Action-chunk subsampling makes SmolVLA robots run 8x faster

Subsampling action chunks can dramatically boost a robot policy’s execution speed without changing the core model, making learned control far more practical on real hardware.

when you have a working action-chunk-based robot policy, try to subsample the action chunk to make the robot go brr brr brr here is me running SmolVLA at 8x speed, short explanation at the end