Adaptive control

Change the controller while the robot runs#

An adaptive controller contains an update rule as well as a control law. Measurements such as tracking error drive the update rule, which changes parameter estimates or controller coefficients while the robot is operating. The control law then uses those changing values to calculate the next command.

This is useful when a fixed model is not accurate enough for every operating condition. A manipulator may pick up an unknown payload, for example, so the inertial and gravity terms used by an inverse-dynamics controller no longer match the plant. The adaptive-control study by Wang addresses nonlinear robot manipulators with parametric uncertainty and constructs adaptive controllers with analysed closed-loop dynamics.

Adaptation is part of the feedback design#

Adaptive control differs from offline system identification. Identification can estimate a model before a task, while an adaptive controller updates quantities during closed-loop operation. The two can be combined, but an online parameter estimate need not converge to the robot's true physical parameter values merely because tracking improves.

It also differs from gain scheduling. A scheduled controller selects gains from a map built around a measured condition, such as speed or payload class. An adaptive controller uses observed behaviour to update its internal quantities. Rapid motor adaptation is a learned approach to fast online adjustment, whereas classical adaptive control usually starts from an explicit plant structure and a stability analysis.

The assumptions still matter#

An update law cannot recover information that the motion does not reveal. Parameter convergence may require sufficiently informative excitation, and several physical parameters can produce similar measured behaviour. Sensor noise, unmodelled friction, actuator saturation, delays, and contact changes can therefore bias or destabilise an update.

Stability guarantees also apply to the model class and constraints used in their proof. They do not automatically cover impacts, hard joint limits, perception failures, or every payload. A practical system needs bounded updates, safe command limits, and tests that separate tracking improvement from reliable physical-parameter estimation.