SiMDex

“SiMDex retrieves only ~1.49M samples (under 5% of a 32M-sample egocentric video pool) and improves overall dexterous manipulation success from 47.7% to 61.1% — a +13.4 point gain”

“The paper's central finding is that *which* human video data you feed a VLA matters more than *how much*. This directly challenges the prevailing 'just scale data' philosophy in robotics.”