Robotics never inherited the internet’s gift of free training data, so somebody has to build the dataset from scratch. Mecka AI founder Josh Gao runs one of the largest robot data operations and walks through what that takes: custom baseball caps and tactile gloves because off-the-shelf hardware cannot capture force, 100-gigabit upload lines moving hundreds of petabytes, and quality gates designed to catch bad collection runs.
He lays out the recipe the field is converging on, roughly a million hours of egocentric human footage, tens of thousands of hours of robot data and a few hours of in-domain teleoperation, and argues that tactile and simulated data still have to enter the same pot. His read on timing is sober: no single GPT-3 moment for robotics, just a steady unlock of commercial tasks starting with constrained industrial work.
