Robust Autonomous Optical Navigation for Asteroid Proximity Operations

Abstract

Future asteroid missions require autonomous navigation to handle ephemeris uncertainties, irregular gravity, and communication delays. This thesis demonstrates that realistic simulation through data augmentation enables robust optical learning-based navigation for asteroid proximity operations. A pose estimation pipeline based on deep learning is developed for asteroid 433 Eros. An SSD MobileNetV2-FPN detector localizes the asteroid, an LPN-101 network identifies 129 keypoints, and a CEPPnP solver estimates the six-degree-of-freedom satellite pose from 2D-3D correspondences. Cross-domain evaluation of six synthetic datasets reveals that augmentation-based training is crucial for realistic scenarios. The augmented-trained model achieves optimal performances (meeting <10% position and <5° rotation requirements in 99.78% of cases), while non-augmented models fail against realistic data. Hardware-in-the-loop validation using a 1:100000 mockup confirms sim-to-real transfer with centimeter-level accuracy. Evaluation of actual NEAR-Shoemaker imagery validates the model's ability to generalize to authentic mission data without the need for retraining.