Single-Antenna ESP32-S3 Wi-Fi Angle Estimation Limits

Low-cost ESP32-class channel state information (CSI) capture enables embedded Wi-Fi sensing, but obtaining reliable azimuth-estimation results requires strict evaluation discipline and repeated-session validation. This paper investigates a single-antenna ESP32-S3 setup for azimuth sensing in a controlled line-of-sight environment, positioning it as a compact reference for broader low-cost embedded sensing research. The dataset comprises 4,796 valid packets collected across 16 capture subsets, two experiments, and eight nominal angle positions. Four lightweight regression approaches are evaluated: RSSI linear, RSSI polynomial, CSI-based k-nearest neighbors (kNN), and a fused CSI+RSSI kNN model. The evaluation follows an experiment-wise protocol in which models are trained on one experiment and tested on the other, ensuring full out-of-fold assessment across all angles. The CSI pipeline employs summary statistical features and train-only principal component analysis (PCA). Results from pooled out-of-fold predictions show that CSI-kNN achieves the best performance, with a mean absolute error (MAE) of 73.8° and a within-10-degree accuracy of 28.1%, while CSI+RSSI-kNN follows closely with 75.4° MAE and 27.3%. Angle-wise analysis reveals strong sector dependence: although the geometry captures exploitable directional structure at certain azimuths, errors remain large in several regions and do not support stable practical azimuth estimation. These findings establish quantitative bounds for the considered setup and provide a reproducible reference for future research on embedded azimuth sensing, improved sensing geometries, coordinated low-cost receiver configurations, and more rigorous repeated-session evaluation protocols.