- Abstract: The development of machine learning approaches can benefit significantly from the availability of high quality simulation environments and from architectures that allow an easy integration of the learners and simulation environments. The main research benefits of the availability of such learning-simulation settings are the enabling of fast architecting of the learning approaches, and exercising of infrequent or anomalous (edge/corner) cases that cannot be experienced and experimented with enough in real-world settings. In particular, preliminary machine learning experiments in highly regulated and controlled environments can benefit from a simulation setting tremendously until we, researchers, become familiar with the feature space, and the basics of the application problem. One such domain is airplane operations both on airport grounds (taxiing) and the air. The X-Plane simulation environment is an ideal simulation platform for that domain - for visual perception tasks in airport environments, and for decision making tasks. This paper describes an open source architecture for learning different tasks that enable the operation of airplanes, in conjunction with X-Plane. We focused primarily on developing an architecture that works for perception tasks, but have used this architecture and we discuss possible approaches for employing it in decision making tasks. We believe that opening our setting to the AI/ML, sensing, simulation, and pilots' communities will enable formulating. advancing, understanding better and solving a significant set of learning and decision tasks in airport and flight environments.
- TL;DR: An architecture and setting for developing and testing learning algorithms and decision algorithms.
- Keywords: learning architecture, supervised learning, reinforcement learning, simulation environment, MobileNetV2, X-Plane, XPlaneConnect