Machine Learning for Docking (ML4DOCK)
Sample images from the Cosmo Photorealistic Dataset, developed in Blender as part of the project, prior to postprocessing.
Project overview

The project entailed a feasibility analysis on the use of Neural Networks for estimating the pose of a known non-cooperative satellite from a monocular image. I undertook this initiative immediately following my graduation from the Master's Degree in Aerospace Engineering, from February 2021 to June 2021. The research scholarship was co-funded by the University of Bologna and Thales Alenia Space Italia.

The key phases of the project included:

  • A comprehensive literature review.
  • The creation of a database of images.
  • The establishment of a simulation environment.
  • An accuracy assessment phase.

This project marked my initial venture into neural networks, pose estimation, and computer vision.

Throughout this endeavor, I acquired valuable skills, including:

  • Generating realistic spacecraft image datasets in Blender, while controlling parameters such as relative position, attitude, and illumination conditions.
  • Training neural networks using TensorFlow.
  • Utilizing cloud services to scale computing power efficiently.

Following the official conclusion of the project, the research scholarship was extended to facilitate further algorithm optimization and testing. This involved utilizing a Coral Dev Board Mini equipped with a Tensor Processing Unit (TPU) through quantized models.

📰 For more detailed information, please refer to the Engineering note published in the Journal of Spacecraft and Rockets.

Sample image from the Cosmo Photorealistic Dataset, prior and after postprocessing.