Adaptive Control

Estimating unknown or changing parameters online while running, and automatically adjusting the controller accordingly.

Adaptive control means a controller estimates unknown or time-varying parameters online from the actual response while running, and adjusts itself accordingly; its foundation is parameter estimation. Common forms include model reference adaptive control (MRAC, which drives the system to track an ideal reference model's response) and self-tuning control, further split into direct methods (adjusting controller parameters directly) and indirect methods (estimating plant parameters first, then computing the controller). It differs from robust control: robust control fixes a range of parameter variation in advance and uses one controller to withstand the worst case, while adaptive control needs no such range and instead corrects itself as it goes. A classic in robotics is the 1987 algorithm by Slotine and Li: PD feedback plus full dynamics feedforward, with unknown parameters such as payload estimated online; it exploits the structure of arm dynamics, requiring no joint acceleration measurement and no inversion of the estimated mass matrix. Rapid motor adaptation (RMA) in reinforcement-learning locomotion does something similar with learned methods.

ExampleAn arm picks up a workpiece of unknown mass; a controller using the old model shows tracking error, while an adaptive controller updates its online estimate of the payload mass from that error, and the error shrinks over repeated motions.

Also called

Model Reference Adaptive Control, MRAC, Self-Tuning Control

Related

Robust Control · Computed Torque Control · Dynamic Parameter Identification · System Identification · Rapid Motor Adaptation · Active Disturbance Rejection Control

Sources

Wikipedia: Adaptive control
Slotine J.-J. E., Li W. On the Adaptive Control of Robot Manipulators. IJRR, 1987