Open-loop Control

Executing a pre-computed sequence of actions straight through, without checking feedback along the way.

Open-loop is originally a control-theory term: the control action doesn't depend on the system's output, and results are never checked during execution — a clothes dryer that just runs for a fixed time is one example. Closed-loop, by contrast, keeps measuring the result and correcting for it. In robot learning, “open-loop” usually means not reading any new observation for the duration of an action segment. Take action chunking: a model predicts several future steps of action at once, and the π0 paper states explicitly that the whole chunk runs open-loop — on a 50 Hz robot, it runs inference once every 0.5 seconds and doesn't look at new images for the 25 steps in between. Open-loop execution gives smooth, coherent motion and needs fewer inference calls, but reacts slowly to sudden changes — if an object gets bumped, the robot has to wait for the next inference call to correct course. Common compromises are executing only part of the predicted chunk before replanning, as in Diffusion Policy's receding-horizon approach, or using real-time chunking to compute the next segment while the current one is still playing out.

ExampleOpen-loop playback: a robot arm follows a pre-recorded trajectory to grab a cup, and still closes its gripper even if someone has moved the cup a few centimeters away; a closed-loop policy would adjust its hand position based on the new image.

Also called

Open-loop Execution

Related

Closed-loop Control · Action Chunking · Action Horizon · Real-Time Chunking · Temporal Ensembling · Open-loop Evaluation

Sources

Open-loop controller (Wikipedia)
π0: A Vision-Language-Action Flow Model for General Robot Control (arXiv:2410.24164)
Real-Time Execution of Action Chunking Flow Policies (arXiv:2506.07339)

As of

2025-06