Supervised Residual NMPC in Force-Disturbed Quadrotors
Residual deployment can affect closed-loop tracking. We compare two ways to deploy one supervised acceleration-residual network in nonlinear model predictive control (NMPC) of a quadrotor. Cyclewise deployment repeats one correction across the horizon. Stagewise deployment evaluates shifted prediction nodes before the solve. Network parameters remain fixed during control. On prescribed force pulses, both schemes improve average tracking mainly after force removal, but worsen hover; some prediction checks also fail. Seed 1 reinforcement-learning benchmarks add no mean benefit under pulses or prescribed random forces. On 220 conditions, cyclewise and stagewise supervision reduce mean integral absolute position error (IAE) by 4.38% and 7.77%. On a separate 32-condition bank, fixed gain 1.5 lowers IAE by 27.47% relative to stagewise supervision but raises angular-rate root-mean-square error by 10.19%. In one synthetic computational fluid dynamics field, stagewise supervision lowers four-task mean IAE by 2.01% against a same-graph zero-residual control; some secondary metrics worsen. These simulations show that residual deployment affects tracking, with gains and limitations that depend on task and motion metric.