The PraktikApp project, running from 1 October 2019 to 31 December 2022, aimed to create a “smart” table that could recognise objects placed on it and trigger corresponding applications, thereby enriching everyday domestic practices. The core technical contribution came from the company Nyris, which was responsible for developing and integrating machine‑learning models that detect avatars and determine their positions on the table. The project was carried out in close cooperation with the University of Siegen, Hochschule Düsseldorf (HSD), Häfele SE & Co KG, IOX GmbH, tennagels Medientechnik GmbH, and spek Design, with regular bi‑weekly teleconferences and annual in‑person workshops to coordinate progress.

Nyris delivered avatar‑recognition models in three successive phases. In the first year, a prototype was built that could recognise conical and cylindrical avatars. The initial training set was only partially annotated, and the model frequently confused cylinders with cones, especially from a top view, because the base areas were identical. Efforts to improve detection by enlarging the avatars in the second half of 2020 did not yield the expected gains, so new avatars were designed by spek Design. The dataset was re‑annotated with bounding boxes and the model was retrained multiple times, but limited data still constrained performance.

The most significant technical progress occurred in 2021 during the MK1 and MK2 phases. New avatars were introduced with distinctive visual features to enhance recognisability. After thorough annotation and cleaning, the data were formatted into JSON for training. The MK1 model achieved a mean average precision (mAP) of 96.5 % with an average inference time of 0.410 seconds. The subsequent MK2 model improved to 97.3 % mAP and reduced inference time to 0.245 seconds. MK2 proved robust under low‑light conditions and could detect small objects that were often invisible to the human eye. Further refinements included reducing the intensity of white light beneath avatars, which lowered false‑positive detections and further boosted accuracy.

In 2022, the project expanded the avatar set by adding new designs and removing two older ones. The updated model was retrained on the expanded dataset, which now included twelve new class labels such as “Ferris Wheel,” “Spider,” and various “memorabilia” states. The annotation process followed the same pipeline as earlier phases, ensuring consistency across the dataset. The final model maintained high precision while handling the increased class diversity.

Throughout the project, Nyris worked closely with HSD, which coordinated the technical integration of the overall system, and with spek Design, which supplied the avatar designs that the models needed to recognise. The University of Siegen led the project, organising the bi‑weekly meetings and annual workshops that facilitated continuous information exchange among all partners. The extended three‑month period at the end of 2022 allowed for final testing and validation of the smart table’s sensor‑based event triggering and application management capabilities.

In summary, the PraktikApp initiative successfully combined embedded sensor technology with advanced object‑recognition models, achieving a 97.3 % mAP and sub‑quarter‑second inference times. The collaborative effort among academic institutions, industry partners, and design studios produced a functional prototype of a smart table that can detect and respond to a diverse set of avatars, paving the way for future applications in domestic environments.