This paper evaluates an archival Bluetooth Low Energy received signal strength indicator (RSSI) dataset collected for bin localization on an industrial shop floor. The installation used 24 fixed relays positioned on walls and Bluetooth beacons attached to bins, with each relay applying a Kalman filter to its own readings in firmware before publishing them. The primary survey produced 19,034 usable two-minute fingerprint samples across 90 known positions, with 23 relay measurements available as model features. We compare calibrated path-loss multilateration with two radio-map estimators: distance-weighted $k$-nearest neighbours and Random Forest regression. A random 70/30 row split measures recognition at surveyed positions, while a coordinate-grouped split reserves 27 entire positions for testing and measures interpolation to unseen positions. Random Forest achieved a median error of 0.66~m for surveyed positions, with 77.9% of estimates within 3~m and 96.3% within 10~m. For unseen positions, median error increased to 9.42~m and 55.5% of estimates were within 10~m. Path-loss multilateration produced a 19.41~m median error on unseen positions, close to the 19.62~m mean-position baseline. The results support RSSI fingerprinting for a fixed inventory of surveyed bin positions, but not precise continuous localization between survey points. Git history dates the original implementation to September 7, 2019. The evaluation code in this paper was written in October 2026 and is reported separately from the seven-year-old prototype.

翻译：暂无翻译