Flow-guided Localization: Methods & Applications

• Flow-guided Localization is a technique that uses dynamic flow cues instead of static landmarks to infer positions in diverse settings.
• In nanoscale scenarios, FGL leverages circulation time and event detection from energy-harvesting nanodevices operating under strict size and power constraints.
• Recent advances employ graph neural networks and set models to enhance accuracy while managing trade-offs in detection, transmission, and energy consumption.

Searching arXiv for papers on flow-guided localization and related usages. Flow-guided Localization (FGL) refers, in the cited literature, to localization formulations that use flow, motion, or iterative transport as the principal signal for inferring location. In in-body nanoscale localization, FGL associates the location of a sensed biological event with raw reports produced by energy-harvesting nanodevices that drift passively with blood flow and communicate through THz links; a raw report is modeled as $X=(t,b)$, where $t$ is a circulation-related time quantity and $b\in\{0,1\}$ is an event bit (Bartra et al., 2023, Pascual et al., 2023). In other research lines, the same label or a closely related formulation is used for optical-flow-guided visual sound source localization, VLM-guided flow matching for lesion-mask refinement, flow-field-based robot localization and mapping, and iterative flow prediction for cross-view geolocalization (Singh et al., 2022, Wang et al., 7 Apr 2026, Song et al., 2019, Lehyeh et al., 23 Mar 2026). A common structural feature across these usages is that localization is derived from dynamic observables rather than from static appearance or discrete landmarks alone.

1. In-body nanoscale FGL: system model and observables

In the bloodstream setting, nanodevices are injected into the bloodstream, drift passively with blood flow, and operate under stringent size and power constraints. Their physical size is on the order of red blood cells ($\sim 5\,\mu\text{m}$), they have no onboard localization sensors, and they rely on pulse-based THz ultra-short-range communication to a single on-body anchor placed at the torso near the heart (Galván et al., 2024). The bloodstream is abstracted either as a directed graph $G=(V_R,E_R)$, whose nodes are anatomical regions and edges encode vascular adjacency, or as $r$ disjoint regions $R_1,\dots,R_r$ with fixed per-lap travel times $T_i$ and physiology-determined region-choice probabilities $P_{R_i}$ (Galván et al., 2024, Pascual et al., 2023).

The localization target is a single biological event in exactly one region $r^\*$ or $t$0. Nanodevices periodically sample for a binary diagnostic event, such as the presence or absence of a biomarker, and after each detection cycle emit a single positive-bit impulse if they detected the event (Galván et al., 2024). In the formal definition used for THz-based nanoscale FGL, the on-body anchor collects raw tuples

$t$1

where $t$2 is the inter-arrival time since the last heart passage and $t$3 indicates whether the event was detected on that circulation (Bartra et al., 2023). In the standardized benchmarking workflow, anchors log quadruples

$t$4

which are subsequently streamed into a localization algorithm (López et al., 2023).

The observable statistics are shaped by two non-idealities. First, event detection is imperfect because the device turns on only when its harvester reaches a turn-on threshold, which is modeled through a detection probability $t$5. Second, communication is unreliable because each heart passage only offers a limited upload opportunity, with successful THz transmission probability $t$6 that accounts for path loss, molecular absorption, self-interference, and energy availability (Pascual et al., 2023). This immediately distinguishes nanoscale FGL from formulations that assume perfect sensing or deterministic communication.

2. Analytical raw-data modeling and standardized benchmarking

The analytical model in "Analytical Modelling of Raw Data for Flow-Guided In-body Nanoscale Localization" represents the observed time field as

$t$7

where $t$8 is the number of laps through region $t$9 before upload and $b\in\{0,1\}$0 captures turbulence and physiological variability (Pascual et al., 2023). The joint probability of $b\in\{0,1\}$1 conditioned on the event being in region $b\in\{0,1\}$2 is constructed from a multiset permutation factor

$b\in\{0,1\}$3

a path term

$b\in\{0,1\}$4

a final-lap upload term

$b\in\{0,1\}$5

and event-detection terms

$b\in\{0,1\}$6

For $b\in\{0,1\}$7, the pmf at exactly $b\in\{0,1\}$8 laps is

$b\in\{0,1\}$9

whereas for $\sim 5\,\mu\text{m}$0,

$\sim 5\,\mu\text{m}$1

This formulation makes explicit how failed uploads create compound-lap timing and how intermittent sensing creates false negatives (Pascual et al., 2023).

The paper contrasts this model with an ideal scenario in which $\sim 5\,\mu\text{m}$2, so every lap is reset at upload, $\sim 5\,\mu\text{m}$3 for the first passage through $\sim 5\,\mu\text{m}$4, and $\sim 5\,\mu\text{m}$5 (Pascual et al., 2023). When $\sim 5\,\mu\text{m}$6, failed uploads prolong the reset time and $\sim 5\,\mu\text{m}$7 becomes a compound sum over multiple region times; when $\sim 5\,\mu\text{m}$8, a device may traverse the true region multiple times without setting the event bit (Pascual et al., 2023). This suggests that raw-data distortion is intrinsic to the sensing-and-link budget, not merely a post-processing artifact.

Validation against the BloodVoyagerS-plus-TeraSim simulator used the Mann–Whitney test for $\sim 5\,\mu\text{m}$9 distributions in regions with $G=(V_R,E_R)$0, ECDF maximum vertical distance $G=(V_R,E_R)$1 for $G=(V_R,E_R)$2, KL divergence between per-region $G=(V_R,E_R)$3 ratios, and a convergence test based on MSE between empirical frequencies and the analytical pmf (Pascual et al., 2023). Table II reports that for $G=(V_R,E_R)$4, the fraction of regions where the Mann–Whitney null hypothesis is accepted is $G=(V_R,E_R)$5 at $G=(V_R,E_R)$6, while for $G=(V_R,E_R)$7, it is $G=(V_R,E_R)$8 at $G=(V_R,E_R)$9 (Pascual et al., 2023). The same study states that high agreement is reflected by MW pass rates $r$0, ECDF distances $r$1, and $r$2, with discrepancies concentrated in mid-distribution shifts and in low-probability regions where one or two extra or missing detections strongly skew ratios (Pascual et al., 2023).

Standardized evaluation was proposed in "Toward Standardized Performance Evaluation of Flow-guided Nanoscale Localization" as an end-to-end workflow combining BloodVoyagerS mobility, ns-3/TeraSim THz communication, an energy-harvesting capacitor model, and a Python post-processor (López et al., 2023). The workflow specifies vessel topology, anchors, nanodevice population, target event locations, physical-layer parameters, energy-harvesting rates, antenna gains, SINR thresholds, and simulation duration; generates 3-D trajectories at $r$3; models capacitor charging and turn-on thresholds; logs raw packet fields; and computes standardized metrics such as point localization error, region detection accuracy, reliability, circulation-based detection time, synchronization accuracy, and energy metrics (López et al., 2023). The illustrative example reports occasional circulation times $r$4, reliability rising from $r$5 at $r$6 to $r$7 at $r$8, and region-detection accuracy $r$9 (López et al., 2023).

3. Learning-based nanoscale FGL: graph models, personalization, and set models

The GNN-based FGL framework casts the vascular topology as a heterogeneous graph whose nodes are anatomical regions plus an anchor node, and whose edges comprise region–region adjacency together with anchor–region links (Bartra et al., 2023). Region nodes use features such as type, travel time, and region-choice probability; anchor-node features are derived from the collected raw data by fitting a Gaussian Mixture Model to positive-bit circulation times and extracting mixture parameters $R_1,\dots,R_r$0 together with the total count $R_1,\dots,R_r$1 (Bartra et al., 2023). Message passing is implemented in Heterogeneous Graph Transformer style with type-specific projections, attention weights, and aggregation, followed by a per-region sigmoid readout trained by weighted cross-entropy to counter class imbalance (Bartra et al., 2023).

This architecture addresses two limitations of earlier neural baselines identified in the bloodstream literature: summary-statistic classifiers do not exploit vascular topology, and shallow models degrade in high-speed regions with highly erroneous raw data. The reported result is a $R_1,\dots,R_r$2 reduction in median point error, for example from $R_1,\dots,R_r$3 to $R_1,\dots,R_r$4 at $R_1,\dots,R_r$5 run time, a region-accuracy increase of $R_1,\dots,R_r$6, and bloodstream coverage extending to regions where a fully connected baseline had essentially zero accuracy; with additional anchors at axillary veins, region accuracy improves by up to another $R_1,\dots,R_r$7–$R_1,\dots,R_r$8 (Bartra et al., 2023). A plausible implication is that vascular topology acts as a strong structural prior when individual measurements are sparse or corrupted.

"Tailoring Graph Neural Network-based Flow-guided Localization to Individual Bloodstreams and Activities" extends this graph formulation by adding physiological context. The proposed adaptation pipeline introduces a master node $R_1,\dots,R_r$9 with features $T_i$0, optionally augmented by inverses $T_i$1, and probability edges from heart regions to all others with Monte-Carlo-estimated weights $T_i$2 (Galván et al., 2024). The baseline design, without master node and without probability edges, attains validation region accuracy of approximately $T_i$3, whereas the best extended design, "Graph c," reaches approximately $T_i$4, corresponding to $T_i$5 percentage points and an approximately $T_i$6 relative gain (Galván et al., 2024). Across profiles, the reported gains are Tall $T_i$7 pp, Short $T_i$8 pp, Light $T_i$9 pp, Active $P_{R_i}$0 pp, Big $P_{R_i}$1 pp, and Small $P_{R_i}$2 pp, while Heavy is $P_{R_i}$3 pp and Inactive is slightly below baseline (Galván et al., 2024).

A different architectural direction appears in "Set Transformer Architectures and Synthetic Data Generation for Flow-Guided Nanoscale Localization," which formulates FGL as permutation-invariant set classification over unordered circulation-time sets $P_{R_i}$4 (Hube et al., 22 Aug 2025). The Set Transformer uses scalar embeddings, two ISAB layers, PMA, a final SAB, and a linear head trained with regularized cross-entropy. In the reported 94-region synthetic vasculature, the best GNN baseline achieves $P_{R_i}$5 region accuracy and $P_{R_i}$6 point error, whereas the best Set Transformer achieves $P_{R_i}$7 and $P_{R_i}$8, albeit with substantially longer training times of approximately $P_{R_i}$9–$r^\*$0 versus approximately $r^\*$1–$r^\*$2 (Hube et al., 22 Aug 2025). Conditional synthetic-data augmentation using CVAE slightly improves the best GNN to $r^\*$3 with $r^\*$4 point error, but degrades the best Set Transformer to $r^\*$5 with $r^\*$6 (Hube et al., 22 Aug 2025). The paper attributes this to attention heads already extracting fine-grained patterns that the synthetic samples fail to cover.

4. Design-space exploration, protocol trade-offs, and evaluation practice

The design-space study in "Insights from the Design Space Exploration of Flow-Guided Nanoscale Localization" evaluates two neural FGL solutions under a common simulator configuration that includes THz carrier $r^\*$7, bandwidth $r^\*$8, $r^\*$9, $t$00, $t$01, $t$02, receiver sensitivity $t$03, one anchor at the heart, and BloodVoyagerS topology with 94 vessels and organs (Lemic et al., 2023). Design variables are the number of nanodevices $t$04, sampling granularity $t$05, and event-detection threshold $t$06 (Lemic et al., 2023).

Both evaluated solutions use the same signaling framework: the anchor near the heart sends a THz beacon every $t$07 seconds, an active nanodevice backscatters a response containing $t$08 and $t$09, and the anchor feeds $t$10 into a classifier (Lemic et al., 2023). Solution 1 is a fully connected network trained by L-BFGS and predicting among 24 loops; Solution 2 is a three-layer network with PRELU, BatchNorm, Dropout, hidden size 512, LogSoftMax over 25 classes, and Adam with negative log-likelihood (Lemic et al., 2023). The reported outcome is that Solution 2 consistently outperforms Solution 1 in point and region accuracies, with an example median error of approximately $t$11 versus approximately $t$12 at $t$13 (Lemic et al., 2023).

The same study identifies several protocol-level trade-offs. Increasing $t$14 improves accuracy and accelerates reliability convergence, but gains diminish beyond $t$15; sampling frequency has only marginal impact because the sensing energy cost is small and detection eventually converges; and $t$16 drastically degrades point accuracy through spurious detections in adjacent loops (Lemic et al., 2023). The paper summarizes these trade-offs by illustrative regimes: a low-energy regime with $t$17, $t$18, $t$19, reliability around $t$20 in $t$21, and median error around $t$22; a balanced regime with $t$23, $t$24, $t$25, reliability approximately $t$26 in $t$27, and median error around $t$28; and a high-accuracy regime with $t$29, $t$30, $t$31, reliability approximately $t$32 in $t$33, and median error around $t$34 (Lemic et al., 2023).

A recurring theme across the nanoscale literature is that evaluation methodology materially affects conclusions. The standardized-workflow paper argues against early-stage fragmentation in metrics and simulators, explicitly invoking the lessons of indoor-localization benchmarking and ISO/IEC 18305:2016 (López et al., 2023). The design-space study similarly emphasizes heterogeneous metrics, realism in THz propagation and nanodevice energy, and objective comparison under shared assumptions (Lemic et al., 2023). This suggests that, within in-body FGL, methodological standardization is part of the research problem rather than a purely administrative concern.

5. Motion-guided perception: visual sound localization and lesion refinement

In "FlowGrad: Using Motion for Visual Sound Source Localization," Flow-guided Localization extends the RCGrad framework by explicitly incorporating motion cues through optical flow (Singh et al., 2022). The RCGrad backbone uses a ResNet-18 visual encoder pretrained on ImageNet, a ResNet-18 audio encoder, and an InfoNCE contrastive objective on unlabeled videos. Optical flow is computed between consecutive grayscale frames at $t$35 using Gunnar Farnebäck, converted to a motion-magnitude map $t$36, and min–max normalized to $t$37 (Singh et al., 2022). Three variants are defined: FlowGrad-H, where the RCGrad heatmap is multiplied elementwise by $t$38; FlowGrad-IC, which stacks flow as a fourth image channel; and FlowGrad-EN, which uses three parallel encoders and pairwise contrastive losses $t$39, $t$40, and $t$41, with the fused inference map $t$42 (Singh et al., 2022).

Evaluation on Urbansas uses 5,704 image–audio pairs in which the sounding vehicle is visible, with consensus IoU at threshold $t$43 and AUC as metrics (Singh et al., 2022). The reported quantitative results are: Vision-only+CF+TF at IoU $t$44, AUC $t$45; optical-flow only at IoU $t$46, AUC $t$47; RCGrad at IoU $t$48, AUC $t$49; FlowGrad-H at IoU $t$50, AUC $t$51; FlowGrad-IC at IoU $t$52, AUC $t$53; and FlowGrad-EN at IoU $t$54, AUC $t$55 (Singh et al., 2022). The strongest gain is the $t$56 IoU improvement of FlowGrad-H over RCGrad, while qualitative examples show an important limitation: at traffic signals, vehicles may be sounding but static, so short-term flow can suppress true positives (Singh et al., 2022).

A different but related usage appears in "Unifying VLM-Guided Flow Matching and Spectral Anomaly Detection for Interpretable Veterinary Diagnosis," where FGL denotes an iterative mask-localization process guided by a VLM and trained by conditional flow matching (Wang et al., 7 Apr 2026). The refinement is formulated as the ODE

$t$57

with interpolation path

$t$58

target velocity

$t$59

and flow-matching loss

$t$60

The implementation uses OpenCLIP ViT-B/32 features, a U-Net decoder, a 4-level U-Net-style flow CNN with cross-attention blocks, Adam with learning rate $t$61, $t$62 epochs for segmentation, $t$63 epochs for flow matching, $t$64 time steps per image, and $t$65 NVIDIA RTX 4090 GPUs (Wang et al., 7 Apr 2026). On the test set, the best competing Unet-based model, PolypFlow, reports $t$66, whereas FGL reports $t$67 and $t$68 (Wang et al., 7 Apr 2026). Here, as in FlowGrad, the localization signal is refined by dynamics rather than inferred in a single static pass.

6. Flow fields, iterative displacement prediction, and conceptual boundaries

In robotics, "Concurrent Flow-Based Localization and Mapping in Time-Invariant Flow Fields" reformulates localization as simultaneous estimation of robot trajectory and ambient flow field, replacing classical SLAM landmarks with a continuous flow map (Song et al., 2019). The robot state is

$t$69

the flow field is represented on a rectangular mesh, and observations are relative flow velocities

$t$70

A factor graph combines prior, motion, and flow-measurement factors, and the batch objective is minimized by iterated Gauss–Newton with Levenberg–Marquardt damping (Song et al., 2019). In a steady single-gyre flow, dead reckoning accumulates $t$71–$t$72 drift over a $t$73 mission, whereas FGL reduces 2D positioning RMS to under $t$74; in a turbulent double-gyre, dead-reckoning error reaches $t$75 and FGL reduces it to approximately $t$76 RMS (Song et al., 2019). The paper also states that residual errors are driven by unmodeled time-varying turbulence.

A further extension of the flow-guided idea appears in cross-view geolocalization. "GeoFlow: Real-Time Fine-Grained Cross-View Geolocalization via Iterative Flow Prediction" learns a conditional displacement distribution $t$77, decouples displacement into polar coordinates, predicts $t$78, and refines multiple initial hypotheses through Iterative Refinement Sampling (Lehyeh et al., 23 Mar 2026). The single-step mean update is

$t$79

and the published system uses uniform averaging over the final refined poses. With $t$80 seeds and $t$81 rounds on a V100 GPU, the reported inference speed is $t$82; KITTI Same-Area mean error is $t$83, median $t$84; and VIGOR Same-Area mean is $t$85, median $t$86 (Lehyeh et al., 23 Mar 2026). Although the paper does not use the same biomedical terminology, it exemplifies the same methodological motif: localization via learned flow-like updates over hypotheses.

A common misconception is that "Flow-guided Localization" names a single architecture or a single application domain. The cited literature does not support that reading. Instead, the term is reused for at least four distinct methodological families: bloodstream localization from circulation-time reports, motion-guided audiovisual localization, flow-matching-based lesion localization, and flow-field-based robot localization and mapping (Bartra et al., 2023, Singh et al., 2022, Wang et al., 7 Apr 2026, Song et al., 2019). Another misconception is that flow cues are sufficient by themselves. The cited results are more nuanced: optical-flow-only heatmaps underperform FlowGrad-H and FlowGrad-EN on Urbansas; short flow windows miss static but sounding vehicles; low-probability vascular regions exhibit higher KL sensitivity; and time-invariant flow assumptions break down under turbulence (Singh et al., 2022, Pascual et al., 2023, Song et al., 2019).

Open directions are correspondingly domain-specific. In nanoscale FGL, the literature points to frequency-selective THz-channel refinement, interference- and MAC-aware $t$87, adaptive energy management with time-varying duty cycle $t$88, multi-anchor deployments, online adaptation with sliding windows of heart rate, hybrid graph-plus-set models, and better generative models for rare-region tail behavior (Pascual et al., 2023, Galván et al., 2024, Hube et al., 22 Aug 2025). In audiovisual localization, proposed directions include longer temporal windows, learned flow estimators such as RAFT, binaural or spatial audio cues, and revised evaluation protocols for multi-class, multi-source urban scenes (Singh et al., 2022). In robotic and geolocalization settings, the stated extensions include incremental smoothing, adaptive mesh refinement, Gaussian-process or dynamical flow models, and multi-robot cooperation (Song et al., 2019). Across these lines of work, FGL remains less a single algorithm than a research pattern: localization by exploiting the structure of motion, circulation, or transport.