I spent 2.5 years at Astroscale building computer vision for ELSA-M. The system had to look at a camera image and estimate where its Eutelsat OneWeb target was and which way it was facing so that another spacecraft could approach it.
Real orbital images of the target were scarce. I could generate synthetic training images with exact position and orientation labels, but the images still came from a simulator. Lighting, shadows, reflections, contrast and camera noise behave differently in space. A model can learn those artificial details along with the spacecraft.
Within the wider ELSA-M team, I built the pose-estimation system, investigated the synthetic-to-real gap and developed AstroGAN myself. By the time I left, the pose system was producing orientation estimates with about four degrees of error.
About three degrees was the launch-ready comparison point based on other systems. It was not a formal flight-readiness determination for ELSA-M, but it gave me a practical indication of how close the pose system was. At roughly four degrees when I left, it was approaching that point.
One part of the investigation used SPEED+. It pairs simulated and laboratory images of the Tango reference spacecraft from the PRISMA mission. It was a useful proxy for measuring the domain gap, not a dataset of ELSA-M's OneWeb target.
I compared the simulated and laboratory images under different lighting. Brightness, contrast, fine detail and edge information all shifted. Under the harshest lighting, much of the visual detail that the pose model relied on disappeared.
I then built AstroGAN by substantially adapting an open-source image-translation model for satellite imagery. It preprocessed synthetic images before pose estimation. On the proxy measures I used, it substantially reduced the measured gaps between simulated and laboratory images.
That was useful progress, but it was not the final test. A GAN can make synthetic and laboratory images look closer under proxy measures without improving the position and orientation estimates that the spacecraft needs. I left Astroscale before completing enough downstream work to establish whether AstroGAN reduced the pose error.
I thought the remaining gap was tractable. My next step would have combined GAN preprocessing with much heavier augmentation and further work on the pose model. I believe that programme could have brought the system close to the three-degree comparison point, but I did not run it and cannot present it as a result.
The measured domain-gap reduction told me that AstroGAN was changing the images in the intended direction. The pose error would have determined whether that change was useful to the system. Astroscale has published more about the ELSA-M and Eutelsat OneWeb mission.