Multi-Task Wi-Fi Sensing Integration

• Multi-task Wi-Fi sensing integration is a unified framework that extracts, fuses, and infers diverse task-relevant features from CSI to improve localization, activity, and presence detection.
• It leverages advanced neural architectures, such as Transformer and LSTM models, alongside synthetic data modeling to simulate complex environmental scenarios.
• Experimental benchmarks reveal notable improvements in accuracy, with localization and activity recognition gains up to 8% over traditional single-task methods.

Multi-task Wi-Fi sensing integration refers to the unified extraction, fusion, and inference of multiple environmental and device-centric tasks—such as localization, activity recognition, presence detection, gesture recognition, and user identification—over the same physical-layer Wi-Fi radio signals and infrastructure. This paradigm departs from single-task (siloed) sensing by sharing data, feature extraction, and inference architectures to exploit the contextual and statistical dependencies among tasks. Recent advances span theoretical unification, hardware and protocol adaptations, robust learning pipelines, and plug-and-play extensibility, and draw on the IEEE 802.11bf/bf standardization, advanced neural architectures, and multi-band cross-modality feature fusion.

1. Unified Theoretical Foundations

Classic Wi-Fi sensing algorithms operate on a single-task basis, wherein the system models an event set $\mathcal{S}(q)$, extracts a feature set $\mathcal{F}(q)$ from channel state information (CSI), and applies an inference mapping $$\mathcal{M}_i(q): \mathcal{F}(q) \to \mathcal{S}(q)$$. This is encapsulated by

$$\mathcal{S}(q) \xrightarrow{\mathcal{M}_f(q)} \mathcal{F}(q) \xrightarrow{\mathcal{M}_i(q)} \mathcal{S}(q).$$

In multi-task (MTL) settings, one must aggregate the union of event sets $$\mathcal{S}(\mathcal{Q}) = \bigcup_i \mathcal{S}(q_i)$$, features $$\mathcal{F}(\mathcal{Q}) = \bigcup_i \mathcal{F}(q_i)$$, and construct a nontrivial set-to-set inference operator $$\mathcal{M}_{i}(\mathcal{Q}): \mathcal{F}(\mathcal{Q}) \to \mathcal{S}(\mathcal{Q})$$. Critically, $\mathcal{M}_{i}(\mathcal{Q})$ is not just a collection of single-task solvers; it must resolve conflicts and leverage latent cross-task correlations (Li et al., 16 Jan 2026).

2. Integrated Pipeline Architectures

Contemporary multi-task Wi-Fi sensing frameworks implement a modular pipeline consisting of:

• Event/Feature Set Definition: A comprehensive $\mathcal{S}$ and $\mathcal{F}$ spanning all current and anticipated tasks (e.g., presence, activity, tracking, location).
• Synthetic Data Modeling: Event sequence simulation via Markovian motion models and kinematics, then mapping these to CSI features (subcarrier correlation, DSER, PLCR) for synthetically generated training sets (Li et al., 16 Jan 2026).
• Shared Inverse Models: Aggregated neural architectures, typically Transformer or LSTM encoder-decoders, are trained to perform sequential inference $\mathcal{F}_{1:T} \to \mathcal{S}_{1:T}$, with multi-head output decoders for per-task inference.
• Plug-and-Play Extensibility: New tasks or features are added by expanding $\mathcal{F}$ and $\mathcal{Q}$ and retraining the inverse model, without changing pipeline structure.

This architecture allows elastic integration of tasks such as localization, activity classification, and presence detection, while abstracting away sensor-specific preprocessing, thus facilitating rapid extension to vital sign monitoring, gesture recognition, and multi-user tracking (Li et al., 16 Jan 2026, Zhang et al., 2021).

3. Multi-task Feature Fusion and Cross-band Integration

Deployment of multi-band and multi-modal sensing techniques has further enhanced robustness and generalization:

• Granularity-Matched Fusion: Multi-tier neural networks extract features from sub-6 GHz (fine-grained CSI) and mmWave (mid-grained beam SNR) in parallel, then hierarchically fuse them via learnable cross-layer matching (Yu et al., 2021). Granularity-matched fusion substantially outperforms naïve feature concatenation or single-modality methods, particularly under limited label regimes.
• Time-aware Attention and Irregular CSI Streams: Handling heterogeneity in timestamp intervals and packet types, a time-aware attention network encodes irregularly sampled, cross-band CSI, providing a unified embedding for multi-task inference without the need for injected probes (Dong et al., 14 Dec 2025).

These methods enable accurate multi-task inferences such as pose, occupancy, and indoor localization, even with scarce labeled data, showing gains of 5–8% absolute accuracy over baseline approaches (Yu et al., 2021, Dong et al., 14 Dec 2025).

4. Joint Sensing and Communication in Wi-Fi Standards

The integration of sensing into IEEE 802.11bf/ax/be protocols entails enhancements at PHY, MAC, and Application layers:

• PHY Layer: Sounding NDP PPDUs, configurable LTFs for fine delay-Doppler maps, expanded (up to 320 MHz) RUs for finer range/Doppler resolutions. Beamforming and multi-link operation (MLO) for spatial diversity (Meneghello et al., 2022, Tai et al., 11 Apr 2025).
• MAC Layer: Sensing-aware scheduler interleaves PPDUs and data, aggregate CSI reports with minimal overhead, explicit resource allocation optimization ($U=\alpha R_\mathrm{comm}+\beta R_\mathrm{sense}$) to balance throughput and sensing quality.
• Application Layer: Shared CNN backbones process joint features with dedicated task heads. Joint multi-task objective functions optimize error per task and system-level latency (Li et al., 16 Jan 2026, Meneghello et al., 2022).

A fundamental trade-off is observed: as system parameters (e.g., $\alpha$) favor sensing, localization error and task accuracy improve, but aggregate throughput may drop (Meneghello et al., 2022, Tai et al., 11 Apr 2025). Cooperative and non-cooperative approaches employing Kalman-filter-based prediction and CRLB-driven trilateration for STA selection provide formal trade-off surfaces for throughput, fairness, and sensing accuracy (Tai et al., 11 Apr 2025).

5. Experimental Benchmarks and Quantitative Performance

State-of-the-art multi-task Wi-Fi sensing achieves high accuracy and low error in diverse environments:

System	Localization Error (median)	Activity Accuracy	Presence Accuracy	Tasks Integrated
Uni-Fi (Li et al., 16 Jan 2026)	0.52 m	98.34%	98.57%	Localization, Activity, Presence
Wimuse (Zhang et al., 2021)	1-2 m (appx) (varies)	95.70% (gesture)	99%+	Gesture, Loc., Identification
MILAGRO (Picazo-Martinez et al., 30 Jul 2025)	N/A (grid)	95–100%	99.9% (pose)	Presence, Pose, Tracking
ISAC-Fi (Chen et al., 2024)	1.12 m (monostatic)	82% (HAR)	N/A	Ranging, Doppler, HAR, Imaging
Multi-Band GM (Yu et al., 2021)	95.8% (loc. class.)	94.4% (pose)	95.5% (occupancy)	Pose, Occupancy, Localization
UniFi (Dong et al., 14 Dec 2025)	N/A (benchmark tasks)	96.88% (HAR)	N/A	HAR, Gesture, Fall, Counting

Most pipelines plateau in localization error beyond sampling rates of 500 Hz. Transformer architectures typically surpass LSTM in tracking. Generalization across packet types and environments requires feature and model design sensitive to statistical variation (e.g., cross-band alignment, attention-based fusion) (Li et al., 16 Jan 2026, Dong et al., 14 Dec 2025, Yu et al., 2021).

6. Security, Privacy, and Practical Implications

Multi-task Wi-Fi sensing raises significant privacy and security concerns, including unauthorized inference of human presence or activity (via unencrypted CSI), AP spoofing, or CSI manipulation. Mitigations include local training data usage, CSI perturbation, encrypted training headers, and SNR-thresholding (Picazo-Martinez et al., 30 Jul 2025, Meneghello et al., 2022).

Efficient, backward-compatible integration is achievable. For example, ISAC-Fi demonstrates that inserting adaptive analog and digital self-interference cancellation within a standard 802.11 PHY allows concurrent monostatic radar-like sensing and legacy communication, with 77 dB SI suppression and negligible communication impact (Chen et al., 2024). Multi-band and multi-link operation further enable robust, fine-grained sensing in real-world topologies (Picazo-Martinez et al., 30 Jul 2025, Meneghello et al., 2022).

7. Open Challenges and Future Directions

Despite the maturation of integrated multi-task Wi-Fi sensing, several challenges persist:

• Scalability to New Tasks: Adding tasks (e.g., respiration, gesture recognition, vital sign monitoring) requires extensible pipelines and feature sets and continued advances in synthetic data and continual learning (Li et al., 16 Jan 2026, Yu et al., 2021).
• Multi-user, Multi-task Tracking: Most systems remain single-user; adaptation to dense, multi-user, or multi-activity scenarios demands graph-based latent representations and robust user-task association (Li et al., 16 Jan 2026).
• Sparse or Dynamic Traffic: Sensing quality degrades with sparse commercial traffic. Hybrid methods fusing injected probes and ambient CSI, and extensions to uplink mobile clients or multi-AP fusion are proposed (Dong et al., 14 Dec 2025).
• Domain Adaptation and Edge Inference: Continual, few-shot adaptation and real-time edge deployment across diverse, changing environments remain open; solutions leverage autoencoder pretraining and rapid transfer learning (Yu et al., 2021).
• Standardization and Regulatory Alignment: Next steps include aligning with IEEE 802.11bf extensions for encrypted LTFs, tighter SI calibration, and large-scale MIMO or distributed monostatic deployments (Chen et al., 2024, Meneghello et al., 2022).

This suggests the trajectory of the field is toward modular, resource-adaptive, and privacy-aware Wi-Fi networks that deliver simultaneous, scalable, and resilient sensing and communications, catalyzed by advances in unified architectures, robust feature fusion, and platform-driven standardization.