Hidden labor powers humanoid robots, critics warn

The real engine behind humanoid robots isn’t silicon—it’s humans in exosuits.

A Technology Review investigation pulls back the curtain on a truth many demos don’t reveal: the flashy “physical AI” we’re seeing in labs and stage demos relies on weeks of grueling, largely invisible human work. Nvidia CEO Jensen Huang has publicized a shift toward a future where AI doesn’t just talk or analyze—it moves and handles objects in the real world. But the story makes a stark counterpoint: the work that makes those demonstrations possible is often hidden, outsourced, and paid in ways that readers and customers rarely see.

In Shanghai, a worker wore a virtual-reality headset and an exoskeleton while repeatedly opening and closing the door of a microwave to train the robot next to him—hundreds of times a day, the piece notes. Across North America, robotics firms such as Figure are pursuing similar human-in-the-loop data collection strategies. The implication is not that the robots don’t work; it’s that the data and demonstrations that teach them new skills are generated by people in controlled environments, under supervision, and at scale. The result is not just a better arm trajectory or a smoother grip—it’s a training regime that looks more like a factory workflow than a purely autonomous software stack.

That revelation has broad implications for how we assess what “physical AI” can actually deliver. On one hand, human-guided learning can accelerate capability in perishable, real-world tasks (opening doors, manipulating small objects, coordinating with other machines). On the other hand, it raises questions about cost, scalability, and ethics. If a robot’s cleverness hinges on rounds of human demonstration and annotation, the long-run economics resemble a hybrid model: hardware plus a labor-intensive data layer, with all the wage, fatigue, turnover, and regulatory considerations that entails. And as consumer-facing demonstrations proliferate, the public risk is a mismatch between what’s marketed and what’s actually happening behind the scenes—a mismatch that could fuel disillusionment or regulatory scrutiny.

From a practitioner’s lens, two concrete realities stand out. First, the economics of humanoid robotics are intertwined with labor costs. A week-long VR/exoskeleton training sprint for a single task can be a nontrivial line item, potentially eroding the efficiency gains that a pure hardware-focused play promises. Second, there is a quality and reliability concern: human performance varies. If the same skill is demonstrated by multiple people or in different contexts, the resulting robot behavior must be harmonized, tested, and validated to avoid drift in performance once deployed at scale. This implies robust QA pipelines and governance around who trains the models, how the data is curated, and how updates are rolled out across fleets of robots.

Two more takeaways for product teams: be prepared for disclosure and compliance to catch up with capability claims; and design roadmaps that prioritize tasks with clear ROI from human-guided learning, or invest in strategies to reduce dependence on heavy human-in-the-loop data through simulation, self-supervision, or modular transfer learning. The industry may still race toward impressive demos, yet the real gating items—labor costs, worker protections, data provenance, and verification—will shape whether these systems ship at scale this quarter or linger in pilot programs.

As the industry markets “physical AI” as the next leap, the article reminds us that people remain the most consequential variable. The tech itself may be transformative, but its adoption will hinge on governance, transparency, and the ability to scale responsibly without leaving workers exposed to the margins of the automation economy.

Sources

The human work behind humanoid robots is being hidden