The artificial intelligence startup has released a tool that transforms recordings of humans performing everyday work into structured datasets used to teach robots how to move. The system, called Memories Upload, generates timed labels and 3D hand pose data from first-person footage to help physical AI better handle real-world tasks and small corrections. The company says 3 of the top 5 frontier robotics labs have already begun using the platform to train their models.
The network utilizes post-trained Nvidia Cosmos for video annotation and relies on recordings from the LUCI Pin, a 4K wearable camera for which users are paid to record tasks. The current dataset covers more than 1,000 different tasks across 20 industries on five continents, providing a search and retrieval pipeline for labs to select specific clips via API or MCP server.
