Development of a Robotic Testbed and Initial Testing for Machine Learning-Based Spacecraft Pose Estimation

Abstract

Abstract. Vision-based navigation is a key enabling technology for autonomous in-orbit operations. However, current deep learning-based pose estimation algorithms face a critical sim-to-real gap, where performance degrades significantly when transitioning from synthetic training images to real ones. This has driven the need for ground-based validation facilities to bridge this gap. This paper presents a robotic testbed installed at the Microsatellites and Space Microsystems Laboratory of the University of Bologna for hardware-in-the-loop validation of vision-based spacecraft pose estimation algorithms. The facility features a 7-DoF robotic system, an optical table with a spacecraft mockup, a sun simulator, and an RGB camera, all integrated within a controlled dark room environment. The software architecture is based on ROS 2, with custom nodes enabling coordinated control of all system components. An initial application demonstrates the facility's capability through the development of a dataset generation pipeline for the ALMASat spacecraft mockup. While designed within constrained budget and space requirements, the testbed provides high precision and supports both fundamental research and educational activities in vision-based spacecraft navigation.