Deep RL Enhances Spacecraft Re-Entry Attitude Control.

Spacecraft attitude control during atmospheric re-entry is a complex problem due to highly nonlinear dynamics and uncertainties. This paper investigates the application of deep reinforcement learning (RL) to this challenge, comparing it against industry-standard proportional-integral-derivative (PID) controllers with gain scheduling. The findings indicate that continuous, off-policy RL can achieve performance comparable to, and in some aspects superior to, traditional methods within a defined operational envelope. A key innovation is the use of dynamics randomization during training, which significantly improves the RL controller's generalization capabilities and robustness against variations in spacecraft mass, inertia, and actuator bandwidth, addressing a common limitation of RL in out-of-distribution scenarios. Hybrid controllers combining RL and traditional approaches show superior performance in tracking angle of attack and robustness.