In collaboration with ESA’s Advanced Concepts Team, we published our work on end-to-end guidance and control networks in the June 2024 issue of Science Robotics.
This review presents the results of training end-to-end neural architectures for interplanetary transfers, planetary landings, and close-proximity operations. To test these architectures on real robotic platforms and boost confidence in their use for future space missions, drone racing is explored as an ideal gym environment. Drone racing shares with spacecraft missions both limited onboard computational capabilities and similar control structures derived from the desired optimality principle. Additionally, it involves different levels of uncertainties, unmodeled effects, and a significantly different dynamical timescale.
The paper summarizes the work at the Micro Air Vehicle Laboratory on end-to-end neural control of quadcopters (see paper list below). In this research, supervised learning and reinforcement learning have been employed to train guidance and control networks (G&CNets) to control the Parrot Bebop drone. A major challenge in bringing these controllers to life is the “reality gap” between the real platform and the training environment. To address this, we combine online identification of the reality gap with pre-trained corrections through a deep neural controller, which is orders of magnitude more efficient than traditional computation of the optimal solution.