Why Labs Are Buying 1M Hours of Human Video for Robots | Varun Nair

In this conversation, we sit down with Varun Nair to dig into the high-stakes world of data for robotics and physical AI. We talk about what actually matters for training robot intelligence. From egocentric video to UMI over teleop, simulation and the gaps between them. You will also learn why the field is still figuring out which data scales and the limits of human-operated data, evaluation challenges, and what it will take to move from demos to real deployment. If you’re building, investing in, or just obsessed with robotics and embodied AI, this one’s for you. #robot #data #ai 0:00 Podcast Overview 0:30 Data Market 1:30 Varun's Background 3:01 Robotics & Crypto 6:27 Physical AI Definition 8:32 Simulation for Deployments 10:45 World Models 11:45 What is good data ? 12:42 GPT 6 Astra Evaluation 13:50 Data Life Cycle 17:01 Modalities & Pseudo Labeling 19:40 Depth for LLMs 20:19 How labs use data 21:38 Deployment Data = Post Training Data 22:17 Teleoperation 24:00 Patient Capital vs Impatient Capital 24:47 Benchmarks 26:35 UMI Data for Mid Training 27:35 Has Openai been training on robot data? 30:59 Humanoid vs other Embodiments 31:36 Hard vs Easy Tasks for Robots 32:51 Simulation 33:59 Niantic Spatial 34:42 Is robotics data broken? 36:17 Model Companies 37:38 Robot Everest 38:11 Avea 38:47 What Varun looks for 40:17 Outro Guest https://x.com/_varunnair Léo https://x.com/LeoKharon Oskar https://x.com/hundreddaysx Podcast https://x.com/roboticsstack Spotify: https://open.spotify.com/show/033M6Ih... Apple Podcasts: https://podcasts.apple.com/us/podcast...