control//adaptive control//robust adaptation

Robust adaptation is the set of modifications that keep an adaptive controller's parameters bounded and meaningful when the plant contains dynamics, noise or disturbances the adaptive law does not model, and it exists because the textbook laws, left alone, fail in a characteristic way: **parameter drift and bursting**. With weak excitation the adapted parameters wander slowly, pushed by noise and by unmodelled effects they try to explain; they drift into a region where the loop has little margin, and then a transient sets off a burst of oscillation or instability. The loop looks perfect for minutes and then fails.

Robust adaptation is the set of modifications that keep an adaptive controller's parameters bounded and meaningful when the plant contains dynamics, noise or disturbances the adaptive law does not model, and it exists because the textbook laws, left alone, fail in a characteristic way: parameter drift and bursting. With weak excitation the adapted parameters wander slowly, pushed by noise and by unmodelled effects they try to explain; they drift into a region where the loop has little margin, and then a transient sets off a burst of oscillation or instability. The loop looks perfect for minutes and then fails.

The warning came early. In 1985 Rohrs, Valavani, Athans and Stein showed that classic adaptive algorithms, proved stable for their model, could be destabilized by small unmodelled dynamics the design ignored, such as a fast actuator lag, when driven by certain reference or disturbance signals. That result is known as the Rohrs counterexample and it founded the field of robust adaptive control. The same lesson applies to the recursive estimators behind indirect schemes: with forgetting and no excitation their covariance grows until the first manoeuvre makes the parameters leap (recursive least squares).

The standard defences each keep the adapter from believing what the data cannot support:

Parameter projection keeps the estimates inside a physical range (the mass between 0.8 and 2 kg), so whatever happens the controller never acts on an absurd model.

Sigma modification adds a small spring term that pulls the parameters back toward their nominal values, trading a little bias for immunity to drift.

An adaptation dead zone stops adapting while the error is at noise level, since a noise-sized error carries no information about the parameters.

Adaptation freeze stops updating when the data carry no information (steady hover, a plant at constant setpoint), and natural manoeuvres are used as excitation when they come (persistent excitation).

Each defence costs something: projection needs prior knowledge of the range, sigma modification biases the estimate, dead zones and freezes slow the learning. Together with a robust baseline that flies alone if adaptation is switched off, they are what separates an adaptive controller that can be deployed from one that only works in the paper (MRAC, adaptive control).