Your examples help Pip choose a path.
Changing a label changes what Pip learns. Game encounters use other photos.
The child labels real reference photos, then Pip encounters different photos. A local MobileNet model extracts image features; a nearest-neighbour classifier learns the child’s labels. Predictions depend on those labels, so mistakes affect Pip’s fictional adventure.
Keep the real-world rule separate from the model: never approach, touch, collect, move or investigate a suspicious or unfamiliar object. Stay away and tell a trusted adult. Do not use this game or its model to identify hazards in the field.
The robot may power down in the game. There are no graphic injuries. Choosing a detour is always acceptable, including when the AI is unsure.
Use varied training examples; compare the nearest examples after a mistake. A small training set cannot identify every explosive item. Museum and reference photos differ from damaged, buried or partly hidden objects.
Protective guidance: MAG risk education · UNICEF Lebanon awareness materials. These are optional sources for adults; gameplay needs no internet. Mabni is independent and is not endorsed by those organisations.
All photographs and model files are bundled locally. The reference library includes several countries and museum images; it is not a catalogue of every danger in South Lebanon. Labels, embeddings and progress stay on the classroom computer.
Full photo attribution · Model licence and provenance
