Instead of the classic sense-act loop that emits one action per tick, a chunking policy looks at the scene and commits to the next fifty or hundred actions in one shot, optionally blending overlapping chunks for smoothness. Borrowed from the psychology of human motor control, the idea was popularized in robotics by ACT and is now standard in vision-language-action models.
The benefit is statistical as much as mechanical: compounding error in imitation learning grows with the number of decisions the policy makes per episode, and predicting chunks divides that number by the chunk length. It also frees a slow, semantic backbone from having to run at motor rates, since one forward pass can cover a second of motion.