To address real-time safety planning for tube-launched UAVs under highly dynamic launch scenarios, this paper proposes a trajectory planning method integrating offline NMPC and online inference through behavioral cloning (BC). First, NMPC generates expert trajectories offline under dynamic initial conditions while satisfying multi-stage constraints. BC then trains networks to learn state-to-trajectory mappings, implicitly inheriting safety constraints while enabling real-time inference. Simulations demonstrate feasible trajectories balancing computational efficiency and constraint compliance under uncertainties (e.g., attitude deviations, aerodynamic disturbances), significantly enhancing launch safety.