I suspect you and GP are talking at different layers. I think you are using robust on specific tasks with measurable confusion matrix. I think GP is talking about robust in more complex and diverse workflows, with the ability to self correct over turns.

Either, please correct me if I'm misinterpreting