I am somewhat of an amateur to this topic so I just want to put forward an idea and I would be happy to be educated what the current state of the art in that direction is.
One of the big questions is how to get an AI that is conscious. It seems to me the surest way to get that would be to train for it, rather then let it emerge by chance. To do that I would imagine to use problems that need one or more properties associated with a "self" to solve it.
E.g. catching a ball that was thrown. To do that the self needs to know its position in space relative to the ball thrown. Maybe an intermediate layer could be trained for the current position, while the output could be trained for the velocity and direction to go to catch the ball. This should help to establish the concept of ones own position in space for the AI. Of course there should be many different problems that help solidify the same property.
Many different properties that can be associated with a self should be trained for. The nodes that correspond to those properties should probably be closely connected, to establish some kind of self core where that information is compressed into.
One could start with problems that require one dimension of the self core to solve, respectively. Then move on to problems that require multiple ones. I suspect the higher the number of dimensions of the "self core" the closer it would get to a state we would recognize as consciousness.
My question would be, whether some kind of approach like that was ever investigated? Or alternatively, why this approach would not be feasable or would not work?