In recent years, artificial intelligence has made important steps towards the development of agents that solve tasks at a human level. However, it is still difficult for artificial agents to solve complex tasks that require understanding, reasoning, and planning over longer timescales. In this project, I envisage intelligent agents that learn deep temporal models of the world and use them to plan long-term actions, grounding on the active inference neuro-inspired theory for perception, cognition, and control. In its development, the project will address several research questions, such as how to learn a hierarchical world model from environment observations, how to exploit the model to plan actions, and how to resolve uncertainty by seeking or retrieving information. All the approaches developed will be implemented with modern machine learning techniques and compared with state-of-the-art techniques from the literature.