Closed-loop Control

Acting while watching: adjusting the next action based on the latest observation at every step.

Closed-loop originated as a control-theory concept: a controller continuously measures a system's actual output, compares it to the target, and uses the difference to correct its action — cruise control that automatically adds power going uphill, for instance. The opposite, open-loop, executes a pre-planned sequence of commands without checking the result. In robot learning, closed-loop execution means the policy keeps reading new camera images and joint states while it runs and decides its next action accordingly, so it can handle surprises like a bumped object or a slipping grasp; open-loop means computing an entire trajectory once and then following it blindly. Many diffusion policies and VLA models output a chunk of actions at a time but only execute part of it before re-observing and re-predicting — receding-horizon control — trading off between smooth motion and timely feedback. “Closed-loop evaluation” also refers to actually running a policy in an environment, rather than just comparing its outputs to a dataset's recorded actions.

ExampleA robot arm reaches for a cup, but someone nudges the cup 5 cm to the side. A closed-loop policy sees the change in the next camera frame and adjusts its path, while an open-loop policy executes the original plan and grasps empty air.

Also called

Closed-loop Execution, Feedback Control

Related

Open-loop Control · Perception-Action Loop · Closed-Loop Evaluation · Action Chunking · Visual Servoing · Model Predictive Control

Sources

Wikipedia: Closed-loop controller
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion