EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data

Technical Analysis "EgoScale" is a human-to-dexterous-manipulation transfer framework built on large-scale egocentric human data. High-level Analogy: Imagine you want to teach a junior chef (the robot) how to perform complex kitchen tasks, like finely dicing vegetables, using specialized tools, and elegantly plating dishes. Traditional way (prior work): You’d teach them one specific task at a time, like 'chop this specific onion with this specific knife.' They learn slowly, and it's hard for them to adapt to new onions or different knives, or to combine skills for a whole meal. They only get to watch a few hours of an expert chef, which isn't enough to build true mastery. EgoScale's way: Massive 'Watch and Learn' (Human Pre-training): First, you let the junior chef watch thousands of hours of master chefs (humans) cooking everything imaginable—from intricate pastry work to quick stir-fries, all filmed from the chef's own perspective. They see a vast array of hand movements, tool uses, and problem-solving, even if some parts are a bit blurry or happen in a messy home kitchen. This teaches them the fundamental 'grammar' and 'vocabulary' of cooking movements, like how to hold a knife, stir, or peel. Personalized 'Kitchen Alignment' (Aligned Human-Robot Mid-training): Next, after all that extensive learning, you give the junior chef a small, focused set of videos. In these videos, a master chef demonstrates a few dishes, but crucially, they do it in the junior chef's exact…