Description
The single most-cited limitation of WiFi-CSI sensing: a model trained in one environment, on one user, with one piece of furniture in one position, degrades sharply when anything about that setup changes. Same room, different time of day. Same task, different room. Same room, same time of day, but a chair has been moved. This is the central practical obstacle the BLE-calibrated CSI thesis is designed to attack — periodic BLE calibration campaigns produce ground truth in the current environment, so the CSI inference model is recalibrated continuously rather than relying on a one-shot training that decays.
Why it's hard
- Multipath propagation is exquisitely sensitive to scatterer geometry; CSI is a high-dimensional fingerprint of the room, not just the people in it.
- The change one cares about (people moving) is small relative to the changes one does not care about (door positions, humidity, antenna orientation).
- Per-environment retraining is expensive and infeasible at deployment scale.
- The problem is unidentified at training time — there is no signal in the training distribution that warns the model it will be deployed in a shifted distribution.
- Standard ML "generalization" theory assumes IID; CSI deployment is fundamentally non-IID.
Common approaches
- Domain-adaptation and domain-generalization methods (adversarial, prototype, contrastive).
- Few-shot / meta-learning to amortize per-site adaptation cost.
- Physics-informed feature engineering (CSI ratio, BVP, Doppler) to suppress environment-specific components.
- Self-supervised pretraining on large unlabeled CSI corpora plus environment-specific fine-tuning.
- Periodic ground-truth calibration campaigns (the thesis approach) — use a complementary sensor (BLE) to label CSI in the current environment.
Source Papers
- chen2023_5cbd ↗ — Cross-Domain WiFi Sensing with Channel State Information: A Survey.
- wang2026_2758 ↗ — Wi-Fi Sensing Generalizability taxonomy survey.
- guarino2026_e72c ↗ — CSI-based WiFi sensing datasets + reproducibility (cross-domain emphasis).
- wang2015_48cf ↗ — Understanding and Modeling of WiFi-Signal HAR (foundational for why environment matters).
- meegahapola2024_a321 ↗ — M3BAT unsupervised domain adaptation for multimodal mobile sensing.
- jiang2018_77f6 ↗ — Towards environment-independent device-free HAR.
- hou2023_bf83 ↗ — DASECount: domain-agnostic few-shot wireless counting.