It would seem to me that this is an oversimplification. Doing screen based video processing is not simple and is not exact especially where camera movements are independent of game dynamics as well as the difficulties inherent in doing fast frame classification- there has been some good work in this direction recently but far from perfect(yolo, ssd). The route deep mind took in partnering with the game manufacturers to gain inputs directly while receiving global information about game state seems the simplest for their goal of training rl agents. We considered doing a project around this but figured it was a massive iceberg.