Event

Tuning Protocol Changes Quadrotor Controller Rankings

1 reports · 1 sources

Xuan-Quy Dao et al. compared cascaded PID, hovering LQR, BL-SMC, linear MPC with LQR fallback (LMPC-F), LMPC-F with preview enabled, and a geometric SO(3) controller on a nonlinear plant that includes rotor gyroscopic coupling, motor lag, and saturation. The study found that after adopting a unified twelve-candidate tuning protocol, the average RMSE ranking under a fixed baseline changed, with BL-SMC rising from last place to first; this ranking was reproduced on 50 new implementations and remained unchanged under a 48-candidate budget. The main study included 556 50 Hz episodes, 400 paired evaluation episodes, and a 48-cell MPC factor experiment; the confirmation study used 1462 episodes to repeat evaluations on 50 new implementations, with tuning budgets of 12, 24, and 48 candidates. In timing tests, replaying measured command-call latency caused three of the four nominal MPC configurations to terminate on a desktop computer, and all four to terminate on a Raspberry Pi 5; after setting the solver iteration cap to 100, no deadline misses occurred among the tested nominal timing configurations.