What Can I Not Do? Towards an Architecture for Reasoning about and Learning Affordances