Alors que l'exploration spatiale humaine n'en est qu'à ses balbutiements, l'humanité a déjà pollué les orbites de la Terre avec de grandes quantités de débris spatiaux. En 2011, plusieurs agences spatiales (NASA, ESA, CNES) ont recommandé l'élimination de ces débris, qui est désormais au sommet de leurs priorités. Cela s'est récemment concrétisé par le projet ClearSpace-1, financé par l’ESA et mené par la startup suisse ClearSpace, qui conduira à une mission de capture d’un débris en 2025.
L'élimination de débris dépend de deux facteurs. Au sol, il faut surveiller la trajectoire et la vitesse de rotation des débris afin de choisir la meilleure stratégie de capture. En orbite, il faut estimer la pose (position et orientation) relative du débris par rapport au satellite de capture de manière à les synchroniser. Ce projet BRIDGE Discovery contribuera à ces deux aspects afin de faciliter l'élimination de débris spatiaux. Nous explorerons les axes suivants :
- Amélioration des bases de données de débris. Le contenu des bases existantes est limité, et nous leur ajouterons des trajectoires précises et de nouvelles informations, telles que la taille, la réflectivité, et la rotation des débris.
- Estimation de la pose de nouveaux objets. Alors que ClearSpace-1 se concentre sur un débris spécifique, nous visons à un service de capture général, capable d’estimer la pose de n’importe quel débris grâce à des modèles d’apprentissage profond.
- Généralisation à de nouvelles conditions. Bien qu'ayant été testés uniquement sur Terre, nous développerons des algorithmes permettant à nos modèles de fournir des résultats fiables dans l'espace.
- Développement de modèles compacts. Nous automatiserons le design de nos modèles afin qu’ils satisfassent des contraintes de budgets reflétant la puissance de calcul disponible.
Notre recherche élargira l'applicabilité de la technologie développée pour la mission ClearSpace-1 et augmentera son impact sur le marché de la surveillance et de l'élimination des débris spatiaux. Notre collaboration avec ClearSpace facilitera son transfert vers l'industrie.
As stated by Stephen Hawking “... the human race has no future if it doesn’t go to space”. Unfortunately, while human space exploration remains in its infancy, mankind has already contaminated Earth’s orbits with vast quantities of satellites and spacecraft debris, thus seriously compromising long-term space operations. As of Feb. 2020, ESA estimates the number of debris measuring more than 10 cm in diameter to be greater than 34’000, about 22’300 of which are tracked by space surveillance networks. In 2011, several space agencies (NASA, ESA, CNES) pulled the red flag and recommended active disposal of space debris at the rate of 5-10 objects per year by 2020. For Low-Earth-Orbit (LEO), active disposal means launching a servicer system that will locate the debris, rendezvous with it, capture and de-orbit it, i.e., put it on a course of Earth atmospheric entry. While this recommendation has not yet become a reality, debris disposal is now at the top of the agencies’ priorities. Specifically, this has recently materialized into the ClearSpace-1 project, funded by ESA and kick-started by an Innosuisse Impulse grant. Led by the Swiss startup ClearSpace, this project, in which both PIs of this proposal are involved, will lead to a mission to capture the 2m-size upper part of the VESPA adaptor in 2025.The success of active debris disposal depends on two key factors. First, on the ground, one needs to monitor the trajectory and rotation speed of the debris so as to put the capture satellite on the correct course and choose the best capture strategy for the observed tumbling rate. Second, in orbit, one needs to estimate the debris relative position and attitude, i.e., 6D pose, w.r.t. the capture satellite, so as to synchronize their motion. This BRIDGE Discovery project will make breakthrough contributions to both aspects so as to facilitate generic debris disposal, thus broadening the applicability of the technology developed for the ClearSpace-1 mission and increasing its impact on the space debris monitoring and removal market.
To this end, we will investigate the following research directions:
- Improving debris knowledge databases. While a number of databases record the orbit parameters of satellites, their content is limited, with accurate orbital parameters only for objects that are actively monitored, e.g. by the United States Space Command, International Laser Ranging Station or the International Scientific Optical Network. We will enhance such databases by adding precise debris trajectories and implement new information, such as size, CAD model and reflectivity, from a combination of literature and astronomical imaging data. Moreover, when the debris tracks in the imaging data are sufficiently bright and with sufficient timeline information, we will aim to build a light curve that reveals information about the debris rotation period.
- Pose estimation for new objects. While ClearSpace-1 focuses on the VESPA upper part, whose 3D model is available prior to the mission, a viable debris removal service needs to generalize to previously-unseen debris, of potentially unknown or altered shape, so as to maximize the chances of success and reduce the costs of each new mission by facilitating the servicer deployment. Existing 6D pose estimation methods, however, cannot handle this scenario. We will therefore develop new deep learning models that generalize to objects unseen during training.
- Domain generalization for pose estimation. Despite having been trained and tested on Earth only, our 6D pose estimation networks will have to yield reliable results in space. To prevent having to perform domain adaptation during every mission, as is currently planned for ClearSpace-1, we will develop domain generalization algorithms that learn domain-invariant 6D pose representations and thus can be applied to new conditions without any re-training.
- Learning compact networks for pose estimation. While the network being developed for ClearSpace-1 will fit in the limited on-board hardware resources, it will result in low frame-rates and most likely be sub-optimal for the given computational budget because of its manual design. To address this, we will develop neural architecture search algorithms dedicated to compact networks that fit in a given budget, and corresponding training strategies for such networks.
In the spirit of open research, most of the data generated in this project will be made publicly available. Our research, aiming to ease the market deployment of active debris removal, will leverage the complementary expertise of both PIs in computer vision, machine learning, space engineering and astronomical observations. Our collaboration with ClearSpace will further facilitate its transfer to the industry.
Last updated:22.04.2022
