Sequoia Podcast: Inside the Robot Data Factory Making Millions of Hours of Human Training Data
Training Data (Sequoia) · rss · 2026-10-20
Sequoia's Training Data podcast hosts Mecka AI founder Josh Gao on manufacturing robot training data at scale:
Inside the data factory
- Robotics never got free internet data, so it must be manufactured: custom baseball caps and tactile gloves (nothing off-the-shelf captures force data), 100-gigabit upload lines moving hundreds of petabytes, custom quality gates and fraud controls.
- First pilot: recruiting "aunties and uncles" via a mobile app to collect egocentric human data.
The industry's converging training recipe
- Roughly 1 million hours of egocentric human data + tens of thousands of hours of robot data + a few hours of in-domain teleop, with tactile and simulation data still needing to enter the mix.
- Josh argues robotics remains orders of magnitude behind Tesla FSD, which spent a decade on just two degrees of freedom while a robot hand has dozens.
Key takes
- No single "GPT-3 moment" for robotics — instead a steady unlock of commercial tasks, starting with constrained industrial work.
- Home data is plentiful but home robots are the hardest; useful robots can deploy today. Long-term: robots mining asteroids.
More from Embodied
- Flipper One pre-orders open as Kickstarter campaign readies to launch — gnukeith · 2026-10-20
- ARK's Tesla 2031 Model Sees $4,400 Base Case, Bull Case $5,000 on Robotaxi Bets — lasas · 2026-10-20
- Tsinghua and AgiBot release RoboValue, a sim-and-real benchmark for robotic value models — Shengbang Liu · 2026-10-20
- AI-assisted robotic glove helps people regain hand control and independence — CurieuxExplorer · 2026-10-20
- Harness Pro: a $299 desk screen to monitor AI agents across all your machines — dee_hw · 2026-10-20
- A fruit fly's 166k-neuron connectome is now driving real robots, no AI training needed — CurieuxExplorer · 2026-10-20