PaperlayerView exact revision on arXiv

Source-linked AI summary

UniCSI Towards a Universal Wi-Fi CSI Encoder for Ubiquitous Human Sensing

Daniel Eckhoff, Hua Kang, Zhitang Chen, Jie Chuai

arXiv:2610.09559v1cs.LGcs.NI

TL;DR

Heterogeneous CSI configurations produce varied tensor dimensions, and existing preprocessing strategies do not provide a unified model that exploits these datasets together. UniCSI natively ingests such measurements with a band-relative tokenizer and fixed-size spectral aggregation. Experiments report improved cross-device and cross-environment generalization, especially where fixed-grid approaches collapse.

  • Problem

    CSI varies across devices and acquisition configurations, while existing preprocessing approaches do not yield one model that exploits heterogeneous datasets within a unified framework.

    Source: Introduction · p. 1

  • Method

    UniCSI uses per-subcarrier band-relative encoding and cross-attention to compress variable-length spectra into fixed representations for native heterogeneous ingestion.

    Source: Introduction · p. 1

  • Results

    UniCSI substantially improves cross-device and cross-environment generalization under supervised and self-supervised settings, with the largest gains where fixed-grid approaches collapse.

    Source: Conclusion · p. 7

  • Takeaways & Limitations

    The findings support handling CSI heterogeneity in the architecture rather than normalizing it away beforehand for cross-domain transfer.

    Source: Discussion · p. 7

    The study compares abstractions of prior ingestion strategies rather than published checkpoints; head-to-head evaluation with each method’s native preprocessing remains future work.

    Source: Discussion · p. 7

Abstract

from arXiv · show

Wi-Fi sensing promises to turn the everyday wireless signals that already surround us into ubiquitous sensors for human sensing. However, a fundamental obstacle is that CSI is acquired under diverse device-specific configurations, including different subcarrier counts, bandwidths, and carrier bands. Consequently, the resulting CSI tensors vary in both spectral resolution and tensor shape, making heterogeneous modeling challenging. Standard architectures struggle with such heterogeneity, forcing lossy pre-processing which compromises the underlying signal. To bridge this gap, we present UniCSI, a unified foundation architecture that directly operates on heterogeneous CSI while preserving the integrity of the native waveform. UniCSI hinges on two core innovations: (1) a physics-informed RF tokenizer that encodes each frequency channel based on its fractional position within the physical spectrum rather than rigid array indices. It preserves intrinsic spectral coherence and enables seamless, frequency resolution-agnostic processing across arbitrary sensing configurations. (2) a spectral aggregator that distills variable-length channel sequences into a fixed-size spectral signature, effectively decoupling the feature dimensionality from the physical subcarrier spacing. Extensive evaluations on a large-scale corpus of 25 heterogeneous public datasets, spanning 14 to 2048 subcarriers, 20 to 160 MHz bandwidth, and the 2.4 and 5 GHz bands, demonstrate that native heterogeneous ingestion substantially improves cross-domain transfer under both supervised and self-supervised training schemes, particularly in regimes where fixed-grid architectures fail to generalize.

Introduction

CSI varies across devices and configurations, while existing approaches either preprocess measurements or remain tied to fixed acquisition settings. UniCSI instead ingests heterogeneous CSI natively, encoding subcarriers by relative position within the band and compressing variable-length spectra for a shared model.

Introduction

  • CSI datasets differ in signal dimensions and acquisition settings, and those differences can also change over time on one device.Configurations include subcarrier count, bandwidth, carrier frequency, device count, and antenna setup.Source: Introduction · p. 1
  • Zero-padding, resampling, and per-device backbones waste capacity, distort signals, or prevent knowledge sharing, without yielding one unified model.The paper frames this as a barrier to using heterogeneous datasets in a single model.Source: Introduction · p. 1
  • UniCSI accepts varying subcarrier counts, bandwidths, carrier bands, and MIMO configurations without padding or resampling, using band-relative subcarrier positions and fixed-size spectral compression.Its per-link tokenizer handles subcarriers independently, while cross-attention compresses variable-length spectra.Source: Introduction · pp. 1–2
  • Jiang et al. (2025) resample CSI to a fixed 600 × 90 input, while WiSenseNet (Lyons, Pandey, and Santra 2025) and AM-FM (Zhu et al. 2026) also resolve heterogeneity before or beneath the model.The paper argues that resampling discards physical subcarrier spacing; WiSenseNet uses subcarrier-selection preprocessing, and AM-FM uses a fixed padded grid.Source: Related Work · p. 2
  • CSI-JEPA (Luo, Li, and Liu 2026) uses channel-variation-aware masking but retains fixed-grid, index-based tokenization and evaluates within one benchmark.WiFi-JEPA (Kim et al. 2026) instead predicts one antenna-link view from others, but also assumes a fixed configuration and index-based positional embeddings.Source: Related Work · p. 2
  • UniCSI differs by treating configuration variation as an architectural input and comparing against controlled fixed-grid instantiations under matched data and training conditions.Earlier foundation models generally share a backbone across tasks but vary in how they address CSI heterogeneity.Source: Related Work · p. 2

Method

UniCSI processes each antenna link independently, tokenizes every subcarrier, and aggregates the resulting variable-length spectrum into a fixed representation. The same architecture supports supervised multi-task training and self-supervised JEPA pretraining.

Input Representation

  • The method uses amplitude and discards phase because most public datasets provide amplitude, while raw phase is corrupted by offsets and requires chipset-specific sanitization.The authors describe avoiding configuration-dependent phase preprocessing as consistent with UniCSI’s design goal.Source: Method · p. 2

Physics-Informed RF Tokenizer

  • Each subcarrier is tokenized independently with a 1D CNN and time-window projections, so subcarrier count determines sequence length rather than requiring fixed input shapes.The method preserves distinct temporal and frequency roles instead of patching the full CSI matrix in two dimensions.Source: Method · p. 3

Spectral Aggregator

  • A query-based spectral aggregator compresses variable-length subcarrier sequences into a fixed-size representation for the downstream Transformer.Learned queries cross-attend to the subcarrier tokens; repeated attention and Transformer refinement progressively update the compressed representation.Source: Method · pp. 3–4

Physics-Informed RF Tokenizer

  • Subcarriers receive learned positional codes based on normalized frequency within the observed band, making relative positions consistent across bandwidths and channel configurations.The encoding needs neither a lookup table nor a predefined maximum subcarrier count.Source: Method · p. 3

Encoder Overview

  • The resulting latent incorporates temporal-window and band embeddings before Transformer processing, while per-link encoding supports arbitrary MIMO configurations and inference-time prediction aggregation.The architecture overview depicts per-link tokenization, learned-query compression, and a Transformer encoder.Source: Method · pp. 2, 4, Figure 1 · p. 3

Training Objectives

  • The same encoder supports supervised multi-task learning with a shared backbone and lightweight dataset-specific heads, using a weighted sum of task losses.Classification uses cross-entropy among the task-specific losses.Source: Method · p. 4
  • Self-supervised pretraining uses JEPA: temporal-spectral blocks are held out as targets, and a predictor maps context representations to targets produced by an EMA encoder.Training uses smooth-L1 loss on parameter-free layer-normalized targets.Source: Method · p. 4

Experiments

Experiments evaluate UniCSI across heterogeneous CSI datasets and show its strongest gains on difficult cross-domain benchmarks. They also assess pretrained transfer, few-shot adaptation, and the contributions of the tokenizer and spectral aggregator.

Experimental Setup

  • The assembled pretraining corpus contains 425.6 hours and 391,277 measurements after held-out evaluation data are excluded; JEPA pretraining uses a bandwidth-stratified sampler.The corpus spans human-sensing tasks and heterogeneous MIMO configurations; the sampler mixes 20, 40, and at least 80 MHz links in a 50/25/25 ratio.Source: Datasets · p. 4, JEPA Pretraining · p. 6

Multi-task Supervised Training

  • 71.6% macro-F1 across eight HAR test sets beats the padded ViT’s 62.4% by 9.2 points; on unseen-environment Exposing-CSI, UniCSI scores 32.5% while every baseline is at or below chance.On NTUFi HAR, GaitID, and FallDar, UniCSI scores 100.0%, 99.2%, and 99.4%, respectively.Source: Figure 2 · p. 4, Multi-task Supervised Training · p. 5

Transferability of Pretrained Representations

  • Across most downstream datasets, pretrained UniCSI outperforms pretrained Padded ViT under both frozen linear-probe and k-NN evaluation, with the largest gains on harder cross-environment benchmarks.The comparison uses the same corpus and a baseline sharing the fixed patch grid, index-based positional embeddings, and JEPA objective.Source: Transferability of Pretrained… · p. 6, Figure 3 · p. 7

Full Fine-Tuning

Sample Efficiency

  • With 5 labeled samples per class, pretraining raises average fused accuracy from 29.6% to 47.2% and macro-F1 from 17.3% to 43.7%.Pretraining improves performance at every tested label budget, with the largest gains in the 5-shot setting.Source: Sample Efficiency · p. 6, Figure 4 · p. 7

Ablation

  • Removing the RF tokenizer lowers average accuracy from 73.3% to 63.5%, while removing spectral aggregation lowers it from 73.3% to 71.7%.The tokenizer preserves per-subcarrier temporal processing; aggregation restores accuracy on low-subcarrier devices and fixes latent size for any subcarrier count, which JEPA pretraining depends on.Source: Ablation · p. 7

Discussion

Across supervised and self-supervised training, UniCSI transfers across devices and environments where fixed-grid baselines collapse toward chance. The discussion attributes gains to native heterogeneous ingestion while noting limits in frequency encoding and checkpoint comparisons.

Discussion

  • Under supervised joint training, UniCSI alone transfers to unseen environments where fixed-grid baselines collapse toward chance; self-supervised representations also transfer under frozen probing and few-shot fine-tuning.The few-shot comparison is against training from scratch.Source: Discussion · p. 7
  • The ablation attributes gains to the shape-agnostic tokenizer and spectral aggregator rather than scale or data alone.The tokenizer has the larger measured effect; the aggregator recovers accuracy on low-subcarrier devices and fixes latent size for JEPA pretraining.Source: Discussion · p. 7
  • The positional encoding is deliberately simple; richer frequency representations, including absolute carrier frequency, remain follow-up directions.Source: Discussion · p. 7
  • Comparisons use abstractions of prior ingestion strategies rather than published checkpoints; head-to-head evaluation with each method’s native preprocessing remains future work.Source: Discussion · p. 7

Conclusion

UniCSI natively processes heterogeneous CSI configurations without lossy preprocessing, combining a physics-informed RF tokenizer and spectral aggregator. Experiments show improved cross-device and cross-environment generalization, especially where fixed-grid approaches collapse.

Conclusion

  • UniCSI natively accommodates diverse CSI configurations without lossy preprocessing.Source: Conclusion · p. 7
  • Its physics-informed RF tokenizer and spectral aggregator process varying subcarrier counts, bandwidths, and carrier bands in one model.Source: Conclusion · p. 7
  • Across supervised and self-supervised experiments, architectural handling of heterogeneity substantially improves cross-device and cross-environment generalization, with the largest gains where fixed-grid approaches collapse.Source: Conclusion · p. 7
Loading 2610.09559v1…