Collision avoidance algorithms are of central interest to many drone\napplications. In particular, decentralized approaches may be the key to\nenabling robust drone swarm solutions in cases where centralized communication\nbecomes computationally prohibitive. In this work, we draw biological\ninspiration from flocks of starlings (Sturnus vulgaris) and apply the insight\nto end-to-end learned decentralized collision avoidance. More specifically, we\npropose a new, scalable observation model following a biomimetic\nnearest-neighbor information constraint that leads to fast learning and good\ncollision avoidance behavior. By proposing a general reinforcement learning\napproach, we obtain an end-to-end learning-based approach to integrating\ncollision avoidance with arbitrary tasks such as package collection and\nformation change. To validate the generality of this approach, we successfully\napply our methodology through motion models of medium complexity, modeling\nmomentum and nonetheless allowing direct application to real world quadrotors\nin conjunction with a standard PID controller. In contrast to prior works, we\nfind that in our sufficiently rich motion model, nearest-neighbor information\nis indeed enough to learn effective collision avoidance behavior. Our learned\npolicies are tested in simulation and subsequently transferred to real-world\ndrones to validate their real-world applicability.\n