Flexible Trajectory Planning With the Dyna Reinforcement Learning Architecture Under Computation Constraints

Spacecraft autonomy is increasingly critical for missions involving long communication delays, spacecraft fleets, or fault-prone actuators. Prior work demonstrated that Dyna-style reinforcement learning (RL) architectures, which combine model-based planning with model-free updates, can adapt to actuator degradation and changing dynamics. However, these implementations assumed fixed computational availability, limiting robustness under dynamic processor loads. This work extends the prior Dyna framework by incorporating time- and processor-aware planning, dynamic batching of critic updates, and improved activation functions to mitigate neuron inactivation. Simulations demonstrate that the modified architecture reduces control cost, lowers failure probability under thruster degradation, and decouples performance from processor speed. These results highlight the importance of integrating computational awareness into RL-based controllers and provide a practical methodology for improving the resilience and adaptability of spacecraft guidance, navigation, and control systems.