Labeling multi-task teleop data or egocentric data is a pain 😟 As we million-x the data collected, we cannot hand label every scene ... BUT, today's tools are really powerful ... we passed 104 seconds of raw robot video into a simple pipeline, and got out: - 12 task segments