adaptive control

Adaptive control is a robot's ability to retune its own controller on the fly, as it runs, when it discovers that the world is not quite what it expected. Most controllers are set up once with fixed numbers based on an assumed model of the robot — how heavy its arm is, how much friction sits in its joints. But the real robot might pick up a heavier tool, run a motor warm, or wear down over months. Instead of failing or needing a human to re-tune it, an adaptive controller watches how it is actually doing and quietly adjusts its own settings until performance comes back.

The way it works is a constant feedback loop on top of the normal control loop. The controller compares what it expected the robot to do with what the robot actually did. If the arm keeps drooping lower than commanded, the controller infers 'I must be underestimating the load' and increases the right internal parameter; the droop shrinks; it nudges again; and so the estimate creeps toward the truth. Over a little time the controller's idea of the robot lines up with the real robot, and the errors fade away on their own.

This is exactly what you do learning to lift an unfamiliar box: your first lift overshoots or undershoots because you guessed the weight wrong, but within a heave or two your muscles recalibrate. Adaptive control gives robots that same self-correcting knack, which is why it shows up wherever the load or the dynamics keep changing — arms that grasp objects of unknown weight, aircraft that burn off fuel and grow lighter mid-flight, or aging machines that must keep working as their parts drift.