Experimental Verification of a Monocular CNN-Based Pose Estimation Algorithm for the SROC Mission

This work presents a Convolutional Neural Network (CNN)-based framework for monocular pose estimation of spacecraft in the context of the ESA Space Rider Observer CubeSat (SROC) mission. The mission, led by a consortium of Italian institutions, involves autonomous proximity operations including inspection and docking maneuvers with the Space Rider vehicle. Given the constraints of CubeSat-class hardware, we explore an efficient and adaptable CNN-based solution for relative navigation using only optical input. In order to address the challenges of domain gap between synthetic and real data, a high-fidelity synthetic dataset was generated using Blender with realistic lighting conditions and applying dataset augmentation techniques. Real image datasets were acquired at ESA’s GRALS facility with a COTS camera to test the trained models with a 1:20 mockup of Space Rider. A hyperparameter optimization (HPO) was conducted using the Optuna framework to enhance training performance. Multiple training strategies were evaluated, including fine-tuning techniques and dataset mixing approaches, to enhance performances on a specific maneuver that requires greater accuracy. The trained networks were deployed on the Payload Processing Unit using the ONNX Runtime framework for inference. Tests demonstrated consistent and reliable performance across multiple scenarios. These results confirm that, when supported by domain adaptation and hyperparameter optimization, the proposed CNN architecture can achieve accurate pose estimation under limited computational resources, making it a strong candidate for future autonomous CubeSat applications.