Mission planning for distributed multiple agile Earth observing satellites by attention-based deep reinforcement learning method

Abstract

The autonomous coordination and integrated planning of observation and data downlink missions for the distributed agile Earth observation satellite (AEOS) constellation hold significant importance in practical applications. In order to address this issue, we introduce an abstract universal mission model and present an algorithm rooted in deep reinforcement learning (DRL), termed the Attention-based Distributed Satellite Mission Planning (ADSMP) algorithm, which is designed to generate effective planning solutions. This algorithm employs a neural network that utilizes the attention mechanism, enabling each satellite to independently make decisions with equal intelligence. Furthermore, a mission priority adjustment method is devised to facilitate the coordination of data download and observation scheduling. The ADSMP is trained using the REINFORCE algorithm with Rollout Baseline. By conducting comparative experiments, we demonstrate that the proposed algorithm attains the highest revenue rate in corresponding scenarios, while simultaneously ensuring fast inference speed.