betaFeedback

WorksComputer Science

MambaLoc: Lightweight Indoor Localization Using Cross-Modal Knowledge Distillation

Abstract

Accurate indoor localization on resource-constrained edge devices remains challenging due to the tension between modeling fidelity and computational efficiency. Wi-Fi Channel State Information (CSI) enables low-cost deployment using existing Wi-Fi infrastructure but is heavily distorted by multipath propagation. Transformer-based solutions mitigate these effects but their quadratic complexity limits real-time edge deployment. Ultra-Wideband (UWB) provides geometrically precise ranging yet requires dedicated infrastructure. We present MambaLoc, a cross-modal distillation framework that improves the accuracy-efficiency balance in CSI-based localization. MambaLoc uses a linear-time Selective State Space Model (Mamba) as a lightweight CSI student, guided during training by a UWB-Transformer teacher. A Feature Alignment Module bridges the representational gap between heterogeneous RF modalities, allowing the CSI model to inherit UWB-derived geometric structure without UWB hardware at inference. A probabilistic Gaussian Mixture regression head models multi-modal spatial ambiguity from multipath effects, avoiding the physically implausible averaging of deterministic regression. Evaluations on a real-world dataset show MambaLoc achieves a median localization error of 0.230 m, reducing error by 22.2% over a standalone Mamba baseline, while requiring 65.1% fewer FLOPs and 58% lower inference latency than Transformer-based approaches.

Introduction

Accurate indoor localization is critical for autonomous systems in GPS-denied environments, including warehouse robotics, drone navigation, and wearable IoT devices. While GNSS provides robust outdoor positioning, signals degrade indoors due to attenuation and multipath propagation [1]. As computation shifts to the edge, localization must deliver both high accuracy and efficient inference under strict hardware constraints.

RF-based approaches leverage existing wireless infrastructure. Wi-Fi CSI provides fine-grained amplitude and phase measurements across multiple subcarriers, implicitly encoding environmental geometry. However, indoor RF propagation is dominated by multipath effects: reflected and diffracted components arrive with different delays and phase shifts, distorting direct range estimates and introducing unreliability [2]. Localization from CSI therefore requires models that integrate information across packet sequences, separating persistent geometric patterns from short-term noise.

Deep learning has substantially improved CSI-based localization. Transformer architectures model long-range temporal structure via self-attention (e.g., TIPS [3]), but self-attention scales quadratically (O(N2)O(N^2)), making continuous inference impractical under Size, Weight, and Power (SWaP) constraints.

To address this trade-off, we introduce MambaLoc, built on Selective State Space Models (SSMs). By leveraging the Mamba architecture [4], our approach achieves linear-time (O(N)O(N)) inference with constant memory state. We further adopt a Cross-Modal Knowledge Distillation strategy inspired by the Learning Using Privileged Information (LUPI) paradigm [5]: UWB data serves as privileged training context, and a UWB-Transformer teacher guides a lightweight CSI-Mamba student via a Feature Alignment Module. The framework can be interpreted as learning a latent spatial representation of RF propagation, where the teacher provides geometric supervision that guides the CSI model toward physically consistent spatial embeddings. The remainder of this paper is organized as follows: § 2 reviews related work; § 3 covers theoretical foundations; § 4 and § 5 describe the MambaLoc architecture and distillation curriculum; Section VI presents empirical results.

Our key contributions are: (1) Cross-modal RF knowledge distillation, a framework transferring geometric structure from UWB CIR to WiFi CSI across both sensing modalities (UWB→CSI) and architectures (Transformer→Mamba), enabling robust localization without UWB at inference; (2) efficient CSI sequence modeling via Selective State Space Models as an alternative to Transformers (0.39 vs. 1.12 GFLOPs); (3) a cross-modal feature alignment module that projects teacher and student representations into a shared latent space, stabilizing distillation and reducing modality mismatch; and (4) a probabilistic spatial regression head using a GMM to model multipath-induced uncertainty.

Related Work

Sensing Modalities for Indoor Localization

Indoor localization modalities carry distinct trade-offs against edge deployment requirements: SWaP, robustness, and infrastructure cost [1]. Vision SLAM [6] and LiDAR [7, 8] approach centimeter-level precision but demand high-throughput processing, degrade in low-light or texture-poor environments, introduce privacy concerns, and add significant payload [9]. UWB offers an excellent accuracy–efficiency balance but requires installing and maintaining a purpose-built anchor network whose geometry critically determines coverage quality; NLOS can still cause outliers even in UWB, making large-scale deployment non-trivial [1]. Geometry-based RF methods (ToA, TDoA, AoA) need no dedicated hardware but are acutely sensitive to NLOS conditions [10]; high-quality AoA additionally demands antenna arrays and tight clock synchronization. Wi-Fi FTM/RTT and BLE lower the infrastructure bar further [11], but FTM depends on compatible access points while BLE suffers RSSI instability and requires dense deployment [1] [12]. Sensor fusion mitigates individual weaknesses but multiplies cost, power, and calibration complexity, which is contrary to lightweight edge deployment.

Commodity Wi-Fi CSI occupies a uniquely attractive position: available on already-deployed infrastructure with no additional hardware, it provides per-subcarrier complex channel estimates encoding rich multipath structure far beyond coarse RSSI. The central challenge is extracting reliable spatial features from noisy, high-dimensional CSI, which brings us to deep learning approaches.

Deep Learning for CSI Localization

Classical RF localization relies on geometric properties (ToA, TDoA, AoA) [10], [13], often assuming reliable LOS propagation; multipath and NLOS conditions introduce severe biases [14]. In contrast, data-driven approaches map CSI fingerprints directly to locations. Early work used MLPs [15] and stacked denoising autoencoders with probabilistic frameworks [16]; hybrid CNN-LSTM architectures [17] added temporal modeling. Attention-augmented CNNs [18] capture both local and global context. Transformer-based models (TIPS [3], Swin-Loc [19]) achieve state-of-the-art accuracy; Transformer-CNN hybrids [20] combine global temporal modeling with local spatial extraction for UWB NLOS classification. However, quadratic complexity is prohibitive for edge deployment. Quantization and distillation [21] compress Transformers for microcontrollers, yet even compressed models struggle under tight memory budgets. This motivates Mamba [4], which offers linear complexity while remaining suitable for edge hardware.

Cross-modal supervision aligns with the LUPI paradigm [5]: high-precision sensors guide models that ultimately rely on cheaper sensors. DistilLoc [22] transfers knowledge from an RTT-based teacher to an RSSI-based student using F-Net, which replaces self-attention with FFT-based mixing (O(Nlog⁡N)O(N \log N)). MambaLoc goes further with a strictly linear O(N)O(N) Mamba student and cross-modal, cross-architecture supervision, using UWB as a geometric teacher to enforce alignment between the student’s CSI latent space and UWB-derived spatial structure that would otherwise be unrecoverable from noisy Wi-Fi data alone.

Background & Theoretical Foundations

Sequence Modeling with Structured State-Space Models

Indoor localization using CSI is inherently a sequential inference problem requiring long temporal dependency modeling in high-resolution wireless signals. Structured State-Space Models (SSMs) have re-emerged as a compelling alternative to attention-based architectures. S4 [23] and Mamba [4] combine continuous-time dynamical systems with the efficiency of recurrent models. Mamba’s selective scan mechanism conditions state transition parameters on the current input xtx_t, enabling adaptive propagation or suppression of information, which is beneficial for discriminating LOS signals from multipath noise. During training, Mamba leverages parallel scan operations; at inference it runs as a recurrent model with constant memory state, processing long sequences in linear time on resource-constrained edge hardware.

Learning Using Privileged Information (LUPI)

We adopt the LUPI paradigm [5], defining the training triplet (xcsi,xuwb∗,ypos)(x_{\mathrm{csi}}, x_{\mathrm{uwb}}^*, y_{\mathrm{pos}}), where xcsix_{\mathrm{csi}} is the deployable modality and xuwb∗x_{\mathrm{uwb}}^* is the privileged modality available only during training. A teacher f∗:X∗→Yf^*: \mathcal{X}^* \to\mathcal{Y} is trained on UWB data to learn a robust geometric mapping; a student f:X→Yf: \mathcal{X} \to\mathcal{Y} then trains on CSI to mimic the teacher’s internal feature representations. By aligning the student’s latent space with the teacher’s, the student effectively learns to filter CSI multipath effects using heuristics derived from the cleaner UWB domain, all without requiring UWB hardware at inference.

Aleatoric Uncertainty in Wireless Localization

Indoor multipath creates scenarios where different locations produce identical signal fingerprints, an inherent ambiguity termed aleatoric uncertainty [24]. Standard regression models confronted with two equally plausible locations (e.g., the true location and a strong reflection) tend to predict their average, often resulting in impossible locations. MambaLoc addresses this with a Gaussian Mixture Model (GMM) head that predicts a full probability distribution, explicitly capturing multi-modal ambiguity. Rather than maximizing likelihood (which can be unstable), we minimize the distance between ground truth and the distribution’s expected value, capturing the uncertainty structure while providing stable gradients [25].

Methodology

System Overview

MambaLoc transfers spatial and geometric knowledge from a high-fidelity UWB-based teacher to a computationally efficient CSI-based student following the LUPI paradigm [5]. As shown in Figure 1, the framework operates in two phases:

MambaLoc system architecture

Figure 1. MambaLoc system architecture. The UWB Transformer teacher guides the CSI Mamba student via the Feature Alignment Module during training; at inference, only the Mamba student runs on CSI data.

  • Offline Cross-Modal Distillation: A UWB-Transformer teacher trained on CIR measurements provides probabilistic supervision and intermediate features. The CSI-Mamba student aligns its latent representations with the teacher through a multi-level distillation pipeline.

  • Online Inference: After training, the teacher and all UWB components are discarded. The CSI-Mamba student performs real-time localization independently, enabling low-latency inference on edge devices.

A key innovation lies in the Feature Alignment Module, which aligns intermediate representations across heterogeneous modalities and architectures rather than distilling only final outputs.

Data Representation and Preprocessing

UWB Teacher Input

The teacher processes UWB Channel Impulse Response (CIR) sequences xu∈RT×Dux_u \in\mathbb{R}^{T \times D_u}, where TT denotes temporal snapshots and DuD_u the feature dimensionality. The input comprises complex CIR magnitude and phase, capturing multipath structure where early peaks correspond to LOS and dominant reflections [26]. We augment with normalized first-path power level and total received power level as NLOS discriminators, since signal attenuation patterns differ significantly between LOS and NLOS states. All features are standardized (zero mean, unit variance).

CSI Student Input

The student consumes Wi-Fi CSI matrices xc∈RT×Dcx_c \in\mathbb{R}^{T \times D_c}, supporting passive, device-free localization [27]. CSI is extracted from IEEE 802.11n frames using a 3×33 \times3 MIMO configuration across Nsc=30N_{sc}=30 subcarriers, yielding 270 complex measurements per packet. Concatenating amplitude and phase gives Dc=2×270=540D_c = 2 \times270 = 540. Raw phase is distorted by sampling frequency offset and packet detection delay; a linear phase sanitization procedure is applied prior to learning [15].

Network Architectures

UWB-Transformer Teacher

An encoder-only Transformer [3] maps normalized CIR inputs to a latent space of dimension dt=256d_t=256 via a linear embedding, followed by L=6L=6 encoder layers with H=8H=8 attention heads. Multi-head self-attention captures global temporal context across variable-length CIR sequences, learning complex multipath and temporal correlations critical for accurate ranging. Its quadratic pairwise interaction (O(N2)O(N^2)) is acceptable in an offline training role. The output feeds a probabilistic GMM regression head.

CSI-Mamba Student

The student uses the Mamba architecture [4] for linear-time complexity. As shown in Figure 2, each Mamba block processes input XX through two parallel branches. The main branch expands features, applies a short Conv1D to capture local temporal dependencies, and then runs the Selective SSM:

Mamba block architecture

Figure 2. Mamba Block Architecture. Processes CSI features using a 1D convolution and a Selective SSM for efficient, linear-time sequence modeling.

ht=Aht−1+Bxt,yt=Cht+Dxt.(1)h_t = A h_{t-1} + B x_t,\qquad y_t = C h_t + D x_t. \tag*{(1)}

Unlike fixed LTI systems, parameters B and C are computed as functions of xtx_t, allowing the model to dynamically vary its memory, selectively propagating relevant multipath features while filtering transient CSI noise or hardware artifacts. Matrix D acts as a direct feedthrough (similar to a residual link). The SSM output is gated by the second branch via element-wise multiplication. This design achieves expressive capacity comparable to Transformers [28] at linear cost.

Feature Alignment Module

Directly distilling final coordinate outputs is insufficient due to the substantial modality gap between UWB and CSI and the scarcity of structural information in a 2D coordinate. Inspired by cross-modal distillation literature [22], [29], [30], we introduce a Feature Alignment Module (Figure 3).

Feature Alignment Module architecture

Figure 3. Feature Alignment Module architecture. Projects teacher (UWB) and student (CSI) features into a shared 144-dimensional space using MSE, Cosine Similarity, and Cross-Attention losses.

Let Fu∈RB×duF_u \in\mathbb{R}^{B \times d_u} and Fc∈RB×dcF_c \in\mathbb{R}^{B \times d_c} denote teacher and student features. The module performs:

  1. Non-Linear Projection: Both feature sets are projected into a shared dad_a-dimensional alignment space via separate two-layer MLPs with LayerNorm and GELU activations: Fu′=ϕu(Fu)F'_u = \phi_u(F_u), Fc′=ϕc(Fc)F'_c = \phi_c(F_c).

  2. Cross-Modal Attention: Student features act as queries against teacher keys and values, allowing the CSI student to selectively extract the geometric features most relevant to its current noisy signal:

F^u=Attention⁡(Q=Fc′, K=Fu′, V=Fu′).(2)\hat{F}_u = \operatorname{Attention}(Q = F'_c,\ K = F'_u,\ V = F'_u). \tag*{(2)}

The alignment loss combines direct feature matching, directional consistency, and attention-guided structural alignment:

Lfeat=λdir∥Fu′−Fc′∥22+λcos(1−cos⁡(Fu′,Fc′))+λattn∥F^u−Fc′∥22(3)\mathcal{L}_{feat} = \lambda_{dir}\lVert F'_u-F'_c\rVert_2^2+\lambda_{cos}\left(1-\cos(F'_u,F'_c)\right)+\lambda_{attn}\lVert\hat{F}_u-F'_c\rVert_2^2 \tag*{(3)}

where λdir=0.4\lambda_{dir}=0.4, λcos=0.3\lambda_{cos}=0.3, λattn=0.3\lambda_{attn}=0.3 (selected by grid search on the validation set). The first term enforces direct feature matching, the second promotes directional consistency, and the third ensures the student aligns with relevant structural information from the teacher.

Probabilistic GMM Regression Head

Both teacher and student use a K=16K=16 component Gaussian Mixture head (KK selected empirically from {4,8,16,32}\{4, 8, 16, 32\} by lowest validation error). The predicted density is:

p(y∣x)=∑k=1Kπk(x)N(y∣μk(x),Σk(x)).(4)p(y|x) = \sum_{k=1}^{K} \pi_k(x)\mathcal{N}(y|\mu_k(x), \Sigma_k(x)). \tag*{(4)}

For simplicity, the covariance is treated as a fixed isotropic variance and is not explicitly trained. Rather than maximizing likelihood directly, the primary objective minimizes the error of the expected coordinate estimate y^=∑kπkμk\hat{y} = \sum_k \pi_k\mu_k:

Ltask=∥y^−ygt∥22.(5)\mathcal{L}_{task} = \lVert\hat{y} - y_{gt} \rVert_2^2. \tag*{(5)}

During distillation, KL divergence between teacher and student mixture weights encourages the student to inherit the teacher’s uncertainty structure: LKD=KL⁡(πT∥πS)\mathcal{L}_{KD} = \operatorname{KL}(\pi_T \mathbin{\|} \pi_S). We distill only the mixture weights π\pi, not the full GMM parameters, because matching which regions of the room are spatially probable is the most critical and transferable aspect of ambiguity structure across the modality gap; enforcing agreement on exact Gaussian shapes is a finer-grained detail that CSI and UWB encode at fundamentally different resolutions. Notably, replacing the GMM head with a deterministic regression head caused the distilled student to underperform the standalone CSI-Mamba baseline, confirming the probabilistic formulation is necessary for stable knowledge transfer.

Training Objective

The total loss combines task supervision, feature alignment, and distribution-level distillation with dynamic weights:

Ltotal=α(t)LKD+β(t)Ltask+γ(t)Lfeat,(6)\mathcal{L}_{total} = \alpha(t)\mathcal{L}_{KD} + \beta(t)\mathcal{L}_{task} + \gamma(t)\mathcal{L}_{feat}, \tag*{(6)}

where α(t)\alpha(t), β(t)\beta(t), and γ(t)\gamma(t) are dynamically adjusted via a curriculum schedule described in § 5. After training, only the CSI-Mamba student is retained, achieving Transformer-level accuracy at linear-time inference cost [31].

Training & Distillation Strategy

Training MambaLoc must bridge two simultaneous gaps: (i) a modality gap between UWB CIR and Wi-Fi CSI, and (ii) an architectural gap between the Transformer teacher and the Mamba student. Naive end-to-end distillation leads to unstable optimization. We adopt a three-stage Cross-Modal Distillation curriculum following LUPI principles, where UWB is available only during training and acts as a geometric oracle.

Stage I: Teacher Pre-Training

The UWB-Transformer teacher is trained independently on UWB CIR measurements and ground-truth coordinates using MSE loss, and although optimized for regression, it implicitly learns a probability distribution over the spatial grid via its GMM weights, producing richer supervisory signals than deterministic regression. Teacher parameters are then frozen.

Stage II: Cross-Modal Aligner Warming

Before student distillation, the Feature Alignment Module is pre-trained for a few epochs, with both teacher and student backbones frozen, to map CSI features toward UWB features. This initialization ensures the alignment loss provides meaningful gradients at the start of Stage III, while full training of the module continues in that phase.

Stage III: Student Distillation with Adaptive Curriculum

The CSI-Mamba student trains jointly under three supervision signals.

  1. Task Supervision: Direct MSE supervision on the student’s own GMM expected output ensures the student remains grounded in real CSI observations rather than over-depending on the teacher, preserving robustness in CSI-only deployment.

  1. Feature Alignment: The weighted Euclidean, cosine, and cross-attention loss from § 4 acts as a geometric regularizer, encouraging CSI features, which are often dominated by noise, to align with the physically grounded UWB representation.

  1. Probabilistic Knowledge Distillation: KL divergence with temperature annealing:

LKD=T(t)2⋅KL⁡(πT(y/T)∥πS(y/T)),(7)\mathcal{L}_{KD} = T(t)^2 \cdot\operatorname{KL}\left(\pi_T(y/T) \mathbin{\|} \pi_S(y/T)\right), \tag*{(7)}

where T(t)T(t) decays linearly from 30.0 to 10.0. A high initial temperature softens the teacher’s distribution, helping the student learn global spatial geometry; progressive annealing then sharpens predictions toward precise localization.

  1. Adaptive Weighting: α(t)\alpha(t) and β(t)\beta(t) are adjusted based on loss trends: if a component deteriorates, its weight increases. The feature alignment weight γ(t)\gamma(t) follows a curriculum decay, starting high to guide early representation learning and decreasing as the student develops CSI-based independence. These strategies were found to stabilize training significantly compared to fixed-weight alternatives through extensive experiments.

Experimental Evaluation

Dataset & Evaluation Metrics

We evaluate on OPERAnet [32], which to our knowledge is the only publicly available dataset providing synchronized UWB and CSI measurements under a shared ground-truth positioning system, making it the canonical benchmark for this problem. Data was collected in two furnished indoor environments (5.09 m×\times10.02 m and 7.20 m×\times7.24 m) over approximately 8 hours, covering six daily activities: walking, sitting, standing, lying down, standing up, and body rotation.

  • CSI: Intel 5300 NICs (5 GHz), 1600 Hz packet rate; one transmitter and two receivers in LOS and bi-static configurations; 3×\times3 MIMO, 30 subcarriers per receiver.

  • UWB: Two synchronized passive systems (Decawave EVK1000 and MDEK1001) providing CIR data at ≈\approx1.0016 ns temporal resolution.

  • Ground Truth: High-resolution 2D positions from a dedicated UWB anchor-tag system; all modalities synchronized via NTP (<< 20 ms).

The primary metric is Median Absolute Error, measured as Euclidean distance between predicted and ground-truth locations. Median Absolute Error is robust to outliers and reflects typical localization performance.

Comparison with State-of-the-Art

We compare against three representative baselines, all re-trained from scratch on OPERAnet for a fair comparison: DeepFi [15], an autoencoder-based compressed CSI representation method; AARES-CNN [18], an attention-augmented residual CNN; and TIPS [3], a Transformer-based system with global self-attention.

As shown in Figure 4, MambaLoc (0.230 m) outperforms DeepFi (0.318 m), AARES-CNN (0.240 m), and TIPS (0.238 m), demonstrating that a lightweight distilled student can match and exceed heavy Transformer-based models when guided by a privileged teacher.

Comparison of Median Absolute Error for DeepFi, AARES-CNN, TIPS, and MambaLoc

Figure 4. Comparison with State-of-the-Art. MambaLoc achieves the lowest Median AE (0.230 m), outperforming CNN and Transformer-based baselines.

ModelGFLOPsLatency (ms)Params
TIPS (Transformer)1.1216.55447k
AARES-CNN0.7012.0767k
MambaLoc (Ours)0.396.94257k

Table 1. Efficiency Comparison. MambaLoc achieves the best efficiency–accuracy trade-off.

Computational Efficiency Analysis

Experimental Setup

All experiments use PyTorch on a single NVIDIA GeForce GTX 1650 GPU (4 GB VRAM), representative of mid-range edge computing hardware. Training uses AdamW with cosine annealing: the UWB teacher is trained for 40 epochs (lr 2×10−32\times10^{-3}); the CSI student uses lr 5×10−45\times10^{-4} for stable convergence during knowledge distillation; batch size is 32. Latency is measured with CUDA synchronization; FLOPs computed via thop; all models benchmarked at B=1B=1 for single-sample inference.

Results

Table I shows that MambaLoc achieves 6.94 ms inference—2.38×\times faster than TIPS (16.55 ms) and 1.73×\times faster than AARES-CNN—a direct consequence of replacing quadratic self-attention with linear-time recurrent inference. MambaLoc requires only 0.391 GFLOPs per forward pass (65.1% reduction vs. TIPS) and has the smallest footprint at 2.57×1052.57\times10^{5} parameters, 42.5% fewer than TIPS and 2.98×\times fewer than AARES-CNN. On battery-powered edge devices where compute, memory, and power are primary constraints, these reductions translate to feasibility of continuous real-time localization without dedicated accelerator hardware.

Ablation Studies

Distillation Strategy

Figure 5 summarizes incremental gains:

Median Absolute Error in the distillation ablation studies

Figure 5. Impact of Distillation Strategies. Progressive improvements in localization accuracy (lower MAE is better) as distillation components are introduced.

  • CSI-Mamba (no teacher): Median AE = 0.2957 m, serving as the lower bound.

  • Same-modality distillation (CSI-Mamba →\rightarrow CSI-Mamba): 0.2757 m, confirming distillation benefits while highlighting the limits of non-privileged teachers.

  • Cross-architecture distillation (CSI-Transformer →\rightarrow CSI-Mamba): 0.2466 m, demonstrating the value of distilling from attention-based architectures.

  • MambaLoc (UWB-Transformer →\rightarrow CSI-Mamba): 0.2300 m, proving UWB provides critical geometric cues inaccessible from CSI alone and validating the cross-modal, cross-architecture approach.

Architectural Ablation

Figure 6 illustrates depth/width sensitivity for both models.

Four line plots showing MAE versus Student Layers, Student Dimension, Teacher Layers, and Teacher Dimension

Figure 6. (a,b) The CSI student favors a “shallow-wide” design to capture high-dimensional CSI features without overfitting noise. (c,d) The UWB teacher exhibits a complexity “sweet spot” (3 layers, 80 dim); exceeding this widens the capacity gap, hindering effective distillation.

Student: Increasing latent dimension dmodeld_{\mathrm{model}} from 60 to 160 monotonically reduces error; a wider latent space is essential to preserve the high-dimensional CSI frequency-diversity cues (Dc=540D_c=540) without premature compression. Depth beyond 2 layers degrades performance, as deeper networks overfit CSI noise artifacts rather than signal structure; a “shallow-wide” architecture acts as an effective regularizer.

Teacher: An optimal point exists at 3 layers and 80 dimensions. Before this, the teacher cannot accurately model UWB geometry. Beyond it, an overly complex teacher learns highly UWB-specific representations that the CSI student cannot imitate across the modality gap, increasing distillation error.

Hyperparameter Sensitivity

  • Alignment dimension (Figure 7): Sharp optimum at D=144D=144. Dimensions below 120 create information bottlenecks; above 150 the student maps noise to sparse dimensions, failing to learn meaningful geometry.

  • Attention heads (Figure 8a): Error is minimized at 8 heads, likely corresponding to dominant multipath components (e.g., LOS and primary reflections). Increasing to 12 heads over-fragments the signal, reducing semantic coherence.

  • Distillation temperature (Figure 8b): Static-temperature analysis reveals a “V-curve” with optimum at T=15T=15; below 10 distributions are too sharp to capture spatial ambiguity, and above 20 they are too flat. These findings motivate our annealing schedule (30→10): the high initial temperature prevents early commitment to specific spatial peaks, forcing global structure learning; the sweep through T≈15T\approx15 provides effective distillation; ending at T=10T=10 acts as a precision refinement phase, unlike starting training with a static T=10T=10, which causes failure, ending there after global structure is encoded achieves high-precision localization.

Line plot of MAE versus Alignment Dimension

Figure 7. A dimension of 144 balances capacity with regularization; lower values cause bottlenecks (< 120) and higher values permit overfitting (> 150).

Two line plots showing MAE versus Teacher Heads and Temperature

Figure 8. (a) Localization error is minimized at 8 attention heads. (b) A minimum at T=15T=15 balances distribution sharpness with entropy.

Conclusion

We presented MambaLoc, a CSI-based localization framework combining linear-time SSMs with cross-modal distillation using UWB as privileged training information. Results demonstrate that the modality gap between UWB and CSI is reduced by aligning intermediate feature representations rather than relying on output-level distillation alone. The three-stage curriculum and adaptive weighting strategy are critical for stable convergence when distilling across sensing modalities and architectures, achieving a median localization error of 0.230 m with 65.1% fewer FLOPs and 58% lower latency than Transformer-based approaches. Future work will evaluate the framework in cross-building scenarios [33] and investigate domain adaptation for cross-environment generalization. The distillation framework may also be extended to other RF sensing tasks such as activity recognition and environmental mapping, and to additional sensing modalities, directions that may clarify the role of cross-modal supervision and efficient sequence modeling in edge-native RF perception systems.

References

  1. [1]F. Zafari, A. Gkelias, and K. K. Leung, “A survey of indoor localization systems and technologies,” IEEE Communications Surveys & Tutorials, vol. 21, no. 3, pp. 2568–2599, 2019.arxiv.org/abs/1709.01015
  2. [2]A. F. Molisch, Wireless Communications, 2nd ed. John Wiley & Sons, 2011.
  3. [3]Z. Zhang, H. Du, S. Choi, and S. H. Cho, “Tips: Transformer based indoor positioning system using both csi and doa of wifi signal,” IEEE Access, vol. 10, pp. 111 363–111 376, 2022.DOI
  4. [4]A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” in First Conference on Language Modeling, 2024.arxiv.org/abs/2312.00752
  5. [5]V. Vapnik and R. Izmailov, “Learning using privileged information: Similarity control and knowledge transfer,” Journal of Machine Learning Research, vol. 16, no. 61, pp. 2023–2049, 2015.
  6. [6]R. Mur-Artal, J. M. M. Montiel, and J. D. Tardós, “Orb-slam: A versatile and accurate monocular slam system,” IEEE Transactions on Robotics, vol. 31, no. 5, pp. 1147–1163, 2015.
  7. [7]J. Zhang and S. Singh, “LOAM: Lidar odometry and mapping in real-time,” in Proceedings of Robotics: Science and Systems (RSS), July 2014.DOI
  8. [8]H. Rizk, H. Yamaguchi, M. Youssef, and T. Higashino, “Laser range scanners for enabling zero-overhead wifi-based indoor localization system,” ACM Transactions on Spatial Algorithms and Systems, vol. 9, no. 1, pp. 1–25, 2023.DOI
  9. [9]Y. Alkendi, L. Seneviratne, and Y. Zweiri, “State of the art in vision-based localization techniques for autonomous navigation systems,” IEEE Access, vol. 9, pp. 76 847–76 874, 2021.
  10. [10]S. Hara, D. Anzai, T. Yabu, T. Derham, and R. Zemek, “Analysis on toa and tdoa location estimation performances in a cellular system,” in 2011 IEEE International Conference on Communications (ICC), 2011, pp. 1–5.
  11. [11]M. Mohsen, H. Rizk, H. Yamaguchi, and M. Youssef, “Timesense: Multiperson device-free indoor localization via rtt,” IEEE Internet of Things Journal, vol. 11, no. 24, pp. 39 593–39 605, 2024.arxiv.org/abs/2409.00030
  12. [12]H. Rizk, A. Elmogy, and H. Yamaguchi, “A robust and accurate indoor localization using learning-based fusion of wi-fi rtt and rssi,” Sensors, vol. 22, no. 7, p. 2700, 2022.
  13. [13]M. A. G. Al-Sadoon, R. Asif, Y. I. A. Al-Yasir, R. A. Abd-Alhameed, and P. S. Excell, “Aoa localization for vehicle-tracking systems using a dual-band sensor array,” IEEE Transactions on Antennas and Propagation, vol. 68, no. 8, pp. 6330–6345, August 2020.
  14. [14]G. Fokin, “Tdoa measurement processing for positioning in non-line-of-sight conditions,” in 2018 IEEE International Black Sea Conference on Communications and Networking (BlackSeaCom), June 2018, pp. 1–5.
  15. [15]X. Wang, L. Gao, S. Mao, and P. Pandey, “Csi-based fingerprinting for indoor localization: A deep learning approach,” IEEE Transactions on Vehicular Technology, vol. 66, no. 1, pp. 763–776, 2017.arxiv.org/abs/1603.07080
  16. [16]M. Abbas, M. Elhamshary, H. Rizk, M. Torki, and M. Youssef, “Wideep: Wifi-based accurate and robust indoor localization system using deep learning,” in 2019 IEEE International Conference on Pervasive Computing and Communications (PerCom). IEEE, 2019, pp. 1–10.
  17. [17]M. T. Hoang, B. Yuen, K. Ren, X. Dong, T. Lu, H. L. Nguyen, R. Westendorp, and K. Reddy, “A cnn-lstm quantifier for single access point csi indoor localization,” arXiv preprint arXiv:2005.06394, 2020.
  18. [18]B. Zhang, H. Sifaou, and G. Y. Li, “CSI-Fingerprinting indoor localization via Attention-Augmented residual convolutional neural network,” IEEE Transactions on Wireless Communications, vol. 22, no. 8, pp. 5583–5597, 2023.
  19. [19]X. Xu, F. Zhu, S. Han, Z. Yu, H. Zhao, B. Wang, and P. Zhang, “Swin-Loc: Transformer-Based CSI fingerprinting indoor localization with MIMO ISAC system,” IEEE Transactions on Vehicular Technology, vol. 73, no. 8, pp. 11 664–11 679, 2024.
  20. [20]H.-J. Hwang, S.-G. Jeong, and W.-J. Hwang, “Ultrawideband non-line-of-sight classification using transformer-convolutional neural networks,” IEEE Access, vol. 13, pp. 86 323–86 332, 2025.
  21. [21]T. Suwannaphong, F. Jovan, I. Craddock, and R. McConville, “Optimising tinyml with quantization and distillation of transformer and mamba models for indoor localisation on edge devices,” Scientific Reports, vol. 15, no. 1, p. 10081, 2025.
  22. [22]H. Rizk, A. Elmogy, M. Rihan, and H. Yamaguchi, “A precise and scalable indoor positioning system using cross-modal knowledge distillation,” Sensors, vol. 24, no. 22, p. 7322, 2024.
  23. [23]A. Gu, K. Goel, and C. Ré, “Efficiently modeling long sequences with structured state spaces,” in International Conference on Learning Representations, 2022.
  24. [24]A. Kendall and Y. Gal, “What uncertainties do we need in bayesian deep learning for computer vision?” in Advances in Neural Information Processing Systems, vol. 30, 2017.
  25. [25]M. Cao, T. Wattyney, and T. Chang, “A gmm-assisted learning approach for ranging with ieee802.15.4 devices,” in 2025 IEEE 23rd International Conference on Industrial Informatics (INDIN), Kunming, China, July 2025.
  26. [26]A. R. Jiménez and F. Seco, “Improving the accuracy of decawave’s uwb mdek1001 location system by gaining access to multiple ranges,” Sensors, vol. 21, no. 5, p. 1787, 2021.
  27. [27]M. Youssef, M. Mah, and A. Agrawala, “Challenges: device-free passive localization for wireless environments,” in Proceedings of the 13th annual ACM international conference on Mobile computing and networking, 2007, pp. 222–229.
  28. [28]B. N. Patro and V. S. Agneeswaran, “Mamba-360: Survey of state space models as transformer alternative for long sequence modelling: Methods, applications, and challenges,” Engineering Applications of Artificial Intelligence, vol. 159, p. 111279, 2025.
  29. [29]G. E. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” ArXiv, vol. abs/1503.02531, 2015.arxiv.org/abs/1503.02531
  30. [30]S. Gupta, J. Hoffman, and J. Malik, “Cross modal distillation for super-vision transfer,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016, pp. 2827–2836.
  31. [31]K. B. Letaief, Y. Shi, J. Lu, and J. Lu, “Edge artificial intelligence for 6g: Vision, enabling technologies, and applications,” IEEE Journal on Selected Areas in Communications, vol. 40, no. 1, pp. 5–36, January 2022.arxiv.org/abs/2111.12444
  32. [32]M. J. Bocus, W. Li, S. Vishwakarma, R. Kou, C. Tang, K. Woodbridge, I. Craddock, R. McConville, R. Santos-Rodriguez, K. Chetty, and R. Piechocki, “Operanet, a multimodal activity recognition dataset acquired from radio frequency and vision-based sensors,” Scientific Data, vol. 9, no. 1, p. 474, 2022.
  33. [33]A. Hesham, E. Samir, H. Rizk, and M. Youssef, “Locatingua: Leveraging language models for cross-building wifi mapping,” in Proceedings of the 32nd ACM International Conference on Advances in Geographic Information Systems, 2024, pp. 723–724.

Paper details

vlnc:W_n53kxac3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents