Physical AI's Next Bottleneck Is Finding the Right Video — Rafael Levi, Bright Data

Rafael Levi of Bright Data addresses the data bottleneck in physical AI training, contrasting the vastness of text data with the scarcity of robot-specific video data. He criticizes staged recordings for bias and simulation/teleoperation for scalability issues. The core contribution is Bright Data's 'search first, collect second' approach: indexing over a billion videos based on actions within them (not titles) and providing an API to retrieve trimmed clips with timestamps, match scores, and frame counts. This method allows for efficient retrieval of real-world physics and interactions, exemplified by searches for dishwashing and clothes-folding clips, and use cases in self-driving and brand discovery. The talk emphasizes that NVIDIA discards 96% of downloaded video for training, while Meta used 1 million hours of public video and only 62 hours of robot data, underscoring the value of intelligent video indexing.