Stable Adaptive Flight Control for Multi-Platform UAV Monitoring: Combining Reinforcement Learning with Lyapunov-Guaranteed Gain Tuning

Unmanned Aerial Vehicles (UAVs) have become vital tools for environmental monitoring and disaster response, particularly in remote regions where ground-based infrastructure is sparse (Erdelj et al., 2017). Yet the practical value of these platforms depends heavily on precise flight control when navigating challenging terrain or conducting time-sensitive surveys after natural hazards. Tuning Model Predictive Control (MPC) weight matrices for such varied operational demands remains tedious and expertise-intensive, which slows deployment during crises.

We present an adaptive control framework merging reinforcement learning with formal stability guarantees. A learning agent tunes controller gains online while Lyapunov-based bounds confine every candidate gain to a provably stable region (Christofides et al., 2011). A projection operator acts as a hard safety layer, clipping any out-of-bounds gain before it reaches the MPC solver. The resulting architecture preserves guaranteed stability regardless of policy network behaviour—essential when aircraft operate beyond visual line of sight in poorly monitored areas.

Validation spans four UAV platforms covering a 200-fold mass range (27 g to 5.5 kg). On an aggressive 3D figure-8 trajectory (±4.0 m on both horizontal axes), tracking improves by 22–27 %. Position root mean square error falls from 0.45–0.55 m to 0.33–0.43 m, with variance reductions of 28–33 %. Across 60 evaluation trials, no stability violations occurred. Sequential transfer learning cuts per-platform training by 75 %, valuable when field crews must swap vehicles mid-campaign—switching, for instance, from a compact quadrotor on initial reconnaissance to a heavier hexacopter carrying hyperspectral sensors.

These results show that rigorous stability guarantees and learning-based adaptation can coexist. For observation campaigns in areas lacking ground-based networks, self-tuning controllers that never risk unstable flight could meaningfully extend what small drone fleets achieve—whether assessing earthquake damage or inspecting infrastructure in informal settlements.

 

References

Christofides, P.D., Liu, J., Muñoz de la Peña, D. (2011). Lyapunov-Based Model Predictive Control. In: Networked and Distributed Predictive Control. Advances in Industrial Control. Springer, London.

Erdelj, M., Natalizio, E., Chowdhury, K.R., Akyildiz, I.F. (2017). Help from the sky: Leveraging UAVs for disaster management. IEEE Pervasive Computing, 16(1), 24–32.