Object sorting is a good robotics project because the world is constrained: objects arrive on a belt, lighting can be fixed and bins are known. That makes it much easier than asking a mobile robot to recognize everything in a room.
Start with color or marker-based sorting before training a model. Once the mechanics work, add a camera classifier. The biggest reliability upgrade is not the neural net; it is consistent lighting and an object spacing gate so the camera sees one item at a time.