Autonomous Spacecraft Threat Avoidance via Trajectory Prediction and Deep Reinforcement Learning

In the increasingly congested orbital environment, spacecrafts are facing growing threats from space debris and adversarial targets, where the challenges of autonomous threat avoidance include dynamic uncertainties and multi-objective optimization of fuel efficiency and evasion effectiveness. To address this issue, this paper proposes an autonomous threat avoidance decision-making strategy for spacecraft. First, a Conditional Variational Autoencoder (CVAE)-based predictor is developed to estimate future trajectories of space targets on the basis of collected trajectories from historical missions. Furthermore, a Deep Deterministic Policy Gradient (DDPG)-based decision framework is developed, incorporating a multi-objective reward function, which enables the maneuver decision-making under the constraints of fuel consumption, relative distance variation and maneuver direction rationality. Experiments validate that the proposed strategy performs better than the prediction-free baseline that makes decisions based only on current states, in terms of the response time, which demonstrates its effectiveness in autonomous threat avoidance within space environments.