This paper addresses the problem of robotic capture of an uncooperative spinning target spacecraft. To do so, a computationally lightweight and real-time implementable guidance, navigation, and control architecture that relies on deep learning as well as pseudospectral optimization is proposed and experimentally validated. Specifically, a convolutional neural-network-driven stereovision pose determination system is first combined with a deep-reinforcement-learning-based guidance algorithm and pose tracking controller to cancel the relative motion between a chaser platform and an uncooperative spinning target platform in real time. Then, real-time tracking of a pseudospectral-based optimal guidance law generated offline deploys a robotic arm while minimizing the overall attitude corrections required to keep the target in view. The integrated experiment carried out using Carleton University’s Spacecraft Proximity Operations Testbed (a state-of-the-art planar air bearing facility, introduced in this work) demonstrates the performance of the developed deep learning architecture.
Optimal Capture of Spinning Spacecraft via Deep Learning Vision and Guidance
2025-05-01
Article (Journal)
Electronic Resource
English
On Deep Reinforcement Learning for Spacecraft Guidance
AIAA | 2020
|ON DEEP REINFORCEMENT LEARNING FOR SPACECRAFT GUIDANCE
TIBKAT | 2020
|