The growing adoption of small satellites introduces significant challenges in terms of system reliability and on-board computational resources. The Attitude Determination and Control System (ADCS) is prone to actuator failures; specifically, the loss of reaction wheels (RWs) may render the spacecraft underactuated, requiring advanced control strategies to recover three-axis stability. While optimal control techniques such as Nonlinear Model Predictive Control (NMPC) can handle coupled nonlinear dynamics and actuator constraints in failure scenarios (e.g., 2 active RWs out of 4, assisted by 3 magnetorquers for attitude support and momentum dumping), their computational burden may hinder real-time implementation on embedded hardware. A data-driven approach based on imitation learning is investigated to bridge the gap between control optimality and computational feasibility.