Sensor Model Identification via Simultaneous Model Selection and State Variable Determination

This work addresses unattended sensor model identification for robotic localization using measurement data alone. We present an unsupervised gray-box approach that selects a single sensor model from a predefined catalog, without prior knowledge of the sensor type or training data. The method jointly determines which calibration states and reference frames are required, while enforcing one-out-of-many decisions within a continuous optimization framework. A dedicated health metric evaluates the reliability of the selected model and enables the detection of explainable false positives, supporting robust plug-and-play integration of sensors, tested with real-world localization systems.