Raw video is not a robotics dataset
Most robotics teams still treat data like a side quest. Record some demos. Label a few hours. Hope the model generalizes.
It does not.
Physical AI fails in the gap between the lab and the job: noisy rooms, odd grips, lighting that never showed up in simulation, workflows nobody teleoped 10,000 times.
What actually moves a system into the field:
Task-specific, first-person human activity — not third-person lab takes
Annotation, QA, and motion understanding before training, not after a failed eval
Evaluation against real environments, not a clean test set
That is the layer Mecka owns. Hardware, models, and commercial partners already exist. The missing piece is the data and integration glue between them.
If you are a lab, a frontier team, or an enterprise trying to get a robot to do real work, start with the task — not the camera.
