視線が通っても点検できたとは限らない、撮影画像の不確かさで宇宙機の観測位置を選択

Illumination-Aware Active Perception for Spacecraft Inspection

Harsh and rapidly varying orbital illumination introduces failure modes during spacecraft inspection. Existing planners rely on line-of-sight inspection metrics which perform occlusion checks against a known target and conflate geometric visibility with task-relevant information. These metrics overlook illumination effects including eclipse, self-shadowing, and specular reflection from spacecraft materials and tend to declare components as inspected despite the corresponding images lacking sufficient information. We propose an inspection-planning framework that replaces geometric visibility with a learned measure of camera information derived from a classifier variational autoencoder (VAE) that operates directly on the captured image. The encoder model is trained to classify spacecraft surface components from rendered patches, and its total latent variance is used to learn the varying illumination of said components, replacing binary visibility checks with a continuous, image-grounded quantity. We evaluate the approach by implementing a Guidance, Navigation, and Control architecture in a photorealistic physics simulator, and vary lighting and viewing geometries with multiple orbits. We show that, although all approaches successfully maximize their own objective, our approach results in better illumination: a mean brightness of $0.57$, with $81.6\%$ of observations above $0.25$, compared with $0.02$ ($3.0\%$) for geometric planning with an illumination band and $0.04$ ($6.0\%$) for purely geometric planning.