Considering the success of neural network image classifiers, it seems garbage sorting would be a slam dunk application for these reasons:
(1) the training data is readily available. there are human sorters who already are classifying everything in real-time as they pick stuff off the conveyor and move the item to the right bin. train cameras on the conveyor belt and the analyze which items are picked out and which bin they get tossed in.
(2) the NN doesn't have to be anywhere near as accurate as, say, a self-driving car. tune the classifier to minimize false positives. If the resulting NN can classify only 70% of the items, but accurately, that means 70% of the garbage stream can be automated and run 24/7. Humans can still go through the remaining 30% if it economically makes sense.
(1) the training data is readily available. there are human sorters who already are classifying everything in real-time as they pick stuff off the conveyor and move the item to the right bin. train cameras on the conveyor belt and the analyze which items are picked out and which bin they get tossed in.
(2) the NN doesn't have to be anywhere near as accurate as, say, a self-driving car. tune the classifier to minimize false positives. If the resulting NN can classify only 70% of the items, but accurately, that means 70% of the garbage stream can be automated and run 24/7. Humans can still go through the remaining 30% if it economically makes sense.