Stability-Guaranteed Reinforcement Learning for Model Predictive Control Gain Tuning with Transfer Learning in Multi-UAV Systems

Tuning Modle Predictive Control (MPC) weight matrices remains a tedious process that demands considerable expertise. A method is presented that enables a reinforcement learning agent to adjust gains online while enforcing Lyapunov-based bounds guaranteeing that every proposed gain lies inside a provably stable region. The resulting adaptive controller maintains stability regardless of policy network outputs. Across four Unamnnaed Aerial Vehicle (UAV) platforms spanning a 200-fold mass range (27 g to 5.5 kg), tracking improvements of 22–27% are observed on an aggressive 3D fgure-8 trajectory spanning ±4.0m on x and y axes. Position root mean square error (RMSE) is reduced from 0.45–0.55m to 0.33–0.43m across all platforms, with variance reductions of 28–33%. No stability violations occurred throughout 60 evaluation trials. A projection operator clips out-of-bounds gains before they reach the MPC solver, acting as a hard safety layer. Sequential transfer learning reduces per-platform training by 75%. These findings demonstrate that formal stability constraints and learning-based adaptation can coexist effectively.