching rules, averaging procedures, and label normalization are provided in Appendix A.1. Baselines. We compare three visual pipelines and twoAirLog semantic-labeling configurations. (1) Color Filling groups connected regions by color and binds OCR labels by spa- tial overlap. (2) Boundary Detection / Su et al. represents the classical edge/region-growing approach in mall-plan pars- ing [55], with OCR applied separately to candidate regions; because no official implementation is used, this is our imple- mentation of the procedure described in the paper. (3) SAM 3 [8] uses the fixed text-prompted configuration documented in Appendix A.2, followed by full-crop OCR with the same spatial binding strategy as Color Filling. We also report a uniform threshold-0.5 SAM 3 sensitivity rerun in the appen- dix. Since these methods differ in both region extraction and label binding, Table 1 presents an end-to-end pipeline com- parison rather than a controlled comparison under a shared semantic-labeling module. The twoAirLogconfigurations use identical frozen pre- dicted polygons and the same visible-label policy. Full-map uses GPT-5.5 to jointly predict labels and coordinates for the entire map before binding them to polygons, whereas the current configuration invokes GPT-4o on each polygon- masked crop and writes the returned label list directly to the corresponding GeoJSON feature. Table 2: Semantic labeling comparison. The GPT-4o per-region run contains 1,920 completed semantic records from 1,927 applica- tion attempts; all seven extra attempts were recovered by retries and no record was excluded. Model InputCalls Tokens (I/O) Time (s) Label F1 GPT-5.5 Full map24/24 127.4k/137.0k1,984.00.734 GPT-4o Region crop 1,920/1,927 524.2k/24.6k1,815.9 0.814 Setting and map complexity. The complete benchmark contains 24 floor-plan images spanning both public web- published mall directories and in-the-wild photographs. Twenty correspond to publicly discoverable malls with official web- sites and public directory/floor-plan images; the remaining four were photographed during our in-the-wild sensing col- lection and may contain realistic perspective distortion, il- lumination variation, and color shift. Region geometry was manually maintained in QGIS against the exact pixel-aligned crop. GPT-5.5 was used only as an auxiliary label-auditing tool, and all final ground-truth labels were manually verified against the source images. All method–image evaluations completed under evaluator v8; the archived benchmark snap- shot and runtime details are documented in Appendix A.2. Figure 10 summarizes the density of OCR text relative to geometry-valid annotated regions, characterizing visual clut- ter rather than recognition accuracy. Region recovery.AirLogsubstantially outperforms the vi- sual baselines in both Area F1 and Region Purity, indicating more complete region recovery and cleaner foreground lo- calization. Area F1 increases from 0.277 for Color Filling to 0.705 forAirLog, while Region Purity increases from 0.510 to 0.767. These scores reflect tolerant region matching and foreground coverage, respectively, rather than exact pixel overlap. The conclusion is unchanged under standard IoU- based metrics:AirLogobtains 0.590 Region F1@0.5 and 0.475 PQ@0.5, compared with 0.229 and 0.199 for Color Filling; Appendix A.1 reports the full sanity check. Semantic-labeling comparison. Table 2 compares the two production semantic-labeling configurations under identical frozen region geometry and the same visible-label policy. Calls denotes completed semantic records/application at- tempts, Tokens reports aggregate input/output tokens, and Time is the summed model-response latency rather than wall-clock batch time. Label F1 is the Matched Label F1 de- fined above. Because the backend model and input scope change together, this is an operational comparison rather than a controlled ablation; we do not attribute the difference independently to model choice or region cropping. Semantic labeling. With region geometry and visible-label policy fixed, the GPT-4o region-crop configuration reaches 0.814 Matched Label F1 (precision 0.789, recall 0.842), com- pared with 0.734 for full-map GPT-5.5 (precision 0.728, recall AirLog : Store-Level Indoor Life Logging Made Easy 50100150 Number 0.0 0.5 1.0 CDF Polygons Labels Figure 10: Floor-plan image complexity. (a) Ground truth(b) AirLog (Ours)(c) SenLLM(d) AutoLife Figure 11: Spatial distribution of subset hit/miss results across localization methods. Green dots indicate waypoints where the GT location falls within the predicted subset; red dots indicate misses. RuleSenLLMAutoLifeOurs Different Methods 0 20 40 60 80 100 Subset Hit (%) Subset Hit Avg Hit Ratio RuleSenLLMAutoLifeOurs Different Methods 0 20 40 60 80 100 Recall (%) Recall Avg SeqSim 0 20 40 60 80 100 Avg Hit Ratio (%) 0.0 0.2 0.4 0.6 0.8 1.0 Avg SeqSim (a) Subset Hit & Avg Hit Ratio(b) Recall & Avg SeqSim Figure 12: Trajectory reconstruction evaluation. 0.740). The 0.080 difference describes the two deployed con- figurations only because model and input scope change to- gether. The GPT-4o run completed all 1,920 semantic records; seven transient failed attempts across six polygons were re- covered within the retry budget, so no sample was excluded from evaluation. 8.2Store-Level Trajectory & Visit Recovery Next, we evaluate how accuratelyAirLogreconstructs a user’s trajectory and recovers the stores visited along the trajectory. Metrics. We evaluate recovery at two complementary lev- els. For point-level trajectory recovery, (1) Subset Hit mea- sures whether the predicted set covers all ground-truth POIs, while (2) Avg Hit Ratio measures the fraction of ground-truth POIs covered, using the five nearest POIs and an 8 m spatial threshold with a±5 s temporal window. These metrics are evaluated on the public benchmark with surveyed ground truth (§7). For store-level visit recovery, we further evaluate self-collected diary cases (§7), where (3) Recall measures recovered ground-truth visits and (4) SeqSim measures or- dering consistency between predicted and ground-truth POI sequences using normalized edit distance. Baselines. We compareAirLogagainst three baselines. (1) SenLLM, inspired by Penetrative AI [62], is a pure LLM rea- soning baseline that receives the same formatted sensing context and infers the POI sequence from motion and Wi- Fi data. (2) AutoLife [63] is adapted to indoor settings by using the provided floor-plan image as VLM location con- text instead of a GPS-queried map. (3) Rule-based Trajectory Reconstruction assigns each timestamp to the POI with the strongest nearby Wi-Fi signal. Results. Figure 12 shows thatAirLogoutperforms all base- lines, improving Subset Hit and Avg Hit Ratio over AutoLife by 24.8% and 24.6%, respectively. It also achieves 84.6% Recall Figure 13: Comparison of generated journals.Underlined italic denotes ground-truth POI names. Red denotes hallucinated or in- correct content. More details in the Appendix C. and 0.778 SeqSim, compared with 63.6% and 0.327 for Au- toLife. The limited gain of AutoLife over SenLLM suggests that floor-plan images alone are insufficient for reliable in- door trajectory reasoning, whereasAirLogcombines map topology, Wi-Fi semantics, and motion continuity for more accurate and physically consistent reconstruction; remaining errors mainly occur in anchor-sparse corridors and dense areas with overlapping Wi-Fi semantics (Fig. 11). 8.3 Journalling Recovery Metrics. To evaluate the quality of generated journals, we measure their similarity to reference journals using chrF [46] and BERTScore [73]. We also report hallucination rate, which is marked as hallucinated if the response contains content that is not supported by evidence and is factually inconsistent with the target context. Settings: We recruit 11 volunteers to collect indoor shopping traces. Each participant carries a smartphone during natural shopping activities while the device records Wi-Fi scans and IMU data. After each session, the participant writes a concise reference journal describing the stores visited, dwell periods, and main activities. Results. Table 3 shows thatAirLogachieves the best per- formance across all metrics. It reduces the hallucination rate from 0.667 for SenLLM and 0.636 for AutoLife to 0.000, while improving chrF by 26.9% and 10.2%, and BERTScore F1 by 53.3% and 23.7%, respectively. Figure 13 further shows that the baselines introduce unsupported stores, whereasAirLog Yun et al. Table 3: Diary generation quality evaluation. Hall. Rate indi- cates the hallucination rate, chrF denotes the character F-score, and P, R, and F1 refer to the precision, recall, and F1 score of BERTScore. Method Hall. Rate↓ chrF↑ BERTScore↑ PRF1 SenLLM0.6670.3310.2630.4030.330 AutoLife0.6360.3810.3630.4580.409 AirLog0.0000.4200.4850.5290.506 Table 4: Token usage and nominal estimated cost per ses- sion. Estimates use GPT-4o-equivalent rates of $2.50/1M input and $10.00/1M output tokens; they are not gateway invoices. ModuleInput Output Cost (USD)Calls Spatial description22.3k1.2k6.8×10 −2 /hr1 per floor Wi-Fi semantic agent118.7k13.2k4.3×10 −1 /hr푁 푤 Trajectory validation62.8k26.9k4.3×10 −1 /hr1 per session Behavior inference1.9k0.8k1.3×10 −2 /hr1 per session Journal generation3.0k1.3k2.1×10 −2 /hr1 per session Total per session208.7k43.4k9.6×10 −1 /hrAll modules generates journals grounded in the recovered POI visit se- quence. 8.4 Ablation Study Figure 14 shows the contribution of each component to store- level recovery. Removing the Wi-Fi Semantic Agent causes the largest degradation, reducing Subset Hit and Avg Hit Ra- tio from 71.2%/75.5% to 38.3%/42.3%, and Recall from 84.6% to 36.0%. This highlights semantic Wi-Fi evidence as the key source for identifying visited stores. Removing the Map Agent or Context Fusion also substantially reduces point- level recovery, with Avg Hit Ratio dropping to 56.0% and 55.2%, respectively. The Motion Agent mainly improves visit- level consistency, while the Planning Agent has the small- est effect on point-level metrics. Overall, the results show that semantic anchoring provides the strongest contribution, with map, motion, and fusion providing complementary con- straints. 8.5 Cross-Model Evaluation We evaluateAirLogwith different LLM backbones to as- sess its robustness to model choice. Figure 15 shows that trajectory recovery remains relatively stable across back- bones, while diary generation is more sensitive to model choice. GPT-5.4 achieves the strongest overall trajectory per- formance (73.9% Subset Hit, 77.3% Avg Hit Ratio, 83.3% Recall, and 0.833 SeqSim), while Qwen3 (32B) achieves the highest Recall (91.7%) but a substantially higher hallucination rate (0.313). This gap highlights that high visit recall alone does not guarantee reliable diary generation. Overall,AirLogre- mains effective across both proprietary and open-weight backbones, demonstrating robustness to the underlying LLM. 8.6 Human Evaluation Study Setting. Following AutoLife [63], we conduct a user study to evaluate generated journals along five dimensions: clarity, conciseness, correctness, completeness, and relevance. Six volunteers rated 11 generated diaries. All dimensions are rated on a 1–4 scale, where 1 indicates that the generated journal completely does not meet the criterion and 4 indi- cates that it completely meets the criterion. Fleiss’휅is used to measure inter-rater agreement. For self-evaluation, users assessed the end-to-end system using their own collected data. Results. Figure 16 shows that the generated journals are generally perceived as clear, relevant, and faithful to the col- lected data. Evaluations from third-party raters are consistent with the authors’ self-assessment, and raters show substan- tial agreement in their judgments. Correctness emerges as a particular strength, while conciseness remains an area for improvement. 8.7 Token Cost We report nominal API-equivalent cost by logging input and output tokens for each LLM/VLM call and applying the nominal GPT-4o rates stated in Table 4. Requests in our experiments were routed through a custom third-party API- compatible gateway, and no billing invoice or verified gate- way pricing was available; the reported dollar values are therefore estimates rather than actual paid cost. Processing one hour of trace consumes 208.7k input tokens and 43.4k output tokens, corresponding to $0.96 per hour. In other words, the token cost is roughly $1 per hour of processed trajectory. The cost is dominated by Wi-Fi semantic anchor- ing and trajectory validation, while behavior inference and journal generation add only a small fraction of the total. 8.8 Mobile and Edge Deployment As location information is privacy sensitive, we evaluate the local deployment cost ofAirLogin terms of sensing overhead and on-device processing. Sensing System Cost. During daytime collection, the phone only records passive Wi-Fi scans, IMU samples, timestamps, and optional venue context using Sensor Logger [10] on commodity phones (Xiaomi Flipmix2, vivo S12, and iQoo12). Across the tested phones, a 1-hour trace consumes 153 mAh (3% battery) and produces 70 MB of local data on average. These measurements indicate modest overhead for hour- scale passive sensing. Running on edge device. We run our agent with Gemma 4 12B [20] locally with LM Studio [14] on a Mac mini equipped with an M4 chip and 24 GB memory. Processing a 1-hour daytime trace takes roughly 75 min end-to-end, making the local setup suitable for deferred trajectory analysis and di- ary generation rather than interactive use. This runtime is AirLog : Store-Level Indoor Life Logging Made Easy Subset HitAvg. HitRecallSeqSim 0.0 0.2 0.4 0.6 0.8 Full Ours w/o Wi-Fi Semantic w/o Motion w/o Map w/o Context Fusion w/o Planning Figure 14: Ablation study. Subset HitAvg. HitRecallSeqSim 0.0 0.2 0.4 0.6 0.8 1.0 DeepSeek-R1 (70B) Gemma 4 (4B) Qwen3 (32B) GPT-OSS (120B) Claude Sonnet 4.6 GPT 5.4 (a) Trajectory Reconstruction Hall. RatechrFBS-PBS-RBS-F1 0.0 0.1 0.2 0.3 0.4 0.5 GPT-OSS (120B) DeepSeek-R1 (70B) GPT 5.4 Claude Sonnet 4.6 Qwen3 (32B) Gemma (latest) (b) Diary Generation Figure 15: Cross-model robustness evaluation. 2 2.5 3 3.5 4 Clarity Conciseness CorrectnessCompleteness Relevance Third-PartySelf-EvalFleiss' κ Figure 16: User study results. therefore suitable for deferred journal generation rather than interactive use during the trip. The map can be parsed once offline and reused across visits to the same venue, so map construction is not included in the per-trace runtime. We further note thatAirLogcan also be deployed on mobile de- vices through the Google AI Edge Gallery framework [18,19]; however, phone-side execution is expected to take longer and is mainly an additional deployment option. 9 Conclusion We have presented the design and implementation ofAirLog, an agentic mobile system that converts a floor plan intended for human interpretation into a machine-readable spatial structure to which ambient Wi-Fi can be anchored, making store-level indoor logging possible without dedicated infras- tructure. We believe this points to a broader direction: agents that interpret the spatial artifacts already present in built environments and ground everyday signals against them. Acknowledgments This work was supported in part by the National Science Foundation under award No. 2554332, No. 2337537, No. 2433914, and No. 2302724. References [1]Abdulrahman Alarifi et al.2016. Ultra wideband indoor positioning technologies: Analysis and recent advances. Sensors (2016). [2] Tuo An, Yunjiao Zhou, Han Zou, and Jianfei Yang. 2025. IoT-LLM: En- hancing Real-World IoT Task Reasoning with Large Language Models. arXiv:2410.02429 [cs.AI] https://arxiv.org/abs/2410.02429 [3] Anthropic. 2026. Claude Sonnet 4.6. https://w.anthropic.com/ claude/sonnet. Accessed: 2026-05-28. [4]Paramvir Bahl and Venkata N. Padmanabhan. 2000. RADAR: An in- building RF-based user location and tracking system. IEEE INFOCOM (2000). [5]Nayan Sanjay Bhatia and Katia Obraczka. 2025. Transforming Decoder- Only Transformers for Accurate WiFi-Telemetry Based Indoor Local- ization. arXiv:2505.15835 [cs.NI] https://arxiv.org/abs/2505.15835 [6] Marc Bolaños, Mariella Dimiccoli, and Petia Radeva. 2016. Toward Storytelling from Visual Lifelogging: An Overview. In Proceedings of the 2016 ACM on International Conference on Multimedia Retrieval. [7]Howard Butler, Martin Daly, Allan Doyle, Sean Gillies, Stefan Hagen, and Tim Schaub. 2016. The GeoJSON Format. RFC 7946. Internet Engineering Task Force. https://doi.org/10.17487/RFC7946 [8]Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman Rädle, Triantafyllos Afouras, Effrosyni Mavroudi, Katherine Xu, Tsung-Han Wu, Yu Zhou, Liliane Momeni, Rishi Hazra, Shuan- grui Ding, Sagar Vaze, Francois Porcher, Feng Li, Siyuan Li, Aish- warya Kamath, Ho Kei Cheng, Piotr Dollár, Nikhila Ravi, Kate Saenko, Pengchuan Zhang, and Christoph Feichtenhofer. 2025. SAM 3: Seg- ment Anything with Concepts. https://doi.org/10.48550/arXiv.2511. 16719 arXiv:2511.16719 [cs.CV] [9] Kehan Chen, Yan Huang, Dong An, Jiawei He, Yifei Su, Jing Liu, Ni- anfeng Liu, and Liang Wang. 2026. FloorPlan-VLN: A New Paradigm for Floor Plan Guided Vision-Language Navigation. arXiv preprint arXiv:2603.17437 (2026). [10] Kelvin Choi. 2026. Sensor Logger. https://w.tszheichoi.com/ sensorlogger. Accessed: 2026-05-28. [11] Hongwei Cui, Yuyang Du, Qun Yang, Yulin Shao, and Soung Chang Liew. 2025. LLMind: Orchestrating AI and IoT with LLM for Complex Task Execution. IEEE Communications Magazine 63, 4 (2025), 214–220. https://doi.org/10.1109/MCOM.002.2400106 [12] Decawave. 2017. Decawave UWB positioning system overview. [13]Aiden R. Doherty et al.2013. Passively recognising human activities through lifelogging. Computers in Human Behavior 29, 4 (2013), S40– S48. [14]Element Labs, Inc. 2026. LM Studio: Local AI on your computer. https://lmstudio.ai/. Accessed: 2026-06-05. [15] G. David Forney. 1973. The Viterbi algorithm. Proc. IEEE (1973). [16] Valentin Gabeur, Shangbang Long, Songyou Peng, Paul Voigtlaender, Shuyang Sun, Yanan Bao, Karen Truong, Zhicheng Wang, Wenlei Zhou, Jonathan T. Barron, Kyle Genova, Nithish Kannen, Sherry Ben, Yan- dong Li, Mandy Guo, Suhas Yogin, Yiming Gu, Huizhong Chen, Oliver Wang, Saining Xie, Howard Zhou, Kaiming He, Thomas Funkhouser, Jean-Baptiste Alayrac, and Radu Soricut. 2026. Image Generators are Generalist Vision Learners. arXiv preprint arXiv:2604.20329 (2026). [17] Jim Gemmell, Gordon Bell, and Roger Lueder. 2006. MyLifeBits: Ful- filling the Memex Vision. In Proceedings of the 10th ACM International Conference on Multimedia. [18]Google AI Edge. 2026. AI Edge Gallery Agent Skills. https://github. com/google-ai-edge/gallery/tree/main/skills. Accessed: 2026-06-05. [19]Google AI Edge Team. 2026. Bring State-of-the-Art Agentic Skills to the Edge with Gemma 4. https://developers.googleblog.com/bring-state- of-the-art-agentic-skills-to-the-edge-with-gemma-4/.Accessed: 2026-06-05. [20]Google DeepMind. 2026. Gemma 4 model card. https://ai.google.dev/ gemma/docs/core/model_card_4. Accessed: 2026-06-05. [21]Google Maps. n.d..Indoor Maps – About.Online documenta- tion.https://w.google.com/maps/about/partners/indoormaps/ Accessed: 2026-05-24. Yun et al. [22]Google Maps Help. n.d.. Use indoor maps to view floor plans. Online support documentation. https://support.google.com/maps/answer/ 2803784 Accessed: 2026-05-24. [23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Vincent Car- tillier, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Ham- burger, Hao Jiang, Devansh Kukreja, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Ilija Radosavovic, Santhosh Kumar Ra- makrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina González, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jáchym Kolář, Satwik Kottur, Anurag Kumar, Fed- erico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Man- galam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Rui- jie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Gio- vanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Kr- ishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. 2025. Ego4D: Around the World in 3,600 Hours of Egocentric Video. IEEE Transactions on Pattern Analysis and Machine Intelligence 47, 11 (2025), 9468–9509. https://doi.org/10.1109/TPAMI.2024.3381075 [24] Cathal Gurrin, Alan F Smeaton, and Aiden R Doherty. 2014. Lifel- ogging: Personal Big Data. Foundations and Trends in Information Retrieval 8, 1 (2014), 1–125. [25]Robert Harle. 2013. A Survey of Indoor Inertial Positioning Systems for Pedestrians. IEEE Communications Surveys & Tutorials (2013). [26]Suining He and S-H Gary Chan. 2016. Wi-Fi fingerprint-based indoor positioning: Recent advances and comparisons. IEEE Communications Surveys & Tutorials (2016). [27]Sachini Herath, Saghar Irandoust, Bowen Chen, Yiming Qian, Pyojin Kim, and Yasutaka Furukawa. 2021. Fusion-DHL: WiFi, IMU, and Floor- plan Fusion for Dense History of Locations in Indoor Environments. In 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 5677–5683. [28]Steve Hodges, Lyndsay Williams, Emma Berry, Shahram Izadi, James Srinivasan, Alex Butler, Gavin Smyth, Narinder Kapur, and Ken Wood. 2006. SenseCam: A Retrospective Memory Aid. In Proceedings of the 8th International Conference on Ubiquitous Computing (UbiComp). Springer, 177–193. [29]Sijie Ji, Xinzhe Zheng, and Chenshu Wu. 2024. HARGPT: Are LLMs Zero-Shot Human Activity Recognizers?. In 2024 IEEE International Workshop on Foundation Models for Cyber-Physical Systems & Internet of Things (FMSys). 38–43. https://doi.org/10.1109/FMSys62467.2024. 00011 [30]Martin Korelič, Octavian Machidon, and Veljko Pejović. 2025. SELLMA: Semantic Location through On-Device LLMs and WiFi Sensing. In Proceedings of the 8th International Workshop on Edge Systems, Ana- lytics and Networking (World Trade Center, Rotterdam, Netherlands) (EdgeSys ’25). Association for Computing Machinery, New York, NY, USA, 7–12. https://doi.org/10.1145/3721888.3722091 [31]Jennifer R Kwapisz, Gary M Weiss, and Samuel A Moore. 2011. Activity Recognition using Cell Phone Accelerometers. In KDD. [32]Yong Jae Lee et al.2008. Wearable context-aware systems for lifelog- ging: A survey. In ISWC. [33]Jiaxin Li, Weiqi Huang, Zan Wang, Wei Liang, Huijun Di, and Feng Liu. 2025. FloNa: Floor Plan Guided Embodied Visual Navigation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 14610–14618. [34]Chen Liu, Jiajun Wu, Pushmeet Kohli, and Yasutaka Furukawa. 2017. Raster-to-Vector: Revisiting Floorplan Transformation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV). 2195– 2203. [35]Location Competition. 2020. Indoor Location Competition 2.0 Sam- ple Data and Code. https://github.com/location-competition/indoor- location-competition-20. Accessed: 2026-05-28. [36]Subigya Nepal, Arvind Pillai, Talie Massachi, Eunsol Soul Choi, et al. 2024. Contextual AI Journaling: Integrating LLM and Time Series Behavioral Sensing Technology to Promote Self-Reflection and Well- being using the MindScape App. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems. 1–8. [37] Zheng Ni et al. 2022. Experience: Indoor Localization in the Wild. In Proceedings of the 18th International Conference on Emerging Network- ing Experiments and Technologies (CoNEXT). [38]Open Geospatial Consortium. 2021. Indoor Mapping Data Format (IMDF) Community Standard, Version 1.0.0. Technical Report 20- 094. Open Geospatial Consortium. https://docs.ogc.org/cs/20-094/ Reference/index.html Accessed: 2026-05-24. [39]Open Geospatial Consortium. 2023. IndoorGML Standard – Indoor Spatial and Navigation Data Standard. OGC standards page. https: //w.ogc.org/standards/indoorgml/ Accessed: 2026-05-24. [40]OpenAI. 2025.Introducing gpt-oss.https://openai.com/index/ introducing-gpt-oss/. Accessed: 2026-05-28. [41] OpenAI. 2026. GPT-4o Model. https://platform.openai.com/docs/ models/gpt-4o. Accessed: 2026-05-28. [42]OpenAI. 2026. GPT-5.5 Model. https://developers.openai.com/api/ docs/models/gpt-5.5. Accessed: 2026-08-24. [43]OpenAI. 2026. Introducing ChatGPT Images 2.0. https://openai.com/ index/introducing-chatgpt-images-2-0/. Accessed: 2026-06-05. [44]OpenAI. 2026.Introducing GPT-5.4.https://openai.com/index/ introducing-gpt-5-4/. Accessed: 2026-05-28. [45]Xiaomin Ouyang and Mani Srivastava. 2024. LLMSense: Harnessing LLMs for High-level Reasoning Over Spatiotemporal Sensor Traces. arXiv preprint arXiv:2403.19857 (2024). [46]Maja Popović. 2015. chrF: character n-gram F-score for automatic MT evaluation. In Proceedings of the Tenth Workshop on Statistical Machine Translation. 392–395. [47]Kevin Post, Reo Kuchida, Mayowa Olapade, Zhigang Yin, Petteri Nurmi, and Huber Flores. 2025. ContextLLM: Meaningful Context Reasoning from Multi-Sensor and Multi-Device Data Using LLMs. In Proceedings of the 26th International Workshop on Mobile Computing Systems and Applications (HOTMOBILE ’25). ACM, New York, NY, USA, 13–18. https://doi.org/10.1145/3708468.3711892 [48] Jiageng Qiao et al.2024. Advancements in Indoor Precision Posi- tioning: A Comprehensive Survey of UWB and Wi-Fi RTT. Network (2024). [49] Lawrence Rabiner. 1989. A tutorial on hidden Markov models. [50] Anshul Rai, Krishna Kant Chintalapudi, Venkata N. Padmanabhan, and Rijurekha Sen. 2012. Zee: Zero-Effort Crowdsourcing for Indoor Localization. In Proceedings of the 18th Annual International Conference on Mobile Computing and Networking (MobiCom ’12). Association for Computing Machinery, New York, NY, USA, 293–304.https: //doi.org/10.1145/2348543.2348580 [51]Zhiwei Ren, Junbo Li, Minjia Zhang, Di Wang, Xiaoran Fan, and Longfei Shangguan. 2025. Toward Sensor-In-the-Loop LLM Agent: Benchmarks and Implications. Association for Computing Machin- ery, New York, NY, USA, 254–267. https://doi.org/10.1145/3715014. 3722082 AirLog : Store-Level Indoor Life Logging Made Easy [52]Suranga Seneviratne, Fangzhou Jiang, Mathieu Cunche, and Aruna Seneviratne. 2015. SSIDs in the wild: Extracting semantic informa- tion from WiFi SSIDs. In 2015 IEEE 40th conference on local computer networks (LCN). IEEE, 494–497. [53]Guobin Shen, Zhuo Chen, Peichao Zhang, Thomas Moscibroda, and Yongguang Zhang. 2013. Walkie-Markie: Indoor Pathway Mapping Made Easy. In 10th USENIX Symposium on Networked Systems Design and Implementation (NSDI ’13). USENIX Association, Lombard, IL, USA, 85–98. https://w.usenix.org/conference/nsdi13/technical- sessions/presentation/shen [54]Muhammad Shoaib et al.2014. Fusion of Smartphone Motion Sensors for Physical Activity Recognition. In Sensors. [55]Ming Su, Wei Shi, Dangjun Zhao, Dongyang Cheng, and Junchao Zhang. 2022. A High-Precision Method for Segmentation and Recog- nition of Shopping Mall Plans. Sensors 22, 7 (2022), 2510.https: //doi.org/10.3390/s22072510 [56] Ke Sun, Chunyu Xia, Xinyu Zhang, Hao Chen, and Charlie Jianzhong Zhang. 2024. Multimodal Daily-Life Logging in Free-living Environ- ment Using Non-Visual Egocentric Sensors on a Smartphone. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 8, 1, Article 17 (March 2024), 32 pages. https://doi.org/10.1145/3643553 [57]He Wang, Souvik Sen, Ahmed Elgohary, Moustafa Farid, Moustafa Youssef, and Romit Roy Choudhury. 2012. No Need to War-Drive: Un- supervised Indoor Localization. In Proceedings of the 10th International Conference on Mobile Systems, Applications, and Services (MobiSys ’12). Association for Computing Machinery, New York, NY, USA, 197–210. https://doi.org/10.1145/2307636.2307655 [58] Shihao Wang, Guo Chen, De-an Huang, Zhiqi Li, Minghan Li, Guilin Liu, Jose M. Alvarez, Lei Zhang, and Zhiding Yu. 2025. VideoITG: Mul- timodal Video Understanding with Instructed Temporal Grounding. arXiv preprint arXiv:2507.13353 (2025). [59] Harvey Weinberg. 2002. Using the ADXL202 in pedometer and per- sonal navigation applications. In Analog Devices AN-602. [60]Zhao Wenda, Goudar Abhishek, Qiao Xinyuan, and Angela P. Schoellig. 2024. UTIL: An Ultra-wideband Time-difference-of-arrival Indoor Localization Dataset. The International Journal of Robotics Research 43, 10 (2024), 1443–1456. [61]Shuo Xing, Shuangyu Xie, Zezhou Sun, Kaiyuan Chen, Yanjia Huang, Yuping Wang, Jiachen Li, Dezhen Song, and Zhengzhong Tu. 2025. Can Large Vision Language Models Read Maps Like a Human? arXiv preprint arXiv:2503.14607 (2025). [62]Huatao Xu, Liying Han, Qirui Yang, Mo Li, and Mani Srivastava. 2024. Penetrative ai: Making llms comprehend the physical world. In Proceed- ings of the 25th International Workshop on Mobile Computing Systems and Applications. 1–7. [63]Huatao Xu, Zilin Zeng, Panrong Tong, Mo Li, and Mani B. Srivastava. 2025. AutoLife: Automatic Life Journaling with Smartphones and LLMs. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 9, 4 (2025), 1–29. https://doi.org/10.1145/ 3770683 [64]Chen Yang, Zhenghao Tang, Yasutaka Furukawa, et al.2019. FloorNet: A Unified Framework for Floorplan Reconstruction from 3D Scans. arXiv preprint arXiv:1903.04394 (2019). [65] Jingkang Yang, Shuai Liu, Hongming Guo, Yuhao Dong, Xiamengwei Zhang, Sicheng Zhang, Pengyun Wang, Zitang Zhou, Binzhu Xie, Ziyue Wang, Bei Ouyang, Zhengyu Lin, Marco Cominelli, Zhongang Cai, Bo Li, Yuanhan Zhang, Peiyuan Zhang, Fangzhou Hong, Joerg Widmer, Francesco Gringoli, Lei Yang, and Ziwei Liu. 2025. EgoLife: Towards Egocentric Life Assistant. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 28885–28900. https: //doi.org/10.1109/CVPR52734.2025.02690 [66]Zheng Yang, Chenshu Wu, and Yunhao Liu. 2012. Locating in Fin- gerprint Space: Wireless Indoor Localization with Little Human In- tervention. In Proceedings of the 18th Annual International Confer- ence on Mobile Computing and Networking (MobiCom ’12). Asso- ciation for Computing Machinery, New York, NY, USA, 269–280. https://doi.org/10.1145/2348543.2348578 [67]Shunyu Yao et al.2023. ReAct: Synergizing Reasoning and Acting in Language Models. arXiv preprint arXiv:2210.03629 (2023). [68] Moustafa Youssef and Ashok Agrawala. 2005. The Horus WLAN location determination system. In MobiSys. [69]Xiaofan Yu, Lanxiang Hu, Benjamin Reichman, Dylan Chu, Rushil Chandrupatla, Xiyuan Zhang, Larry Heck, and Tajana S. Rosing. 2025. SensorChat: Answering Qualitative and Quantitative Questions during Long-term Multimodal Sensor Interactions. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 9, 3, Article 148 (2025), 35 pages. https://doi.org/10.1145/3749496 [70] Ozgur Yurur et al.2014. Context-awareness for mobile sensing: A survey and future directions. IEEE Communications Surveys & Tutorials (2014). [71]Faheem Zafari, Athanasios Gkelias, and Kin K. Leung. 2019. A survey of indoor localization systems and technologies. IEEE Communications Surveys & Tutorials (2019). [72]Haopeng Zhang, Yili Ren, Haohan Yuan, Jingzhe Zhang, and Yitong Shen. 2025. Wi-Chat: Large Language Model Powered Wi-Fi Sensing. arXiv:2502.12421 [cs.CL] https://arxiv.org/abs/2502.12421 [73] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. 2020. BERTScore: Evaluating Text Generation with BERT. In International Conference on Learning Representations. [74]Kaiwen Zhou, Kaizhi Zheng, Connor Pryor, Yilin Shen, Hongxia Jin, Lise Getoor, and Xin Eric Wang. 2023. NavGPT: Explicit Reasoning in Vision-and-Language Navigation with Large Language Models. arXiv preprint arXiv:2305.16986 (2023). Yun et al. A Map Parsing Evaluation and Failure Analysis This appendix expands the map-parsing claims made in Sec- tion 4 and the results reported in Section 8.1. The main text gives the headline comparison; here we explain why the three reported metrics are appropriate, how to interpret the corresponding scores, and which qualitative cases support the design choices and limitations discussed in the method. A.1 Metric Design and Interpretation Table 1 in the main text evaluates two questions: whether AirLog recovers each floor-plan region as a usable polygon, and whether the recovered polygon receives the correct read- able labels. We therefore report two geometry metrics and one semantic metric. Area F1: did we recover the regions? For an annotated region푔and a predicted region푝, we first compute the two directed area coverages 푐 푔 (푔,푝)= Area(푔∩ 푝) Area(푔) , 푐 푝 (푔,푝)= Area(푔∩ 푝) Area(푝) . (5) A pair is eligible when max(푐 푔 ,푐 푝 ) ≥ 0.5,min(푐 푔 ,푐 푝 ) ≥ 0.05,(6) after which we form a label-independent one-to-one match- ing. Let푇푃 퐴 be the matched pairs,퐹푃 퐴 the unmatched pre- dictions, and퐹푁 퐴 the unmatched annotations. We compute 푃 퐴 = 푇푃 퐴 푇푃 퐴 + 퐹푃 퐴 , 푅 퐴 = 푇푃 퐴 푇푃 퐴 + 퐹푁 퐴 , AreaF1= 2푃 퐴 푅 퐴 푃 퐴 + 푅 퐴 . (7) This tolerant criterion reflects the downstream require- ment of AirLog: later modules need one queryable polygon for each floor-plan region, while small boundary deviations are less important than missing, splitting, or duplicating a region. Thus, the increase from 0.277 to 0.705 in Table 1 indicates substantially more successful one-to-one region recovery; it should not be interpreted as 70.5% pixel overlap. For reproducibility, this paper-facing Area F1 corresponds to the association_f1 field in evaluator v8. Region Purity: did a polygon absorb unrelated content? LetΩ GT be the union of all evaluator-valid annotated regions. For each predicted polygon 푝, we compute Purity(푝)= Area(푝∩Ω GT ) Area(푝) .(8) Predictions receive equal weight within each image, and image-level means are macro-averaged. Region Purity com- plements Area F1 because a prediction may recover the in- tended region while also absorbing corridor, neighboring- store, icon, or background pixels. The increase from 0.510 to 0.767 in Table 1 therefore indicates cleaner region predictions with less unrelated content. Matched Label F1: are the labels correct once geometry is fixed? Semantic evaluation first forms a separate label- independent one-to-one geometry matching at IoU≥0.5. Label strings are canonicalized identically for all methods and compared as sets. With pooled semantic true positives, false positives, and false negatives denoted by푇푃 퐿 ,퐹푃 퐿 , and 퐹푁 퐿 , we report 푃 퐿 = 푇푃 퐿 푇푃 퐿 + 퐹푃 퐿 , 푅 퐿 = 푇푃 퐿 푇푃 퐿 + 퐹푁 퐿 , LabelF1= 2푃 퐿 푅 퐿 푃 퐿 + 푅 퐿 . (9) Unmatched polygons are left to the geometry metrics, so geometry failures are not counted again as semantic errors. Under frozen geometry, the deployed GPT-4o region-crop configuration obtains 0.814 Matched Label F1, compared with 0.734 for the GPT-5.5 full-map configuration in Table 2. Because the backend model and visual scope change together, this comparison does not isolate the effect of region cropping or model choice. Standard-metric sanity check. Area F1 uses a task-motivated tolerant matching rule. To verify that the geometry conclu- sion does not depend on this rule, we additionally evalu- ate the same frozen predictions using standard one-to-one matching at IoU≥0.5. Table 5 reports Region F1@0.5 and PQ@0.5. The method ordering remains unchanged under both metrics. Table 5: Standard-metric sanity check on map geometry. Re- gion F1@0.5 and PQ@0.5 use one-to-one matching at IoU≥0.5 and are macro-averaged across the 24 maps. MethodRegion F1@0.5↑ PQ@0.5↑ Color Filling0.2290.199 Boundary Detection / Su et al.0.0950.075 SAM 30.0200.016 AirLog0.5900.475 A.2 Benchmark and Comparison Protocol The benchmark used by Section 8.1 contains 24 floor-plan images with 1,488 geometry-valid annotated polygons and 1,400 label strings; 190 polygons contain multiple readable labels. Twenty images correspond to publicly discoverable AirLog : Store-Level Indoor Life Logging Made Easy malls with official websites and publicly available directory or floor-plan images. The remaining four are mall-directory photographs captured during the same in-the-wild sensing collection used elsewhere inAirLog. These photographs may contain perspective distortion, illumination variation, and color shift that are less pronounced in web-published maps. The benchmark therefore covers both web-published floor- plan artifacts and photographed mall directories encountered in the target deployment setting. The multi-label cases motivate evaluating label sets rather than forcing every region to contain a single store-name string. Ground-truth polygon geometry was manually main- tained in QGIS against the exact pixel-aligned source image. GPT-5.5 was used only as an auxiliary tool for label audit- ing; every final ground-truth label was manually reviewed against the source floor-plan image. All reported values use the same archived evaluator-v8 snapshot and prediction artifacts under OpenCV 4.10.0. Baseline- specific settings were fixed globally and applied uniformly across the 24 maps. We did not tune parameters separately for individual maps or select settings against the benchmark ground truth. Baseline configurations. Color Filling uses gray-world normalization, OKLab clustering with푘=6, merge threshold 0.04, random seed 13, and a 30-pixel minimum connected- component area. PaddleOCR (lang=ch) is applied to the full map, and OCR results are bound to predicted regions by maximum overlap with center fallback. Boundary Detection / Su et al. is our implementation of the edge/region-growing procedure described by Su et al., rather than an official author-code reproduction. It applies the fixed threshold/edge pipeline at the original resolution, retains regions of at least 30 pixels, runs OCR on regions of at least 1,200 pixels among at most the 80 largest candidates, and retries OCR once after a 90 ◦ rotation when necessary. For SAM 3, the configuration reported in Table 1 uses the text promptsmap regionandenclosed spatial region, longest-side preprocessing at 2,048 pixels, and a mask-binarization threshold of 0.5. Separately, the wrapper’s omitted object- query score threshold resolves to a detection-confidence threshold of 0.0 rather than the processor default of 0.5. The configuration uses up to 500 masks, no overlap NMS, and exports each residual connected component as a polygon. This permissive configuration produces 24,939 predicted polygons for 1,482 annotated polygons, and 68.05% of the predicted polygons have no evaluator-raster overlap with an- notated foreground. This large amount of over-segmentation explains the unusually low 0.037 Area F1 reported for SAM 3. SAM 3 configuration sensitivity. To test whether the low SAM 3 result is primarily an artifact of the permissive con- figuration above, we reran all 24 maps using the same check- point, text prompts, resize cap, and OCR pipeline, while restoring the detection-confidence threshold to 0.5 and re- taining one residual component per query. This more con- servative configuration removes the large false-positive frag- mentation, but produces predictions on only one of the 24 maps. With empty-prediction maps assigned zero, it obtains Region F1@0.5 of 0.042 and PQ@0.5 of 0.033. Thus, the sen- sitivity check does not indicate robust SAM 3 performance on this benchmark; instead, the more conservative configu- ration suppresses nearly all detections. Table 1 is an end-to-end map-parsing comparison: the vi- sual baselines may differ in both region extraction and label binding. By contrast, Table 2 freezes the predicted polygons and visible-label policy, then compares the full-map GPT-5.5 labeling configuration with the GPT-4o per-region semantic subagent. Because model and input scope still change to- gether, this table compares the two deployed configurations but does not independently isolate either factor. For the GPT-4o per-region run, all 1,920 semantic records were completed from 1,927 application attempts. Seven tran- sient failed attempts affected six polygons, and all six reached a valid terminal response within the retry budget. No seman- tic record was dropped from evaluation. A.3 Downstream Map Preparation Section 4 ends with a labeled region map serialized as a GeoJSONFeatureCollection. Corridor geometry and con- nectivity are constructed afterward, before the trajectory stage in Section 5. Figure 17 visualizes this interface: the parser-produced regions are converted into a complemen- tary corridor mask, region targets are placed in the same metric space, and the corridor is skeletonized into the navi- gability graph used by the Viterbi decoder. This separation also clarifies the scope of Table 1: Area F1, Region Purity, and Matched Label F1 evaluate the parser’s region geometry and labels, rather than the downstream corridor or graph construction. Yun et al. (a) Corridor mask and region tar- gets. (b) Navigability graph. Figure 17: Interface between map parsing and trajectory fu- sion. The labeled GeoJSON region map produced in Section 4 is converted by a separate deterministic stage into the corridor mask and navigability graph consumed in Section 5. These artifacts are not parser outputs and are not included in the map-parsing metrics. A.4 Qualitative Evidence and Remaining Failures Tables 1 and 2 report the aggregate geometry and label re- sults. The examples below illustrate specific failure mecha- nisms discussed in Section 4; they are qualitative evidence rather than additional quantitative comparisons. Exact-region OCR errors. Figure 18 fixes the target geom- etry and gives PaddleOCR only the exact region crop. The readable vertical HMV label is returned as AWH, while O2 is returned as 02. Region isolation therefore removes spatial- assignment ambiguity but does not, by itself, resolve OCR orientation or glyph confusion. (a) Orientation GROUND TRUTH HMV PADDLEOCR AWH (b) Glyph confusion GT O2 → OCR 02 O → 0 Figure 18: Exact-region OCR failures. PaddleOCR reads (a) a visible vertical HMV crop as AWH and (b) O2 as 02. Full-map coordinate-binding error. Figure 19(a) provides an illustrative cached full-map failure case. GPT-5.5 recog- nizes VOGUE, yet its predicted coordinate falls inside the adjacent UBERFONE polygon. Per-region labeling removes this additional coordinate-prediction and polygon-binding step. This example illustrates the failure mechanism and is not used for the quantitative comparison. Geometry-to-semantics cascade. Figure 19(b) illustrates a remaining limitation of the factorized pipeline. For one matched Crew-associated component, the predicted polygon excludes the readable word. The resulting polygon-masked crop therefore removes the visual evidence needed by the semantic subagent. The adjacent Crew component, whose polygon retains the text, is labeled correctly. anchor text outside (a) Correct text, wrong polygon anchor in UBERFONE exact region VOGUE no binding step (b) Geometry removes label evidence predicted polygon masked crop labels = [] Figure 19: Two map-parsing interface failures. (a) A cached full-map GPT-5.5 result recognizes VOGUE but places its coordinate in the adjacent UBERFONE polygon; an exact-region input removes this binding step. (b) For one Crew-associated component, the pre- dicted polygon removes the readable word from the masked crop, leaving no readable evidence for the semantic subagent. B Wi-Fi Semantic Evidence and Collection Context This appendix groups the evidence supporting the use of ambient Wi-Fi as a weak semantic cue: collection coverage, the sparsity and failure modes of lexical SSID–POI matching, and the ambiguity of RSSI-based nearest-signal decisions. B.1 Collection Coverage Figure 20 summarizes representative urban and session-level coverage. These figures document collection context; store- level reconstruction remains grounded in indoor Wi-Fi ob- servations, PDR motion continuity, and structured indoor topology. B.2 Lexical Similarity and Representative Failure Cases The character-level distribution shows a clear non-zero tail, indicating that many SSIDs still contain partial POI-related strings. However, only a small fraction of pairs reach strong lexical overlap: 647 pairs exceed 0.8 at character level with a precision of 52.2%, and 14,516 additional pairs fall between 0.6 and 0.8. Among word-level and TF-IDF pairs, high-similarity matches (>0.8) achieve up to 100% and 90.0% precision re- spectively, but account for only 11 and 40 pairs out of 185,140, AirLog : Store-Level Indoor Life Logging Made Easy (a) Shenzhen density.(b) Shanghai density.(c) Hong Kong session context. Figure 20: Representative Wi-Fi collection context. The density maps summarize two urban collection areas, while the Hong Kong panel provides a session-level view of mobility coverage. Table 6: Dataset statistics across six cities. MetricHong Kong Beijing Shanghai Shenzhen Dalian Nanjing Total Active hours (h)59.35.8208.4225.52.6101.3603.0 GPS points157,3602,137277,356319,8656,696168,041931,455 Wi-Fi observations260,3942,71125,79845,80110,39129,472374,567 Unique SSIDs (raw)21,3039841,7302,1251,6041,27429,020 Unique BSSIDs (raw)56,6151,7422,5413,1992,8321,92268,851 Avg. SSIDs per scan81.046.73.05.948.86.259.4 Table 7: Number of SSID–POI pairs by similarity range and precision. Precision is computed over manually annotated same- store labels. Range Cosine (Char) Cosine (Word) Cosine (TF-IDF) 0.09,731184,841184,871 0.0–0.233,45433 0.2–0.472,31813165 0.4–0.654,474137 (34.3%)100 (9.0%) 0.6–0.814,51617 (58.8%)61 (44.3%) 0.8–1.0647 (52.2%)11 (100%)40 (90.0%) Total185,140185,140185,140 showing that exact token matches are extremely sparse in practice. Table 8 illustrates a complementary failure mode: surface string similarity can rank misleading SSID–POI pairs highly even when the semantic explanation points elsewhere. These examples justify using semantic reasoning over names, local context, and map constraints; they are not presented as a standalone quantitative benchmark. B.3 RSSI Ambiguity Figure 21 expands the Wi-Fi anchoring example from the main paper. Nearby access points overlap substantially along a single walk, and one signal window can remain compatible with multiple neighboring stores. This motivates treating the top-퐾SSIDs as soft evidence rather than selecting the strongest signal as the user’s location. Table 8: Representative bad cases of Wi-Fi SSID–POI match- ing. String similarity algorithms fail where LLM semantic reasoning succeeds. Wi-Fi SSIDPOIBest Sim. SKH-SchoolDeviceSKH Holy Carpenter Church0.693 BeautyA Boutea0.833 MNSGUESTMonster Sushi0.820 푎 SKH-SchoolDevice: SKH is a dominant institutional prefix of the Hong Kong Sheng Kung Hui; “SchoolDevice” identifies a campus device network, pointing to an affiliated lodging facility rather than the church itself. 푏 Beauty: “A Boutea” is a beverage brand whose name superficially resembles “Beauty” in spelling; the SSID more plausibly belongs to a cosmetics or personal-care venue. 푐 MNSGUEST : MNS follows a typical three-letter property code convention; combined with its location in the Miramar cluster in Tsim Sha Tsui, the network is better attributed to Miramar Shopping Centre’s guest Wi-Fi than to a restaurant abbreviation. C Additional Journal Examples The following examples supplement the journal comparison in the main paper.Underlined italicdenotes ground-truth POI names, and red denotes hallucinated or incorrect con- tent. Yun et al. t = 0 entrance t = 60 KFC t = 120 Uniqlo t = 180 atrium t = 240 Starbucks t = 300 H&M zone t = 360 rebound Time (s) -90 dBm -80 dBm -70 dBm -60 dBm -50 dBm -40 dBm RSSI (dBm) ambiguity zonesignal rebound KFC_FREE_WiFi Uniqlo_Store Zara_Guest Starbucks_WiFi in top-10 (LLM visible) out of top-10 (LLM invisible) (a) RSSI variation. Solid lines indicate SSIDs in the top-퐾 window; faded dashed lines indicate lower-ranked sig- nals. (b) Spatial layout. The corresponding mall trajec- tory remains compatible with multiple nearby stores. Figure 21: RSSI ambiguity and spatial layout. The profiles and map show why semantic anchoring is formulated as soft multi-label inference rather than a deterministic nearest-signal decision. Figure 22: Journal generation case 2: restaurant visit. Figure 23: Journal generation case 3: shopping and dining visit. AirLog : Store-Level Indoor Life Logging Made Easy D Prompt Templates D.1 Spatial Description Generation This prompt converts GeoJSON store features into normal- ized names, neighbor relations, and natural-language posi- tion descriptions. The output is a compact spatial context that later prompts can use without directly reasoning over raw polygon coordinates. Figure 24: Prompt template for spatial description generation from GeoJSON floor-plan features. D.2 Wi-Fi Semantic Anchoring The following prompt maps Wi-Fi SSIDs to candidate store locations via semantic similarity, relative signal strength, and spatial context. Figure 25: Prompt template for Wi-Fi semantic anchoring using SSID semantics and signal evidence.