Humanoid Robots Trained by Gig Workers Filming at Home

Humanoid robot training for Tesla, Figure AI, and Agility Robotics now relies on a hidden global gig workforce — raising serious questions about privacy, labor rights, and the true cost of AI automation. Zeus, a Nigerian medical student, straps an iPhone to his forehead every week and records himself washing dishes, folding laundry, and cooking. He earns $15 an hour. His footage goes to Micro1 — a U.S.-based robotics data firm — and from there, it trains humanoid robots being built by Tesla, Figure AI, and Agility Robotics. "This would be a real nice opportunity to set a mark and give data that will be used to train robots in the future," he told MIT Technology Review.

But he also admitted: "I really do not like it so much." Zeus is a gig worker (a freelancer taking short-term digital tasks through an online platform, like a global Uber for data work) — and he represents a quiet revolution in AI labor the robotics industry has barely acknowledged. Thousands like him, spread across more than 50 countries, are doing the repetitive, invisible work that makes humanoid robots possible. This is the hidden workforce powering the $100 billion humanoid robot boom of 2026.

Why Humanoid Robot Training Requires Real-World Home Data

Teaching a robot to fold a shirt is genuinely hard. Unlike chess, where every rule is fixed, household tasks involve endless variation — different fabrics, lighting conditions, table heights, grip angles. The most effective training approach is to show robots thousands of hours of real humans actually doing these things in real homes.

This is called embodied data (real-world physical movement footage used to teach robots body mechanics — how people grip objects, balance, pivot, and adapt on the fly). Generating this in a controlled lab is slow and expensive. Outsourcing it to gig workers across 50+ countries is fast and cheap.

Micro1 coordinates this network. Workers strap head-mounted smartphone harnesses to their foreheads and perform ordinary household tasks. The footage is labelled and sold to:

• Tesla (Optimus) — training its bipedal robot for factory and home environments
• Figure AI — building human-like movement models for logistics and household automation
• Agility Robotics (Digit) — focused on warehouse and retail use cases

Each company ultimately wants its robot to operate autonomously in human environments. To get there, it needs millions of data points showing how real human bodies move through real human spaces. Gig workers are providing exactly that — at scale, globally, at a fraction of what in-house lab data would cost.

Privacy Risks in AI Training Data Collection

Here's where things get uncomfortable. Workers like Zeus are required to use pseudonyms and are explicitly not authorized to discuss their work publicly. When MIT Technology Review tracked him down, he was speaking in violation of his agreement. His real name is not Zeus.

The informed consent picture is murky. Workers know their recordings will be used for AI training — but the full scope of commercial downstream use (exactly which companies buy it, which products it ends up in, for how long it's stored) isn't made transparent. And all of this happens inside workers' actual homes: their kitchens, living rooms, hallways. Family members can appear in frame. Personal spaces are captured in detail.

Zeus doesn't love his situation. He's "the kind of person that requires a technical job that requires me to think," he told MIT Technology Review. The task — repeating the same domestic motions for hours while an iPhone films from his forehead — is intellectually unfulfilling. Long-term, that's also a data quality problem: disengaged workers produce noisier, less consistent training footage, which makes the robots they're training less reliable.

Are AI Automation Benchmarks Measuring the Right Things?

Angela Aristidou, a professor at University College London and researcher at the Stanford Digital Economy Lab, argues there's a deeper structural problem: "We need new benchmarks that assess AI's performance over longer horizons within human teams, workflows, and organizations."

Traditional AI benchmarks (the standardized tests used to score how capable an AI system is — roughly equivalent to a school exam for robots and software) evaluate models on isolated, self-contained problems. A coding challenge. A single image question. A math puzzle. They don't test how an AI system performs embedded in a real organization over weeks, navigating shifting priorities and human dynamics. The result: systems that score impressively on paper while underperforming badly in practice — and no benchmark warned you this would happen.

The Supply Chain Crisis Behind AI Automation Hardware

While the humanoid robot industry chases massive funding, a supply chain crisis is quietly building that will raise the cost of every component involved — including the robots themselves.

Crude oil has recently topped $100 per barrel, driven by disruptions near Iran's Strait of Hormuz. U.S. gasoline is averaging $4/gallon, the highest since 2022. But the bigger downstream problem is plastics.

Plastics are made from naphtha (a petroleum-derived liquid extracted during oil refining — the primary raw material for plastic production worldwide). The Middle East produces approximately 20% of global naphtha and supplies roughly 40% of Asia's market. When oil supply gets disrupted, naphtha prices spike. This past month, naphtha prices in Asia have already surged 50%.

The downstream consequences are already materializing. One Indian water bottle manufacturer raised prices by 11% after its packaging costs jumped 70%. Global plastics production stands at 431 million metric tons per year — and there is no scalable substitute ready.

Bio-based plastics (plastics made from agricultural feedstock like corn or sugarcane, rather than petroleum) make up just 0.5% of the global market today and are projected to reach only 1% by 2030. They cost 2–4x more than conventional plastic and compete directly with food production for agricultural land. Mechanical recycling (the process of melting down used plastic to form new material) degrades the plastic each cycle and cannot be repeated indefinitely. Chemical recycling is highly polluting and rarely converts material back to usable plastic at commercial scale.

This matters for the tech and AI industry specifically: everything runs on plastic. The smartphones gig workers use to record training data, the robot chassis components, PCB casings (the protective plastic housings around circuit boards inside every computer), cable insulation, packaging for every hardware shipment — all of it depends on petrochemical stability that is currently anything but stable.

The average U.S. resident consumes over 250 kg of new plastic annually — more than 4 times the global average of 60 kg. The U.S. tech industry sits at the center of that consumption, and a 50% naphtha price surge doesn't stay in the water bottle aisle. It ripples through every manufactured good.

$122 Billion, 18 Threatened Companies, and AI Automation's Fragile Foundation

OpenAI just closed a $122 billion funding round — the largest in Silicon Valley history — with an IPO expected in 2026. That capital signals enormous confidence in AI's commercial future. But the foundation underneath that future has visible cracks.

Iran's Islamic Revolutionary Guard Corps (IRGC) has explicitly threatened 18 U.S. tech companies — including Nvidia, Apple, Microsoft, and Google — with retaliation for operating in the Middle East, stating via Telegram: "From now on, for every assassination, an American company will be destroyed." The same regional tensions destabilizing oil prices are now directly targeting the chip manufacturers, cloud infrastructure providers, and software platforms that AI runs on. These risks don't cancel each other out — they compound.

The picture that emerges is this:

• AI training pipelines rely on gig workers operating under opaque consent conditions in 50+ countries
• The benchmarks used to evaluate AI quality don't measure real-world workflow performance
• Hardware supply chains for every AI device depend on petrochemical markets being disrupted right now
• The geopolitical forces threatening oil supply are the same forces threatening cloud and chip infrastructure

For Zeus, who genuinely hoped this work would let him "set a mark" on the future of robotics, the reality is more complicated than the brochure. The job is repetitive and intellectually unfulfilling. The data he produces will likely help automate work in his own country before anywhere else. His privacy is protected by a pseudonym, not by meaningful consent design. And at $15/hour, he is an inexpensive solution to an expensive problem — until a better one comes along.

The practical takeaway: test AI tools against your actual workflows rather than trusting benchmark scores. And keep an eye on oil prices — they will tell you where hardware and component costs are heading before any tech publication does.

Related Content — Get Started | Guides | More News