Hybrid Deep Reinforcement Learning for Real-Time Multi-Satellite Mission Planning under Dynamic Constraints

In multi-satellite mission planning, limitations are exhibited in addressing scenarios with dynamically changing mission constraints and further improvement are still required for handling the integration of local and global information. A method trained with deep reinforcement learning is proposed to solve the dynamic mission planning problem. The proposed method effectively adapts to dynamic changes by continuously updating the mission's information through the integration of static and dynamic satellite information, supported by a global critic network that evaluates overall performance and allows individual satellites to make independent decisions based on local information. Case studies validate that the proposed method achieves enhancements compared to baseline methods in terms of mission completion efficiency, adaptability to dynamic priorities, and computational speed, making it a robust solution to large-scale multi-satellite mission planning scenarios.