Neural Networks for Minimum Time Satellite Reorientation

This work investigates the application of imitation learning, and specifically behavioural cloning, to solve the time-optimal satellite reorientation problem. Traditional optimal control methods, while accurate, are often computationally demanding and unsuitable for real-time execution on resource-constrained systems. To address this limitation, we propose a learning-based framework where neural networks are trained offline on expert trajectories generated with the GPOPS-II optimal control solver. Several architectures are explored, each designed to infer optimal control commands from state error inputs. The models are trained on a dataset of state-action pairs and evaluated in terms of manoeuvre performance and inference speed. Importantly, the trained networks are tested on commercial off-the-shelf CPU-only platforms, in order to demonstrate their suitability for embedded applications. Results show that the proposed models are capable of replicating the time-optimal behaviour while achieving inference times in the order of milliseconds, even on low-power hardware.