The rapid expansion of the orbital population and the increasing demand for on-orbit servicing have placed autonomous proximity operations at the forefront of space research. A primary bottleneck in navigating around non-cooperative or unknown targets is the lack of a priori geometric information, such as CAD models or fiducial markers. While traditional computer vision methods rely on multi-view consistency or depth measurements, the recent advent of large-scale generative AI offers a paradigm shift toward generalized 3D inference from extremely limited data. In this talk, we introduce DreamSat, an open-source, model-agnostic pipeline designed for the single-view 3D reconstruction of spacecraft and asteroids. We will discuss the first implementation of fine-tuned large-scale 3D generative models specifically for the space domain, demonstrating how these models can synthesize high-fidelity digital twins from a single monocular RGB image. We evaluate the reconstruction fidelity across diverse synthetic datasets comprising high-quality NASA and ESA spacecraft models, as well as asteroid geometries. Furthermore, we showcase the practical applicability of the reconstruction pipeline through qualitative assessments using real-world in-orbit imagery of non-cooperative targets, highlighting the system’s ability to handle actual flight conditions and harsh lighting. Building on this geometric foundation, we present DreamSat-Pose, a framework that reformulates 6-DoF pose estimation into a cross-modal correspondence problem. By leveraging frozen vision transformers for 2D feature extraction and graph-based networks for geometric encoding, we show how these machine-learned proxies can effectively substitute for ground-truth CAD models to achieve state estimation. We evaluate DreamSat-Pose through benchmarking on the SPE3R dataset, demonstrating reliable pose estimates using only a single image and reconstructed geometry and strong generalization to unseen spacecraft. Finally, we will briefly discuss DreamSat-Bench, an ongoing effort to develop a modular software- and hardware-in-the-loop robotic testbed. This facility aims to bridge the sim-to-real gap, providing a rigorous environment for maturing autonomous vision-guided navigation in the complex and unpredictable modern orbital environment.
Speaker:
Dr. Giovanni Lavezzi is a Research Scientist in the Department of Aeronautics and Astronautics at the Massachusetts Institute of Technology (MIT), where he joined the Astrodynamics, Space Robotics, and Controls Lab (ARCLab) in 2022. He received a Bachelor’s degree in Aerospace Engineering (2016) and a Master’s degree in Space Engineering (2018) from Politecnico di Milano, followed by a Ph.D. in Mechanical Engineering from South Dakota State University (2022). Dr. Lavezzi possesses interdisciplinary expertise in nonlinear and optimal spacecraft attitude control, astrodynamics, multi-body trajectory optimization, nonlinear and sequential convex programming, space sustainability, and artificial intelligence. He is a primary developer for the MIT Orbital Capacity Assessment Tool (MOCAT), utilizing source-sink evolutionary models to investigate long-term orbital sustainability, orbital capacity and policy interventions. His research further explores the intersection of AI and space operations, where he leveraged and adapted deep learning techniques for early time-series classification of space objects from astrometric data and engineered DreamSat, a novel 3D spacecraft and asteroid reconstruction framework utilizing fine-tuned generative diffusion models. At MIT, he manages a portfolio of concurrent research projects and has contributed to numerous research proposals, successfully securing funding for space sustainability and policy studies. He has authored 17 peer-reviewed journal articles and 23 conference papers. An experienced educator and mentor, Dr. Lavezzi has served as the Instructor for graduate-level Astrodynamics and a Co-Instructor for undergraduate Dynamics at MIT, and has supervised over a dozen graduate and undergraduate students. Since 2020, he has been an active technical peer reviewer for leading journals, including AIAA, IEEE, and Springer.