Great read, valuable information about the real bottlenecks with egocentric data collection. I agree the long‑term goal should be to get as close as possible to well labeled, outcome‑linked egocentric data.

I’m not an expert in this space, but do we mostly write off unannotated egocentric footage that exists today? My intuition is that with newer training approaches (self-supervised learning, better VLMs, etc.), messy egocentric footage can be mined for patterns over time, especially in repetitive settings like kitchens and factories.