Zero-Fi: Zero-Shot Wi-Fi-Based Human Activity Recognition via Contrastive Signal-Language Alignment

Traditional Wi-Fi-based human activity recognition (HAR) systems are limited by a closed-set assumption, requiring labeled training data for every activity they intend to recognize. This is impractical given the diversity of human behavior. Zero-Fi addresses this by enabling zero-shot recognition, allowing a system to identify activities it has never encountered during training.

The authors propose a contrastive signal-language alignment framework. Instead of mapping Wi-Fi signals to discrete class labels, Zero-Fi maps them to a semantic embedding space shared with natural language. To achieve this, the researchers:

• Denoise Wi-Fi signals: They use conjugate multiplication between antennas to cancel out hardware-induced phase noise and a domain discriminator to suppress environment-specific signal variations.
• Extract multi-modal features: The system processes Doppler frequency shifts, phase, and amplitude using dedicated Transformer-based encoders to capture comprehensive motion dynamics.
• Leverage LLMs: They use large language models to decompose activity labels into detailed, attribute-level descriptions (e.g., specific body part movements), which provides a richer semantic bridge between Wi-Fi signals and unseen activity categories.

Zero-Fi demonstrates that by aligning Wi-Fi sensing with language, models can generalize to new activities without requiring additional labeled Wi-Fi samples. In experiments on large-scale public benchmarks, the framework achieved an average accuracy of 69.58% on held-out activity classes. This approach significantly lowers the barrier for deploying HAR systems, as it eliminates the need for exhaustive data collection for every new activity category, making Wi-Fi sensing more scalable and adaptable to real-world environments.