Navigation of autonomous Unmanned Aerial Vehicles (UAVs) in unknown, obstacle-cluttered environments requires motion planning under limited sensing and strict safety constraints. Although learning-based navigation approaches have demonstrated promising performance, many rely on purely reactive obstacle avoidance or indirect reward shaping (e.g., end-to-end learning or deep-learning based sensor encoding), without explicitly incorporating predictive collision information into the learning process. This work introduces a conceptual framework for learning collision-aware UAV navigation over a predicted motion horizon.