DeepRL Sim-to-Real

Project FORCIS investigates solutions to the challenging problem of few-shot object detection. FORCIS leverages synthetic training data generated by simulation engines which are dynamically parameterized to maximize utility to a downstream deep-learning algorithm. We explore this approach, inspired by advances in the Reinforcement Learning (RL) branch of machine learning (ML), and contrast it against other advanced ML concepts aimed at addressing few-shot detection. Our approach on FORCIS leverages a flexible simulation environment based upon a modern game-development engine (e.g. Unreal or Unity). A programmatic interface to this simulator allows an optimizing agent to control the probability distributions of a wide range of simulation parameters. The control agent is designed to explore this high-dimensional parameter-space in order to maximize the performance of a mission-specific computer-vision “Main Task Model” (MTM). During this Phase II Enhancement, we propose to fully mature and apply the innovative techniques established under the prior phases, demonstrating their usage against real flight-test events.