Work

From Classical Control to Deep Reinforcement Learning: A Reproducible Framework for Quadrotor Modeling and Control

Year: 2026
Type: article
Abstract: Unmanned aerial vehicles used in delivery applications must maintain reliable trajectory tracking despite payload variations that alter the vehicle dynamics. This study compares seven control strategies—proportional–integral–derivative (PID), linear quadratic regulator (LQR), linear quadratic integral (LQI), linear model predictive control (MPC), nonlinear MPC (NMPC), Adaptive NMPC, and Soft Actor–Critic (SAC)—for a nonlinear quadrotor under nominal conditions, staged payload-mass variation, and measurement noise. All controllers are evaluated using the same vehicle model, actuator constraints, and four benchmark trajectories. The comparison considers tracking accuracy, robustness, estimated mechanical propulsion power, actuator saturation, benchmark completion, noise sensitivity, and computational burden. Under nominal conditions, NMPC and Adaptive NMPC achieve nearly identical tracking, with a maximum difference in three-dimensional root-mean-square error (3D RMSE) of approximately 4.5 × 10−4 m. Under staged payload-mass variation, Adaptive NMPC completes all four trajectories without divergence, whereas fixed-model NMPC remains bounded only in Hover and diverges on the three moving trajectories. In the Hover case, adaptation reduces the 3D RMSE by 80.2% (from 0.3085 m to 0.0611 m). SAC exhibits pronounced trajectory- and seed-dependent performance; the representative policy (Seed 1) completes only 3 of the 8 nominal and staged-mass benchmarks at full duration. These results demonstrate that strong nominal tracking alone does not guarantee robustness to payload variation, and they highlight the benefit of online model adaptation for quadrotor control under changing mass conditions.
Source: Machines
Language: English
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7234180868
FWCI:
Cites: 61
Cited by:
Related to: 10