Ambient intelligence seeks to continuously understand human presence, activity, and physiology in physical spaces, enabling smart environments, health monitoring, and human-computer interaction. WiFi infrastructure offers a ubiquitous, always-on, and privacy-preserving sensing substrate across billions of IoT devices. Yet ambient sensing with WiFi remains largely fragmented, with most systems relying on task-specific models, labeled data collection, and customized training pipelines. We present AmbientFM, a foundation model for ambient sensing through ubiquitous WiFi signals. AmbientFM is pre-trained on 9.2 million unlabeled Channel State Information (CSI) samples collected over 439 days from 20 commercial device types deployed in real homes, the largest real-world WiFi CSI pre-training corpus reported to date. It learns transferable wireless representations through contrastive learning, masked reconstruction, and physics-informed objectives tailored to wireless signals. With lightweight adaptation, a single backbone transfers to 9 downstream tasks, achieving above 0.90 AUROC on all classification benchmarks and supporting dense spatial reconstruction. AmbientFM reduces the need for labeled data and task-specific design, enabling scalable ambient intelligence on existing wireless infrastructure.
Figures & tables
Figure 1: AmbientFM overview. AmbientFM pre-trains a shared sensing backbone on in-the-wild CSI with complementary self-supervised objectives, then adapts it to diverse downstream tasks with lightweight tuning.
Method
Application domain
Foundation model
Pre-training scale
Real/Sim.
#Tasks
SSL-HAR Xu et al. (2025)
Sensing
No
≤ 45K / setting
Real
1
CIG-MAE Liu et al. (2025)
Sensing
No
2.2K–3.8K / dataset
Real
1
CAPC Barahimi et al. (2024)
Sensing
No
5.5K
Real
2
CSI-MAE Jiang et al. (2026)
Wireless channel modeling
Yes
1.45M CSI
Sim.
3
WiFo-CF Liu et al. (2026)
Wireless communication
Yes
160K CSI
Sim.
2
AmbientFM
Ambient sensing with WiFi
Yes
9.2M / 439 days
Real
9
Table 1: Comparison with related self-supervised and foundation-model work on wireless signals.
Method
Fall
HAR
Gesture
UID
Localization
MSR
Occupancy
Proximity
Imaging †
AmbientFM + BA
0.919
0.923
0.999
0.993
0.995
0.974
0.989
0.936
0.726 / 18.87
(0.902, 0.935)
(0.915, 0.932)
(0.999, 0.999)
(0.990, 0.995)
(0.992, 0.997)
(0.970, 0.977)
(0.986, 0.992)
(0.927, 0.946)
AmbientFM + TP
0.873
0.915
0.634
0.996
0.990
0.992
0.989
0.928
0.722 / 18.43
(0.851, 0.893)
(0.906, 0.922)
(0.623, 0.645)
(0.994, 0.997)
(0.986, 0.994)
(0.991, 0.993)
(0.985, 0.992)
(0.919, 0.938)
Supervised baseline
0.916
0.527
0.564
0.995
0.631
0.593
0.566
0.494
0.726 / 18.41
(0.899, 0.932)
(0.513, 0.541)
(0.562, 0.568)
(0.993, 0.997)
(0.606, 0.653)
(0.581, 0.606)
(0.544, 0.587)
(0.477, 0.510)
Table 2: Cross-task performance across downstream WiFi sensing tasks. AUROC and 95% confidence intervals (CI) are reported for classification tasks and SSIM / PSNR are reported for the reconstruction task Imaging † .
Figure 2: Cross-task transfer performance with small data sample sizes on four downstream tasks.
Figure 3: Effect of augmentation strategy on downstream performance (left). Loss component ablation on the 1/5 pre-training corpus (right).
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Collection
Reported size
Duration
Devices/chipsets
Env.
Subjects
Band/BW
SignFi [ 19 ]
Scripted
8,280
–
1/1
2
5
5 GHz/20 MHz
WiFiCam [ 25 ]
Scripted
18,148
10 min
1/1
1
1
2.4 GHz/20 MHz
Widar3.0 [ 44 ]
Scripted
17K
–
1/1
3
16
5.8 GHz/20 MHz
XRF55 [ 29 ]
Scripted
42.9K
59.6 h
1/1
4
39
5 GHz/20 MHz
CrossSense [ 42 ]
Scripted
606K
–
1/1
3
100
5 GHz/20 MHz
CSI-Bench [ 46 ]
In-the-wild
231.6K
461 h
16 configurations
26
35
2.4/5 GHz, 20–80 MHz
Appendix
Table 3: Comparison of the AmbientFM pre-training corpus with representative public WiFi sensing datasets.
Task
Source
#Classes
#Samples
Category
Fall Detection
CSI-Bench [ 46 ]
2
6,700
Physiological
Human Activity Recog.
CSI-Bench [ 46 ]
5
41,503
Activity
Motion Source Recog.
CSI-Bench [ 46 ]
4
60,900
Activity
Occupancy Detection
CSI-Bench [ 46 ]
2
∼ 6,000
Spatial
Localization
CSI-Bench [ 46 ]
6
7,062
Spatial
Proximity Recognition
CSI-Bench [ 46 ]
4
64,154
Spatial
Appendix
Table 4: Summary of downstream evaluation tasks.
Hyperparameter
Small (2M)
Base (5M)
Large (12M)
Embedding dimension ( dmodel )
192
256
768
Encoder layers
4
6
12
Attention heads
6
8
12
Feed-forward dimension
384
512
2048
Dropout
0.1
0.1
0.1
Max sequence length
512
512
512
Appendix
Table 5: Transformer Encoder Hyperparameters across model scales.
Augmentation
Parameters
p
Rationale
Additive Gaussian Noise
σ=0.005
0.5
Receiver thermal noise
Amplitude Modulation
factor =0.05
0.5
AGC and path-loss fading
Frequency Perturbation
scale =0.02 (FFT)
0.5
Hardware variation
Temporal Crop
crop ratio =0.8
0.5
Partial observation
Frequency Masking
mask ratio =0.1
0.5
Frequency-selective fading
Gaussian Blur
σ=1.0 , kernel =5
0.5
Spectral smoothing
Appendix
Table 6: Contrastive augmentations applied during pre-training. The first eight are each applied independently with p=0.5 ; time reversal is applied with p=0.2 .
Task
Method
AUROC
95% CI
Accuracy
F1
FAR
Fall Detection
Supervised baseline
0.916
(0.899, 0.932)
0.827
0.844
0.117
AmbientFM + TP
0.873
(0.851, 0.893)
0.801
0.833
0.252
AmbientFM + BA
0.919
(0.902, 0.935)
0.834
0.852
0.125
Human Activity Recognition
Supervised baseline
0.527
(0.513, 0.541)
0.520
0.137
0.200
AmbientFM + TP
0.915
(0.906, 0.922)
0.765
0.685
0.072
AmbientFM + BA
0.923
(0.915, 0.932)
0.800
0.729
0.059
Appendix
Table 7: Test-set performance across downstream tasks. AUROC is reported with 95% CI. FAR denotes false alarm rate.
Method
Fall
HAR
Gesture
UID
Loc.
MSR
Occ.
Prox.
Imaging
Task-specific reference
0.949
0.754
0.948
0.995
1.000
0.996
0.988
0.775
0.749 / 20.39
AmbientFM + BA
0.834
0.800
0.939
0.936
0.938
0.886
0.958
0.796
0.726 / 18.87
Appendix
Table 8: Comparison with published task-specific references. Classification tasks report accuracy and Imaging reports SSIM / PSNR. References are from CSI-Bench [ 46 ] for the seven CSI-Bench tasks, SignFi [ 19 ] for Gesture, and WiFiCam [ 25 ] for Imaging.
Figure 4: t-SNE visualization of learned representations for fall detection. (a) Random initialization shows overlapping clusters with poor class separation, indicating the model has not learned meaningful features. (b) SSL pretrained model demonstrates separation between fall and non-fall classes, revealing that self-supervised pretraining learns discriminative representations beneficial for downstream tasks.
Figure 5: t-SNE visualization of learned representations for localization. (a) Random initialization exhibits highly overlapping clusters across all six classes, with no discernible boundaries between different locations, indicating the absence of meaningful spatial features. (b) SSL pretrained model shows improved cluster formation with reduced inter-class overlap, demonstrating that self-supervised pretraining captures spatial patterns that facilitate location-specific classification.
Model
Fall
HAR
UID
Loc.
MSR
Occ.
Prox.
Pre-train loss
Small (2M)
0.898
0.840
0.984
0.979
0.971
0.980
0.901
3.63
Base (5M)
0.919
0.923
0.993
0.995
0.974
0.989
0.936
3.09
Large (12M)
0.627
0.524
0.990
0.985
0.515
0.651
0.496
3.27
Appendix
Table 9: Downstream performance (AUROC) across model scales, with the converged pre-training loss.
Metric
Fall
Loc.
UID
Prox.
HAR
MSR
Pre-trained
3.61
4.12
5.16
3.64
1.57
2.58
Random Init
0.25
0.24
0.21
0.14
0.03
0.12
Ratio ( × )
14.4
17.4
24.2
26.5
52.0
21.0
Average
Pre: 3.45 Rand: 0.16 Ratio: 25.9 ×
Appendix
Table 10: Inter-class distance in the representation space.
r
Trainable Params
Accuracy
AUROC
64
1.98M
0.778
0.928
128
2.77M
0.787
0.935
192
3.55M
0.794
0.936
256
4.34M
0.764
0.932
Appendix
Table 11: Bottleneck dimension ( r ) trade-off on Proximity Recognition. Trainable Params includes adapter parameters across all 6 encoder layers plus the LSTM temporal classifier and task head.
Parameter
Lower
Default
Higher
Mask ratio
0.05: 0.647
0.10: 0.687
0.25: 0.675
Loss weights ( λCL:λREC:λACF )
1:1:1: 0.639
1:4:1: 0.687
1:8:1: 0.672
Appendix
Table 12: Sensitivity of HAR AUROC to mask ratio and loss weights (1/5 pre-training corpus, bottleneck adaptation).
Task
Band
Test samples
Accuracy
AUROC
F1
FAR
HAR
2.4 GHz
1,067
0.876 ± 0.025
0.958 ± 0.007
0.827 ± 0.040
0.039
HAR
5 GHz
2,333
0.789 ± 0.013
0.927 ± 0.004
0.716 ± 0.022
0.068
Proximity
2.4 GHz
527
0.820 ± 0.019
0.953 ± 0.006
0.821 ± 0.021
0.060
Proximity
5 GHz
1,201
0.776 ± 0.021
0.933 ± 0.011
0.778 ± 0.022
0.076
Appendix
Table 13: Performance by carrier band. Accuracy, AUROC, and F1 are mean ± standard deviation.
Rate
Packet loss
Accuracy
AUROC
F1
100 Hz
0%
0.839 ± 0.019
0.917 ± 0.009
0.860 ± 0.017
95 Hz
5%
0.802 ± 0.023
0.878 ± 0.032
0.829 ± 0.013
90 Hz
10%
0.789 ± 0.028
0.862 ± 0.045
0.813 ± 0.015
80 Hz
20%
0.760 ± 0.041
0.843 ± 0.058
0.777 ± 0.026
60 Hz
40%
0.720 ± 0.069
0.813 ± 0.100
0.729 ± 0.056
Appendix
Table 14: Fall Detection under reduced effective packet rates. Mean ± standard deviation across repeated evaluations; the 100 Hz mean is consistent with Table 2 (0.919).
Task
Split
0-shot
1-shot
5-shot
10-shot
20-shot
HAR
Cross-device
0.411 / 0.604
0.388 / 0.591
0.483 / 0.676
0.537 / 0.769
0.575 / 0.808
Cross-environment
0.425 / 0.537
0.399 / 0.584
0.400 / 0.615
0.430 / 0.654
0.492 / 0.704
Cross-user
0.498 / 0.605
0.451 / 0.626
0.477 / 0.670
0.508 / 0.718
0.550 / 0.753
Proximity
Cross-device
0.432 / 0.638
0.430 / 0.661
0.497 / 0.705
–
0.605 / 0.820
Cross-environment
0.357 / 0.600
0.366 / 0.586
0.405 / 0.618
–
0.516 / 0.738
Cross-user
0.303 / 0.558
0.320 / 0.560
0.323 / 0.571
–
0.390 / 0.631
Appendix
Table 15: Accuracy / AUROC under CSI-Bench’s cross-domain protocols. Proximity has no 10-shot setting.
Task
Device-macro Acc.
Device-macro AUROC
Device-macro F1
HAR
0.730 ± 0.073
0.920 ± 0.038
0.683 ± 0.088
Proximity
0.834 ± 0.076
0.952 ± 0.033
0.833 ± 0.088
Appendix
Table 16: Device-macro performance. Values are mean ± standard deviation across three seeds.
Channel state information (CSI) provides a widely available sensing modality for human and environment perception, but existing CSI sensing models usually rely on task-specific supervised training and require substantial labeled data for each task, device, user, or environment. This limits their scalability in practical deployments where unlabeled CSI is abundant but labeled data is costly to collect. In this paper, we present CSI-JEPA, a self-supervised predictive representation learning framework for label-efficient and transferable Wi-Fi sensing. CSI-JEPA learns reusable temporal-spectral representations from unlabeled CSI samples by predicting latent features of masked channel regions from visible context. To better match the physical structure of CSI, it jointly tokenizes CSI along the time and subcarrier dimensions and introduces a channel variation-aware masking strategy that prioritizes regions with stronger local channel dynamics. After pretraining, the encoder is frozen and adapted to downstream sensing tasks using lightweight task-specific adapters. We evaluate CSI-JEPA on seven real-world Wi-Fi sensing tasks spanning diverse objectives and deployment settings. The results show that CSI-JEPA improves downstream sensing performance over competitive baselines, achieving up to 10.64 percentage points (pp) mean accuracy gain over the supervised Transformer and matched-budget label savings of up to 95.0%. Across unseen tasks, environments, hardware platforms, and CSI organizations, frozen CSI-JEPA improves downstream adaptation by up to 22.92 pp over non-pretrained baselines, demonstrating strong representation transferability.
WiFi sensing based on Channel State Information (CSI) promises ubiquitous, device-free perception, yet current research remains trapped in a Tower of Babel - fragmented into isolated silos where models are tailored to specific hardware dialects, fixed environments, and narrow tasks. The primary bottleneck is the Heterogeneity Gap: the disparity in signal dimensions, sampling rates, and semantic labels that prevents cross-system understanding. To bridge this gap, we propose a foundation-model framework that treats CSI not merely as raw signals but as a structured language with a learnable universal grammar. We first curate and standardize a large collection of heterogeneous real-world CSI datasets, establishing a unified infrastructure that allows incompatible signal formats to be treated as a single corpus. Second, we introduce a modular architecture that acts as a universal translator where lightweight dataset-specific adapters tokenize diverse signal inputs into a shared latent vocabulary, while a shared self-supervised Transformer backbone learns the temporal syntax of human motion and environmental dynamics. This design decouples sensing semantics from hardware syntax. Extensive evaluations show that by mastering this universal language, our approach consistently outperforms task-specific baselines and exhibits strong generalization capability in new environments, achieving superior efficiency in few-shot scenarios. By effectively absorbing heterogeneity, the framework offers a path toward robust, general-purpose wireless sensing, mirroring the linguistic generalization observed in Large Language Models. The code implementation is available at: https://github.com/cjychenjiayi/WiLLM.
Jiayi Chen, Weiting Ou, Guangxu Zhu
Shenzhen Research Institute of Big Data · The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute (SLAI)
Wi-Fi-based human activity recognition has advanced substantially, but most existing methods assume a closed set of activities and require labeled Wi-Fi samples for every target class, limiting their ability to recognize unseen activities. We present Zero-Fi, a contrastive signal-language alignment framework for zero-shot Wi-Fi-based human activity recognition. Zero-Fi learns unified representations from complementary Wi-Fi signal features and aligns them with the semantic representations of natural-language activity descriptions in a shared embedding space. This cross-modal alignment enables Zero-Fi to recognize new activity classes without requiring labeled Wi-Fi samples or model adaptation for those classes. Experiments on large-scale public benchmark datasets demonstrate effective zero-shot recognition of held-out activity classes, highlighting the potential of signal-language alignment to extend Wi-Fi sensing beyond predefined activity classes.
Yitong Shen, Cheng Guo, Peiliang Wang +5
University of South Florida · University of North Carolina at Charlotte