Onboard Deep Reinforcement Learning: Deployment and Testing for CubeSat Attitude Control

Recent progress in Reinforcement Learning (RL), especially deep RL, has created new possibilities for autonomous control in complex and uncertain environments. This study explores these possibilities through a practical approach, implementing an RL agent on a custom-built CubeSat. The CubeSat, equipped with a reaction wheel for active attitude control, serves as a physical testbed for validating RL-based strategies. To mimic space-like conditions, the CubeSat was placed on a custom air-bearing platform that allows near-frictionless rotation along a single axis, simulating microgravity. Unlike simulation-only research, this work showcases real-time hardware-level implementation of a Double Deep Q-Network (DDQN) controller. The DDQN agent receives real system state data and outputs control commands to orient the CubeSat via its reaction wheel. For comparison, a traditional PID controller was also tested under identical conditions. Both controllers were evaluated based on response time, accuracy, and resilience to disturbances. The DDQN outperformed the PID, showing better adaptability and control. This research demonstrates the successful integration of RL into real aerospace hardware, bridging the gap between theoretical algorithms and practical space applications through a hands-on CubeSat platform.