Inside the Robot Data Factory: Mecka's Josh Gao on Sensorizing Humans to Train Robots

Robotics data collection is still orders of magnitude behind Tesla's decade of driving data, so progress will unlock commercial tasks gradually, not in one breakthrough.

• Data constrained field: Unlike LLMs' free internet corpus, robotics has no equivalent, forcing Mecka to manufacture pixels, actions, and logic via custom hardware like tactile gloves and sensorized baseball caps.
• Converging training recipe: Industry is settling on roughly a million hours of egocentric human data, tens of thousands of hours of UMI/teleop data, and a few hours of in-domain teleoperation per task.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.