Robotics data bottleneck: NVIDIA discards 96% of training video — search-first indexing as the fix
AI Engineer · youtube · 2026-09-24
Bright Data's Rafael Levi argues physical AI's next bottleneck is finding the right video: LLMs train on trillions of words, but robotics has only about a million videos of robots acting.
- Staged recordings produce biased data (nobody opens a door naturally on command); simulation and teleoperation don't scale
- The web holds billions of hours of real people handling objects with real physics
- Noise is costly: NVIDIA discards 96% of video downloaded for Cosmos, Stable Video Diffusion discards 74%
- Counterpoint: Meta trained on 1M hours of public video plus only 62 hours of real robot data to control a robot
Bright Data's "search first, collect second" approach indexes 1B+ videos by the actions in them, returning trimmed clips with timestamps, match scores and frame counts via API — applicable to self-driving and brand discovery too.
More from Embodied
- Trainable JWST model boosts exoplanet sensitivity 2–15x in Nature paper — bravo_abad · 2026-11-11
- Clean point clouds: Tersus GNSS MVP S1 scans turned into high-quality splats in LichtFeld Studio — janusch_patas · 2026-11-11
- Flashing cheap ESP32 IoT devices with ESPHome for a smart-home upgrade — haydendevs · 2026-11-11
- NYU's Steertuning Lets a Robot Learn a Chore From 5 Minutes of Human Demo — CyberRobooo · 2026-11-11
- Boston Dynamics Atlas demoed with Blender 5.3 — TomLikesRobots · 2026-11-11
- Lamport Robotics posts robot 'wake-up test' demo video — R4nsack · 2026-11-11