Humanoids are being trained on “real-world data.” Translation: people in dozens of countries are filming themselves doing dishes with an iPhone strapped to their forehead. It’s not sci‑fi. It’s piecework with better lighting.
What’s happening
MIT Technology Review reports on companies recruiting contract workers globally to record first-person videos of household chores. The footage is then reviewed, filtered, and annotated, and sold as training data to robotics companies chasing better manipulation and generalization.
This is the embodied-AI version of the “data labeling” boom, except the labelers are also the actors, the camera rigs are consumer phones, and the set is someone’s actual home.
Why robotics wants this so badly
Robots can learn acrobatics in simulation. Manipulation is nastier. It’s contact, friction, occlusion, and endless edge cases. Simulation helps, but the mismatch between simulated physics and real objects is still brutal.
So the industry is vacuuming up demonstrations: how humans grasp, place, wipe, fold, open, and adjust. The bet is that enough messy, varied, first-person examples will help systems generalize beyond the tidy lab bench.
The part everyone keeps pretending isn’t there
This isn’t just a technical pipeline. It’s a labor system. Workers are paid to produce chore footage, and they often don’t know which robotics companies will use it, how long it will be stored, or what secondary uses it might end up in.
Even if faces are excluded, first-person home video is intimate by default. The privacy risk is not a bug. It’s the dataset.
The Droid Brief Take
“Real-world data” is a euphemism for “somebody’s apartment.”
If humanoids ever become useful in homes, the weird irony is that they’ll be useful partly because thousands of people, many of them far from the customers who will eventually buy the robots, spent their evenings reenacting domestic labor for a training set.
The industry should stop hand-waving this away as a temporary hack. It’s the core of the current approach. That means it deserves real governance: consent, deletion, clear downstream use disclosure, and an honest discussion about whether “teaching robots chores” is just outsourcing chores twice.
What to Watch
Standards for training data: what counts as “safe” demonstration data for robots that will operate near humans.
Privacy and consent policy: whether workers can actually delete footage and what “informed consent” means when clients are undisclosed.
Data versus capability: whether this approach yields durable manipulation gains, or just better-looking demos.
Sources
MIT Technology Review — "The gig workers who are training humanoid robots at home"