Athena Labs’ Robot-Training Data Card

Field notes

The proposed collection spans lab and everyday environments, with separate tiers for bimanual or humanoid capture, UMI handheld capture, and egocentric video. Athena specifies 15–60 second episodes, 30–60 Hz synchronized streams, language instructions, scene metadata, consent releases, and per-episode QA. Its unusual emphasis on failed attempts and operator corrections addresses a type of training material the page says is usually absent from open datasets.