Towards the portability of knowledge in reinforcement learning-based systems for automatic drone navigation