Gig Workers Train Humanoids at Home

Gig workers strap iPhones to their foreheads to train robots.

In a Palo Alto startup’s quiet race to build humanoid assistants, thousands of contract workers in more than 50 countries are becoming the new data engineers. Micro1, a US company, collects real-world footage—snapping chores, folding laundry, washing dishes—captured by workers like Zeus, a Nigerian medical student, who records himself performing everyday tasks with a head-mounted camera. The videos feed robotics firms’ models, a data stream that’s becoming as vital as any silicon chip in the push to move humanoids from lab to living room.

The scale is striking. Micro1 has hired “thousands” of gig workers across continents, paying local-rate wages that look generous in many places. Zeus describes a job that is part opportunity, part industrial choreographer: mount the iPhone, step through routine actions, and hope the frame captures the subtlety needed for a robot to mirror human movement. Companies such as Tesla, Figure AI, and Agility Robotics are chasing the same dream—humanoids that can operate in homes and factories with human-like dexterity—so these at-home recordings have become a hot, contested data asset.

The arrangement is a double-edged sword. On one hand, real-world footage across varied environments creates a more robust training signal than staged data ever could. On the other, it raises thorny issues about privacy, informed consent, and worker protections. The footage often depicts intimate, ordinary moments—laundry, cooking, improvised problem solving—exposing participants to questions about surveillance, consent, and how their likeness and behavior will be used long after a single clip is captured. The conversations about consent aren’t abstract: critics worry that workers may not fully grasp how the data will help train robots that could operate in sensitive spaces, from hospitals to homes with children.

The commentary around these practices arrives at a pivotal moment for AI evaluation. A separate thread in The Download notes that AI benchmarks, historically measured in isolated tests, are failing to reflect real-world performance. The push is for benchmarks that assess AI behavior over longer horizons—what happens when a model keeps operating in a messy, multi-person environment over time. The implication for robot training is direct: even with a mountain of real-world video, the true test is whether a humanoid can plan, adapt, and cooperate across tasks that unfold over minutes and days, not just milliseconds.

Analysts say the data approach is viable but expensive and brittle. The data quality depends on how consistently workers perform tasks, how well the camera setup captures fine-grained motion, and how representative the footage is across homes, kitchens, and workspaces. Data provenance and labeling integrity become mission-critical: one off-day’s lighting or camera angle can ripple into model errors in a real robot.

From a product perspective, the implications are clear. First, data costs are nontrivial: hosting, labeling, and curating thousands of hours of video, plus privacy safeguards, add up. Second, privacy controls and consent mechanisms must be robust and auditable to avoid missteps that could trigger regulatory or reputational backlash. Third, the industry’s current bench market—short, isolated tests—may fail to predict performance in dynamic homes, pushing teams toward more sophisticated evaluation that mirrors Zeu’s world.

Analysts offer one vivid analogy: this is like scaffolding for a skyscraper built with data instead of steel. The scaffolding is essential to reach higher floors, but the real structure depends on how the building materials behave when they finally bear weight. In robotics, that weight is a humanoid’s ability to navigate real rooms with unpredictable people and objects, not just pass a lab test.

What this means for products shipping this quarter is pragmatic but nontrivial. Expect rosters of at-home data collection to remain a foundational, costly channel for training, paired with heightened attention to consent and privacy. Companies will likely accelerate toward stronger, long-horizon evaluation protocols and consider synthetic or simulated data to augment real-world footage and address edge-case failures before hardware hits consumer-facing channels.

Zeus’s work sheds light on a broader shift: AI systems tuned on living-room footage are becoming a staple of the robotics race, even as the industry debates how to measure genuine, long-term capability.

Sources

The Download: gig workers training humanoids, and better AI benchmarks

The gig workers who are training humanoid robots at home