Some mechanism to separate individual items, a conveyor belt, maybe two cameras and then train some neuronal network on classifying the images. Training data is

Some mechanism to separate individual items, a conveyor belt, maybe two cameras and then train some neuronal network on classifying the images. Training data is easily generated by feeding in a couple of boxes with known identical content. Not sure if you can get the neural network to generalize to before unseen dimensions, if not it will only work for sizes in the training set. Maybe some lights for more consistent lighting. Then you still need a bunch of containers to sort things into, maybe along the conveyor belt and pistons to push things into the correct containers or reject them because identification failed. If you want to cover a wide variety of sizes including very similar sizes, then it could maybe become hard for a neural network to get the sizes right? So maybe just let the neural network classify the items and orientation, then measure dimensions in the image, maybe using stereo images. This will however now probably require manually labeling training images. Do you also want to read the grade and specification? Do you want to deal with coated, old, rusty or dirty items? Interesting project, but that could take some time to get working reliable enough to be useful.