An Ego4D alternative for commercial robot training

Ego4D is a large-scale first-person video database. Access goes through a data usage agreement that each user requests on the Ego4D license page. The published draft of that agreement says:

• Use is limited to a listed Purpose. It covers research publications, and using the database to research, develop, train, evaluate or improve technology that detects, classifies, recognises, retrieves or understands objects, events, places or activities, models 3D objects or environments, or improves VR and AR applications.
• For those listed uses, the draft allows academic research and commercial or noncommercial product development.
• You keep the rights to the models and software you develop from the database.
• The data itself may not be sold, sublicensed or provided to third parties, and may not appear in another dataset or product except as the agreement permits.

So Ego4D is often called research-only, but the text is more specific than that. If you plan to use it for a commercial robot, read the agreement you sign and check with your own counsel that your use fits the listed Purpose.

Public data gives breadth. A custom set adds the specific tasks your robot must do, in the places it will work:

• Hotel housekeeping: bed making, linen change, bathroom cleaning, amenity restocking.
• Kitchens: prep, chopping and plating, dishwashing and racking, station clean-down.
• Laundry, reception and short-let apartment turnovers.

Every worker signs informed consent, every venue agrees to the recording, and faces and sensitive details are de-identified before footage is stored. See worker consent and de-identification.

A pilot is typically 20 to 50 hours of accepted footage, so you can test the data in your own pipeline before you scale. Billing is tied to the hours you accept.