PaperlayerView exact revision on arXiv

Source-linked AI summary

Masked Latent Prediction of CSI for Indoor Localization in Integrated Sensing and Communication Systems

Ibtissam Labriji, Mahdi Maleki, Pavan Koteshwar Srinath

arXiv:2609.08509v1eess.SP

TL;DR

CSI fingerprinting faces domain shift, limited labeled data, and multipath-related challenges. The paper uses JEPA to predict masked CSI snapshots’ latent embeddings, then localizes with a frozen encoder and lightweight head; on DICHASUS-005x, mean localization error falls from 0.90 m to 0.42 m against a supervised raw-CSI CNN.

  • Problem

    CSI localization is challenged by domain shift, costly labeled radio maps, and multipath/NLOS, while robust label-efficient representations for CSI remain underexplored.

    Source: §I · p. 1

  • Method

    The framework tokenizes CSI snapshots and pretrains an encoder to predict masked snapshots’ latent representations, then freezes it and trains a lightweight head on labeled positions.

    Source: A. Motivation and Contributions · p. 2, §VI · p. 6

  • Results

    0.42 m versus 0.90 m mean localization error on DICHASUS-005x: the proposed method reduces error by 53% relative to a supervised raw-CSI CNN.

    Source: B. Results · p. 5

  • Takeaways & Limitations

    Snapshot masking makes the encoder operate on partial windows by construction, and the results support masked temporal latent prediction as an augmentation-free route aligned with ISAC objectives.

    Source: §VI · p. 6

    The reported results use a single-antenna transmitter, with spatial signatures coming from the 32-element receive array.

    Source: B. Results · p. 6

Abstract

from arXiv · show

Channel State Information (CSI)-based fingerprinting can enable accurate indoor localization but suffers from domain shift, limited labeled data, and degraded performance in multipath-rich environments. To address these challenges, we propose a self-supervised localization framework built on a Joint Embedding Predictive Architecture (JEPA). Each CSI time snapshot is treated as a token, and the encoder is pre-trained by predicting the latent embeddings of masked snapshots from the visible ones, learning robust channel representations without any labels. The encoder is then frozen, and only a lightweight regression head is trained on a small set of labeled positions to estimate the user location. We evaluate the framework on the measured DICHASUS-005x dataset, a single-antenna transmitter received by a 32-antenna array in a multipath-rich indoor environment. With a 50% masking ratio, the proposed method reduces the mean localization error from 0.90 m to 0.42 m (a 53% reduction, or 0.48 m) relative to a supervised Convolutional Neural Network (CNN) baseline trained on raw CSI. Owing to its label-efficient, frozen-encoder design, the framework aligns with integrated sensing and communication (ISAC) objectives, enabling reliable sensing with minimal labeling and compute overhead.

I. INTRODUCTION

CSI fingerprinting offers indoor localization from communication signals, but domain shift, costly labeling, and multipath complicate robust deployment. The paper applies masked latent prediction to CSI, targeting a gap in label-efficient, environment-invariant representations for localization.

I. INTRODUCTION

  • CSI fingerprinting faces domain shift and costly labeled radio-map collection, with multipath and NLOS further complicating localization.The introduction also notes that supervised models can lose accuracy under environmental change.Source: §I · p. 1

A. Motivation and Contributions

  • The framework treats each CSI time snapshot as a token and predicts masked snapshots’ latent representations, avoiding augmentation design and raw-CSI reconstruction.Only a lightweight regression head is trained on a small set of labeled positions after the encoder is pretrained and frozen.Source: A. Motivation and Contributions · p. 2
  • 0.42 m versus 0.90 m mean localization error: the proposed method reduces error by about 2.1× relative to a supervised CNN trained on raw CSI.The evaluation uses the measured DICHASUS dataset with a single-antenna transmitter and 32-antenna receiver.Source: A. Motivation and Contributions · p. 2

II. STATE OF THE ART

  • Prior approaches either depend on fragile, label-intensive fingerprints, RF augmentations or low-level reconstruction, or specialized ranging hardware.JEPA-style latent prediction is presented as a route to geometry-linked representations without augmentation design or raw-signal reconstruction.Source: §II · p. 2

III. SYSTEM MODEL

The system estimates a user’s position from CSI collected by radio infrastructure under realistic propagation conditions. It separates unlabeled encoder pretraining from labeled localization with a frozen encoder and lightweight head.

III. SYSTEM MODEL

  • The system estimates a user’s spatial position from high-dimensional CSI collected by a base station or access point under multipath, NLOS, and irregular sampling.The proposed pipeline is illustrated in Fig. 1.Source: §III · p. 2

A. Overview

  • Offline pretraining splits each CSI window into visible and masked snapshots and predicts masked latents using an EMA copy of the encoder, without labels.Historical unlabeled logs can supply this stage.Source: A. Overview · p. 2
  • A frozen encoder and lightweight head map embeddings to coordinates from a small set of surveyed points; inference averages predictions across masking patterns.The head can be retrained when a site changes, and inference is described as fast and edge-friendly.Source: A. Overview · p. 3, Figure 1 · p. 3

B. CSI Signal Representation

CSI observations are organized as time-ordered snapshots, which may be irregularly sampled or missing. Each snapshot remains a single token spanning the frequency–antenna structure.

B. CSI Signal Representation

C. Temporal Tokenization

  • Each time-aligned token spans the full frequency–antenna structure, preserving temporal evolution without splitting CSI along frequency or antenna dimensions.Source: C. Temporal Tokenization · p. 3

IV. PROPOSED METHOD

The method predicts masked CSI snapshot embeddings from visible snapshots, using temporal context for self-supervised representation learning. An EMA target encoder and stop-gradient stabilize prediction without negatives or hand-crafted RF augmentations.

IV. PROPOSED METHOD

  • The method’s self-supervised principle and full localization pipeline are presented through masked temporal latent prediction.The window is partitioned into visible and masked subsets.Source: §IV · p. 3, A. Masked Temporal Learning · p. 3

A. Masked Temporal Learning

  • Masking hides entire CSI snapshots, requiring the encoder to infer them from temporal context and making the design tolerant by construction to missing or irregular measurements.The passage states that this pushes the encoder toward geometry-aware structure.Source: A. Masked Temporal Learning · p. 3

B. Encoder–Predictor Architecture

  • The context encoder embeds visible snapshots, while a predictor estimates masked-token latents using visible embeddings and masked-position indices.The targets come from an EMA target encoder, and the encoder is frozen for the later localization stage.Source: B. Encoder–Predictor Architecture · pp. 3–4
  • The EMA target and stop-gradient prevent representation collapse, so latent prediction needs neither negative examples nor hand-crafted RF augmentations.Source: B. Encoder–Predictor Architecture · pp. 3–4

C. Latent Prediction Objective

The objective predicts masked-token embeddings from visible embeddings and trains them in latent space. Layer normalization and smooth-L1 loss stabilize comparisons and limit the influence of outlier dimensions.

C. Latent Prediction Objective

  • Training minimizes latent-space prediction loss between predicted and target embeddings of masked tokens.The predictor estimates masked-token representations from visible embeddings and masked-position indices.Source: C. Latent Prediction Objective · p. 4
  • Layer normalization standardizes each D-dimensional latent to zero mean and unit variance, making comparisons invariant to scale and offset and stabilizing targets.D is the encoder embedding dimension, and ϵ is a small constant for numerical stability.Source: C. Latent Prediction Objective · p. 4
  • Smooth-L1 loss with β = 1 is quadratic for small residuals and linear for large ones, limiting outlier influence from multipath and NLOS fluctuations.The loss is applied element-wise and averaged over the embedding dimension; the latent-space objective captures temporally consistent channel structure without modeling low-level CSI detail.Source: C. Latent Prediction Objective · p. 4

D. Lightweight Localization Module

The localization module concatenates frozen-encoder embeddings of visible tokens in temporal order, then maps that fixed-dimension representation to a position estimate. A small MLP is the only labeled component, and predictions are averaged across deterministic balanced masks at inference.

D. Lightweight Localization Module

  • Visible-token embeddings from the frozen encoder are concatenated in temporal order to form Zwin.With a fixed masking ratio, the number of visible tokens is constant, so Zwin has fixed dimension.Source: D. Lightweight Localization Module · p. 4
  • With fixed masking ratio, Zwin has a fixed dimension, and a small LWLM maps it to a position estimate; only LWLM is trained with labeled data, while the encoder remains frozen.Source: D. Lightweight Localization Module · p. 4
  • At inference, predictions are averaged over deterministic balanced masks to produce a mask-invariant position estimate.The window representation uses only the frozen encoder’s visible-token embeddings.Source: D. Lightweight Localization Module · p. 4

E. Training Strategy

Training separates unlabeled representation learning from labeled localization: the encoder and predictor learn through masked temporal latent prediction, while a frozen encoder feeds a labeled position head. This decoupling is designed to reduce labeling requirements and online latency, and is expected to ease cross-site adaptation.

E. Training Strategy

  • The framework uses a two-stage procedure separating offline unlabeled representation learning from labeled localization.In the first stage, the online encoder and predictor perform masked temporal latent prediction, while the target encoder is updated by EMA.Source: E. Training Strategy · p. 4
  • For localization, the encoder is frozen and only the lightweight module G is trained on labeled positions.Decoupling is designed to reduce labeling requirements and online latency.Source: E. Training Strategy · p. 4
  • The design is expected to ease cross-site adaptation because only the small module G must be retrained for a new environment.This is an expected benefit of separating representation learning from localization.Source: E. Training Strategy · p. 4

V. PERFORMANCE EVALUATION

The evaluation uses measured DICHASUS-005x CSI windows and compares a frozen-encoder localization pipeline with an end-to-end raw-CSI CNN under a shared split and masking setup. Performance is assessed with test-set localization-error summaries and distributions, while relative gain accounts for scenario-specific scale.

V. PERFORMANCE EVALUATION

  • The measured DICHASUS-005x dataset uses a single-antenna transmitter and 32-antenna receiver, yielding 3,200 windows from 25,600 snapshots.Eight consecutive snapshots form each time-ordered window, assigned the median ground-truth position; the resulting sample shape is 3 × 8 × 64 × 32.Source: §V · pp. 4–5
  • The frozen Vision Transformer encoder is pretrained with masked temporal latent prediction, and a small MLP head is trained using a 70/15/15 split and 50% masking.Test predictions average six deterministic coverage-balanced masks; the end-to-end raw-CSI CNN baseline uses the identical split, masking, and ensemble averaging.Source: §V · p. 5

A. Performance Metrics

  • Per-sample Euclidean localization errors are summarized by mean localization error and an empirical CDF, with final error reported on the held-out test set.The CDF gives the fraction of test samples localized within a distance d; Table I covers mean localization error on the DICHASUS test set, and Fig. 3 shows the test-set error CDF.Source: A. Performance Metrics · p. 5, Table I · p. 5, Figure 3 · p. 5
  • Relative gain is reported because absolute error in meters depends on deployment extent and sampling, while same-scenario comparison isolates representation contribution from scenario-specific scale.The provided passages establish the metric rationale but do not provide table values.Source: A. Performance Metrics · p. 5

B. Results

The proposed localizer outperforms the supervised CNN baseline on the held-out test set, while its error distribution is left-shifted and validation error converges to a low, stable floor. The reported evaluation uses a single-antenna transmitter; performance with multiple transmit antennas is expected to improve but is not measured here.

B. Results

  • 0.42 m versus 0.90 m mean localization error: the proposed method reduces held-out test error by 53% relative to the supervised CNN baseline.The reported relative gain is approximately 2.1×.Source: B. Results · p. 5
  • 70% of test windows fall within 0.5 m versus 23% for the CNN, and the proposed method's 90th-percentile error is 0.77 m versus 1.51 m.The error CDF is consistently left-shifted, indicating lower typical error and fewer large-error outliers.Source: B. Results · p. 5
  • Validation error decreases rapidly to a low, stable floor; the initial approximately 2 m error reflects measurement-area extent and non-uniform coverage, not model performance.The validation-error-per-epoch comparison is shown in Fig. 4.Source: B. Results · p. 5, Figure 4 · p. 6
  • The evaluation uses a single-antenna transmitter, so the measured spatial signature comes from the 32-element receive array; improvement with multiple transmit antennas is expected, not demonstrated.The text anticipates that additional spatial degrees of freedom could improve accuracy.Source: B. Results · p. 6

VI. CONCLUSION

The paper presents a self-supervised JEPA framework that predicts masked CSI snapshot embeddings, then freezes its encoder and trains a lightweight localization head on labeled positions. On measured DICHASUS-005x data, it reduces mean localization error relative to a supervised CNN; the authors identify partial-window operation and ISAC alignment as supported implications, and list generalization and label-budget studies as future work.

VI. CONCLUSION

  • The JEPA framework predicts latent representations of masked CSI snapshots from visible ones without labels, then freezes the encoder and trains a lightweight localization module on labeled positions.The paper describes this as a label-efficient, computationally lightweight pipeline.Source: §VI · p. 6
  • On measured DICHASUS-005x data, mean localization error falls from 0.90 m to 0.42 m versus a supervised CNN trained on raw CSI, with a consistently left-shifted error distribution.The reported improvement is 2.1×.Source: §VI · p. 6
  • Masking entire snapshots makes the encoder operate on partial windows, and the authors identify the approach as aligned with ISAC objectives.Future work includes measuring label efficiency across budgets and evaluating cross-site and cross-device generalization.Source: §VI · p. 6
Loading 2609.08509v1…