Speaker
Description
This presentation details the development and verification of a vision-based autonomous Rendezvous and Proximity Operations (RPO) framework designed for satellite inspection and life extension missions. Driven by frameworks such as the ESA Deep Neural Network (DNN) project, this work supports preparation for an upcoming LEO in-orbit demonstration.
The core software stack utilizes a power-efficient AI pose estimation architecture based on target detection and keypoint localization. The pipeline outputs runtime quality metrics for Fault Detection, Isolation, and Recovery (FDIR) gating, feeding an Unscented Kalman Filter (UKF) for precise relative state estimation. For hardware prototyping, the system is tested on an embedded flight-representative processing platform.
Our high-fidelity Digital Twin Visual Simulator for Software-in-the-Loop (SIL) testing is compared against images from an in-house robotic testbed. This facility features a 13-degree-of-freedom robotic setup with custom Sun and Earth albedo illumination. Efforts included material randomisation, camera emulation and calibration, and lighting characterisation. The Learned Perceptual Image Patch Similarity (LPIPS) metric was utilised to demonstrate that the perceptual discrepancy between synthetic and laboratory imagery was minimised.