Action Chunking

Instead of deciding one action at a time, predict the next half-second of motion at once and execute it. This cuts the number of decisions in a task, and with it the compounding error that makes long horizons fail — the same exponent Long-Horizon Agent describes, attacked by shortening the exponent rather than raising the base.

It also smooths the motion, because a chunk is internally consistent in a way a sequence of independent decisions is not, and it hides inference latency: the robot executes chunk k while the model computes chunk k+1. The cost is reactivity — a chunk committed to is a plan that ignores what happens during it — so the chunk length is a direct trade between smoothness and responsiveness.

With chunk length k, a horizon of T steps needs T/k decisions, so compounding error falls from p super ^T  toward p super ^T/k . The overlap between successive chunks decides how much the policy can revise: executing a chunk fully is open-loop within it, while re-planning every step with temporal ensembling recovers reactivity at more inference.