New gig economy earning $15 per hour: Workers from over 50 countries participate in robot action video collection

Caixin April 2 News (Editor: Ma Lan)
Humanoid robot training requires a large amount of real action data.
Previously, the industry mainly collected and integrated these materials by establishing dedicated data factories, but new methods have now emerged.

According to MIT Technology Review,
Micro1 is recruiting gig workers worldwide to record videos of themselves performing household chores, such as folding towels, opening refrigerators, stacking dishes, mopping, and so on.

The company has recruited contractors in countries like Kenya, the Philippines, India, and Brazil,
requiring workers to wear cameras and record hours of household labor, earning about $15 per hour.
This rate is quite competitive in emerging markets, but compared to the billions of dollars invested by companies in this industry, it’s just a drop in the bucket.

However, the gap between huge investments and the relatively modest pay for end workers
is likely a quick solution to obtaining large amounts of real data, and it’s also a key to the future large-scale production of humanoid robots.

Risks

In chaotic and unpredictable real home environments, how humans grasp, twist, lift, and place objects
is an important reference for robot training.
Robot companies have also spent heavily on purchasing such data in the past, such as US startup Scale AI, which collected over 100k hours of training videos.

Meanwhile, some traditional industry companies have developed new ways to make money.
For example, food delivery company DoorDash allows delivery drivers to兼职 contribute training data, which is then sold to humanoid robot manufacturers.

This business model, to some extent, is not very friendly to grassroots workers,
since video providers receive relatively little compensation.
On the other hand, these on-site videos also pose privacy risks.
Robot companies purchasing these video materials rarely disclose their data retention policies, such as whether videos are anonymized, and what happens to these videos after training ends.

Some experts have raised questions, such as a worker in Brazil filming himself installing a washing machine,
whose video not only contains information about the washing machine installation but also reveals visual content of Brazilian family life.
The ownership and access rights to these videos are not clearly defined, and data leaks are a significant concern.

Interestingly, earlier this month, it was revealed that the game Pokémon Go provided data to robot manufacturers.
They partnered with Coco Robotics, a short-distance food and grocery delivery robot,
using over 30 billion photos taken by players to train visual positioning systems, which are then used in robot operations.

This is also a future collaboration trend in the robotics industry:
as robot movements become more precise, everyone may need to contribute their own action data.

Massive information, precise analysis, all on Sina Finance APP