Fractional-Order Meta-Reinforcement Learning for Space Noncooperative Object Active Visual Tracking

This article primarily addresses the challenge of actively tracking noncooperative objects in aerospace, where the chaser continuously tracks noncooperative targets. Traditional approaches treat object tracking and chaser control as separate entities, and make system integration difficult. To overcome these challenges, we use fractional-order meta-reinforcement learning for active visual tracking (FracMRAVT) about noncooperative aerospace objects. FracMRAVT extracts information from the chaser's view image to generate corresponding actions and interact with the noncooperative objects. FracMRAVT demonstrates the capability for a better initial active visual tracking model compared to traditional reinforcement learning algorithms. FracMRAVT offers several advantages, including the ability to quickly adapt to other objects through the pervasiveness potential of meta-learning and memory properties of fractional calculus. Our algorithm demonstrates efficient task completion through minimal adaptation, even when dealing with other objects exhibiting different characteristics. Experimental results validate the efficacy of our algorithm, showcasing superior performance in scenarios involving both a few adaptations and nonadaptation.