Reinforcement Learning for Minimizing Age of Information in Real-time Internet of Things Systems with Realistic Physical Dynamics