Towards Robust And Fair Vision Learning In Open-World Environments