GANS today are used to improve generators.
In a GAN the generator competes with an authentacy identifier in order to generate something that is so good the identifier can't tell a fake from an original.
So with a GAN you are training the generator.
In a Clown and identifier adversarial network you are training the identifier on more general classification tasks. What the Clown does is try to fool the Identifier into missing something important with the kind of issues/tricks you get in the real world as to help the identifier have a more optimal fit for real world tasks like automated driving.
Using a clown should save a lot of training time and human effort when trying to fit the data best accounting well for all possibilities.
A clown can widen perspective and make an identifier system more aware with less human input.