Edge-Deployed Unseen Spacecraft Pose Estimation

• Unité de recherche: SEFRI
• Numéro de projet: P2024-EPFL EDUSPE

Mots-clé (Anglais)

EPFL Mathieu Salzmann Edge-Deployed Unseen Spacecraft Pose Estimation SEFRI SERI SBFI SSO ARF

Description succincte (Anglais)

Funding is requested for Edge-Deployed Unseen Spacecraft Pose Estimation. The objective of this project is to transfer developments in Deep Neural Network (DNN) 6D pose estimation to the Swiss space sector; contextually aiming to support Swiss companies such as Clearspace. The project period is 1 December 2024 - 31 December 2026. 218,964 CHF is requested across two years, primarily allocated to an EPFL postdoc and subcontracting to Swiss partner Klepsydra.
This work directly supports the ESA Zero Debris Charter and the Active Debris Removal / In-Orbit Servicing (ADRIOS) project. A target-agnostic (or unseen) 6D pose estimation system is a key enabling technology for Active Debris Removal (ADR) and On-Orbit Servicing (OOS). A Swiss
consortium of EPFL, Clearspace and Klepsydra are uniquely positioned to support this. This project leverages current Swiss space industry strengths, while paving the way to exploit new developments in OOS.
In brief, five work packages are defined: (WP1) State-of-the-art unseen pose estimation DNN architectures will be evaluated; 3-4 candidates will be selected. (WP2) A fault-injection framework representative of the space environment will be created. The architectures’ sensitivity to injected faults will be assessed; 2 final architectures will be selected. (WP3) With industry partner Klepsydra, the architectures will be deployed to space-grade hardware. (WP4) Once deployed to the sample hardware, the software’s sensitivity to injected faults will again be assessed and tested in the ESA ESTEC radiation facilities. (WP5) The project will support education initiatives.

Objectifs du projet (Anglais)

The expected output of this project is a TRL5 6D pose estimation prototype prepped for transfer to the Swiss space sector (e.g. Clearspace); a publicly released recommendations report for deploying fault-tolerant 6D pose estimation architectures to space; and a space-environment fault-injection assessment pipeline. Understanding neural network failure modes when exposed to the space environment will support the space sector, both locally and abroad.