Trajectory Optimization for Reusable Launch Vehicles Based on a Reinforcement Learning Heuristic Hybrid Algorithm

Rapid trajectory optimization for reusable launch vehicle (RLV) has become a focus in aerospace research; however, its inherent nonlinearities and multiple constraints make it highly challenging. To address this issue, our proposal begins with a glide profile by considering strict terminal and steady-glide constraints, from which analytic expressions were derived to reduce nonlinear coupling and facilitate optimization. Nofly zones were modeled as polygons, and a cubic spline-based lateral avoidance strategy was implemented. Then, a hybrid optimization framework integrating analytic solutions, particle swarm optimization (PSO), and deep reinforcement learning (DRL) was developed. PSO’s exploration capability was retained, while DRL dynamically adjusted swarm hyperparameters based on swarm behavior to improve efficiency and solution quality by balancing exploration and exploitation. In the end, numerical simulations under multiple scenarios validated full constraint satisfaction, with a mean terminal state error below 0.92%. The DRL-PSO framework demonstrated adaptability and efficiency in rapid trajectory computation, achieving an average runtime of approximately 1.0 s. Furthermore, its effectiveness was validated through comparative analysis with alternative methods.