AM-FM: A Foundation Model for Ambient Intelligence Through WiFi

2026-02-13 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Internet of Things (IoT) & Connected Devices, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

AM-FM is the first foundation model designed for ambient intelligence and sensing using WiFi Channel State Information (CSI). It was pre-trained on an extensive dataset of 9.2 million unlabeled CSI samples collected over 439 days from 20 commercial device types across 11 real-world environments and 26 users. The model learns general-purpose representations through a self-supervised framework incorporating contrastive learning, masked reconstruction, and physics-informed objectives tailored to wireless signals. AM-FM's architecture features adaptive frequency aggregation and relative temporal encoding to handle heterogeneous subcarrier quality and translation-invariant activity patterns. Evaluated on nine diverse downstream tasks, including fall detection, human activity recognition, localization, and gesture recognition, AM-FM demonstrated strong cross-task performance, achieving over 0.90 AUROC on all classification tasks, and significantly improved data efficiency compared to models trained from scratch.

Key takeaway

For research scientists developing ambient intelligence systems, AM-FM demonstrates that a single, pre-trained foundation model can replace numerous task-specific models, significantly reducing the need for labeled data and improving generalization across diverse environments and devices. You should consider adopting a foundation model approach for WiFi sensing to achieve robust, scalable, and privacy-preserving perception, especially for applications requiring cross-environment robustness or few-shot learning capabilities.

Key insights

Foundation models can enable scalable ambient intelligence by learning general-purpose representations from unlabeled WiFi CSI data.

Principles

• Large-scale self-supervised pre-training yields transferable representations.
• Domain-specific objectives are crucial for unique signal properties.
• Data diversity enhances generalization across devices and environments.

Method

AM-FM uses contrastive learning, masked reconstruction, and physics-informed autocorrelation prediction on WiFi CSI, with adaptive frequency aggregation and relative temporal encoding.

In practice

• Use existing WiFi infrastructure for ambient intelligence.
• Apply bottleneck adaptation for complex, fine-grained tasks.
• Temporal probing is effective for linearly separable attributes.

Topics

• Ambient Intelligence
• WiFi Sensing
• Foundation Models
• Channel State Information
• Self-Supervised Learning

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• The 10 most viewed blog posts of 2025 — Foundation models are expanding beyond language to transform forecasting, quantum computing, robotics, and data processing.
• Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models — Embodied-R1.5 unifies diverse embodied reasoning capabilities into an 8B-parameter EFM, achieving SOTA performance and real-world generalization.
• Identifying People Using Wi-Fi Routers — Standard Wi-Fi routers can identify individuals by analyzing their unique walking styles through unencrypted signal propagation data.
• Sergey Levine - Building LLMs for the Physical World - [Invest Like the Best, EP.465] — The core idea is building general robotic foundation models for universal control of embodied systems.
• How to Use a Time Series Foundation Model for Anomaly Detection — A single, general time series foundation model can replace many narrow, problem-specific anomaly detectors.
• LatentWave: JEPA Pretraining for Wireless Foundation Models — JEPA pretraining for wireless foundation models yields transferable representations by predicting latent space masks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.