control//robust control

Robust control is the design and analysis of feedback controllers that must remain stable, and keep performing, for every plant in a described family rather than only for the nominal model, and it is how an engineer pays in advance, with some performance, for the certainty that a heavier payload or a slower motor will not destabilize the loop. Every model is wrong; the question is how much and where. Uncertainty arrives in three shapes: **parametric** (mass varies 20 %, thrust per command changes with the battery), **unmodelled dynamics** (motor lag, arm flex, filters), which usually grows with frequency, and **disturbances** such as wind.

Robust control is the design and analysis of feedback controllers that must remain stable, and keep performing, for every plant in a described family rather than only for the nominal model, and it is how an engineer pays in advance, with some performance, for the certainty that a heavier payload or a slower motor will not destabilize the loop. Every model is wrong; the question is how much and where. Uncertainty arrives in three shapes: parametric (mass varies 20 %, thrust per command changes with the battery), unmodelled dynamics (motor lag, arm flex, filters), which usually grows with frequency, and disturbances such as wind.

The tools run from the most used to the most sophisticated. Classical stability margins and the peak of the sensitivity function MsM_sMs​ (a usual target between 1.2 and 2; Ms=2M_s=2Ms​=2 guarantees a gain margin of 2 and a phase margin of 29° at once), checked at every point where the loop can break: each actuator, each sensor. Monte Carlo testing: simulate hundreds of variants (masses, delays, motor constants) and look at the worst (Monte Carlo method); it is what is done most in practice. At the top, H-infinity control synthesizes the controller that minimizes the worst-case amplification of uncertainty.

A robust guarantee covers the uncertainty you described, not the one you forgot.

If the ESC delay was not in the uncertainty model, the certificate says nothing about it. Optimal is not robust either: an optimal controller is best for the model you wrote and the noise you assumed, as minimizing training loss says nothing about data outside the training set (LQG is the classic case).

A robust controller is one fixed design that holds for the whole family of plants; its alternative for a plant that changes a lot is one that finds out online which plant it got (adaptive control).

Bode's sensitivity integral is the no-free-lunch theorem behind every robust design: pushing sensitivity down in one band raises it in another, and an unstable plant pays an extra tax. Robustness is a choice of where error lives across frequency (error relocation).

Domain randomization is robust control by sampling: the randomized simulator ranges are the uncertainty set, and the guarantee is empirical, valid on the samples seen (domain randomization).

Passive fault tolerance is robust control applied to faults: a controller that withstands a weakened motor without knowing it happened. Maturity: margins and Monte Carlo are industry everywhere; H-infinity and mu synthesis are respected niche, mostly aerospace.