As I read that «APOPO’s HeroRATs take around nine months to fully train», the preliminary thought struck, about whether cybernetic deminers, for which training is an "upload", can be worth considering. Maybe the time to build a robot and its costs are competitive. It seems that the more complex part could really be the motility - the landmine sensor may not be something you train, if the revealing substance remains TNT, or the problem seems to remain deterministic, if progressively building a database of all the diverse threats is a goal. On the other hand, other techniques that involve e.g. visual clues and other open factors for decision may enhance the results, and call again for dynamic "training".
Also note that the inspiring entity seems to use automation already in their processes: rats can use automated training rooms (they are rewarded for TNT detection) - «The rat's response is measured by optical sensors and the cage produces an automated click sound with food delivery».
I find the matter interesting and "important" as an understatement, and maybe some good result could come from opening the topic.
--
¹And, in a subsequent effort built on the experience on landmines, tubercolosis.
Aerial detection seems to be the way to go.
Also: Honeybee-based biohybrid system for landmine detection
https://www.sciencedirect.com/science/article/pii/S004896972... [..]Highlights • Honeybees and organic semiconductor films can indicate explosive contamination.
• Conditioned honeybees can subsequently be used to locate explosive plumes.
• Reconditioning prolonged interest of honeybees for target odour.
• Camera-equipped UAVs can remotely track the honeybee trajectory.
• Computer vision algorithms can detect landmines by analyzing honeybee movements in video.[..]