STAN: A Spatio-Temporal Attention Network for Space Debris Multistage Collision Avoidance

Liwen Yang, Xue Bai, Xiaoyi Wang, Ming Xu

Keywords: Space Debris Collision Avoidance, Deep Reinforcement Learning, Spatio-Temporal Attention, Continuous Low-Thrust Control, Policy Network Design

TL;DR: We propose STAN, a spatio-temporal attention network for real-time low-thrust avoidance of multistage space debris collisions using deep reinforcement learning.

Abstract: The rapid expansion of space missions has led to an exponential increase in space debris, posing severe threats to spacecraft. Existing approaches struggle to handle multistage collision risks in cluttered orbital environments, and the use of continuous low-thrust propulsion further complicates avoidance planning. To address these challenges, we propose the Spatio-Temporal Attention Network (**STAN**), which employs novel Spatio-Temporal Attention (**ST-Attention**) layers in place of conventional attention mechanisms. STAN encodes satellite-debris pairs and integrates time and distance into attention weight computation, enabling the model to generate context-aware low-thrust maneuvers. The model is trained using deep reinforcement learning across four representative multistage collision scenarios, jointly optimizing collision probability, fuel consumption, and orbital deviation. Experimental results show that STAN outperforms baseline methods in safety performance, fuel efficiency, and orbit preservation.

Primary Area: reinforcement learning

Submission Number: 8566