Advanced Optimization Techniques for CNN-based visual Navigation in Rendezvous and Docking Operations

Accurately estimating a spacecraft’s position and orientation during proximity operations—such as rendezvous and docking—is critical to ensuring the success and safety of activities including formation flying, on-orbit servicing, and active debris removal during space missions. Depending on the specific mission scenario, the pose determination task entails both theoretical and technological challenges, including the development of robust navigation algorithms. Facing this challenge, the present study proposes an optimized CNN-based architecture for monocular spacecraft pose estima tion. This approach is designed to address the limitations of current deep learning-based approaches by leveraging advanced optimization techniques to improve the accuracy and reliability of estimation under various operational con ditions. Hyperparameter optimization is performed within the Optuna framework, integrating two distinct algorithms: the Tree Parzen Estimator (TPE) sampler dynamically explores and exploits the search space, finding the best set of hyperparameters that significantly improves the network performance, while the Asynchronous Successive Halving al gorithm (ASHA) pruner prematurely trims the more uncompromising combinations in early stages of training, thereby reducing computational overhead. Model compression methods are also used with the objective of reducing the size of the architecture, typically composed of millions of parameters, eliminating redundancies that minimally contribute to the final prediction for deployment on resource-constrained hardware while improving performance and reducing com putational demands. Complementing these methodological advances, multiple custom synthetic datasets are generated using Blender® to simulate distinct, mission-critical scenarios associated with proximity operations. These datasets are employed for training, validation, and testing, ensuring the robustness of the proposed network across varying il lumination conditions and orbital backgrounds. The obtained results, shown and discussed in the paper, demonstrate positional estimation accuracy on the order of centimeters while maintaining angular accuracy within degree limits. The proposed optimization techniques not only preserve the model’s predictive accuracy but also ensure robust performance even with lower input resolutions without compromising output quality. This achievement underscores the feasibility of sustaining high-performance levels in resource-constrained environments while significantly reducing computational costs and convergence time, opening pathways for the deployment of advanced, accurate, and efficient machine learning solutions in the challenging operational environments of space missions.