My PhD thesis on monocular spacecraft pose estimation with deep learning: the domain gap between synthetic and real images, deployability on embedded hardware, uncertainty, and the design of a robotic hardware-in-the-loop testbed for validation. Supervisors: Paolo Tortora and Dario Modenini.

Within the ASI-funded uFASTER project, developed with IMT Srl, this work validates a forward motion compensation concept for CubeSat-class Earth observation payloads: a biaxial piezoelectric actuator micro-shifts the focal plane against the apparent ground motion. MTF and resolution measurements on a slanted-edge target show the system recovers about 90% of the sharpness lost to motion blur.

This paper presents the robotic hardware-in-the-loop testbed built to generate labelled real images of a satellite mock-up for machine-learning pose estimation. It covers the 7-DoF system, the ROS 2 control stack, the calibration procedures, and the measured 0.124 mm positional repeatability.

A pose estimation approach that adds a fail-safe fallback mechanism, keeping relative navigation reliable when the primary deep-learning estimate is not trustworthy.

This work applies stochastic variational Gaussian processes to monocular spacecraft pose estimation, to attach calibrated uncertainty to the predicted pose.

This manuscript highlights the positive impact of AI-based challenges on education and research at the u3S laboratory, University of Bologna, within aerospace engineering degree programs. It discusses three examples from recent academic years, showcasing the benefits and challenges of using AI competitions as hands-on learning tools for final projects.

Improving agricultural practices with AI and Earth observation is crucial for farm sustainability. Non-invasive soil parameter estimation using hyperspectral images can optimize fertilization. However, inconsistent validation across datasets hinders comparison. To address this, we present an AI-ready dataset with airborne hyperspectral images and in-situ soil measurements, along with a standardized validation procedure. This dataset was used in the HYPERVIEW challenge, aiming to advance soil analysis algorithms for satellite deployment. We discuss the challenge's outcomes and offer a reproducible dataset for future research in AI-driven soil analysis.

A close-range navigation method for orbiting a known but uncooperative resident space object, with attention to computational efficiency.

The University of Bologna and Nautilus, in collaboration with ESA, are developing a facility to test nanosatellites' attitude determination and control systems. This facility, to be installed in ESTEC, aims to verify closed-loop attitude control using physical sensor stimuli. The testing environment is expected to uncover issues not typically identified in the compressed verification campaigns of the New-Space approach. The manuscript provides insights into the project status, emphasizing key requirements and design choices.

This study addresses the challenge of pose estimation for in-orbit servicing missions using a monocular camera. Synthetic images are commonly used due to impracticalities in collecting real datasets in space, but a significant domain gap exists. The research explores the use of generative adversarial networks to make synthetic images more realistic.

The paper explores the application of vision Transformers, specifically Swin Transformers, for autonomous onboard pose estimation in space missions. Addressing challenges faced by neural networks, the study details an algorithm incorporating Swin Transformers and adversarial domain adaptation. Notably, this algorithm secured the fourth and fifth positions in the ESA's Satellite Pose Estimation Competition 2021. The research emphasizes the impact of larger models and data augmentations on accuracy. Additionally, a lightweight variant is proposed to overcome limitations without access to test images.

The paper introduces a pose estimation software for uncooperative space objects, leveraging monocular cameras and scalable neural network architectures. The software is designed for on-board implementation, compatible with low power Edge Tensor Processing Units. It achieves state-of-the-art accuracy on both the purposely developed Cosmo Photorealistic Dataset and the Spacecraft Pose Estimation Dataset.