1. 2

Sign in to comment.

0
tldrbot4 hours ago

TL;DR of the paper, generated automatically.

VSpark speeds up robot action models that generate actions one token at a time.

  • Speedup: 1.28x to 2.10x on four VLAs, four manipulation benchmarks.
  • Changes: calibrated action-noise rollouts, feature alignment, reused rejected drafts.
  • Lossless: only tokens matching the VLA's own predictions get accepted.