Vision-Only-Based Control of Approaching Disabled Satellites via Deep Learning

Removing disabled satellites is essential to the efficient utilization of orbital resources. Approaching a target satellite is one of the critical stages of the whole removal process, requiring the chaser satellite to perform accurate control. In this article, we propose a method to establish a neural-network controller that utilizes an optical camera as the sole relative measurement device. To achieve this, we first create a set of optimal approach trajectories and generate a numerical dataset. We then modify this dataset to instruct the neural-network controller to generate additional corrective forces when the relative velocity deviates from its optimal value due to unexpected disturbances. By making use of the modified dataset and 3-D simulations, we create image sequences that are employed as training samples in deep learning. Finally, the neural-network controller established based on the 3D-ResNet-18 architecture is trained and obtained. The simulation results suggest that our approach significantly improves control accuracy under thruster output uncertainty.