The Hidden Labor Problem Inside the Humanoid Robot Boom

Key points

• Humanoid robotics companies raised billions of dollars over the past 18 months, much of it funding large-scale human teleoperation programs.
• Teleoperation datasets, where humans record themselves doing tasks to teach robots, are more than 100,000 times smaller than the datasets used to train today's language models.
• Workers across China, India, Europe, and the US are being hired to film household tasks and operate robots remotely, forming a commercial data-supply industry.
• Nikita Rudin, co-founder and CEO of robotics startup Flexion, argues that reinforcement learning in simulation is the only approach that can break the dependency on human demonstration data.
• Flexion raised 50 million dollars in funding from DST Global and NVentures to build what Rudin calls a general-purpose "brain" for humanoid robots.

The pitch for humanoid robots has always been straightforward: humans won't be able to fill certain jobs in the future, so robots will step in. But a pointed opinion piece published by The Robot Report this week raises an uncomfortable question. What if the robots themselves need a permanent supply of human workers just to function?

The issue is teleoperation. This is where a human operator, often sitting far away, controls a robot through a remote device, the same way you might steer a toy car with a controller. Every movement gets recorded. Then the robot studies that footage and tries to copy it.

Teaching a robot this way sounds sensible. The problem, according to Nikita Rudin, who completed his PhD at ETH Zurich's Robotic Systems Lab and worked on robotic simulation tools at Nvidia, is that it scales terribly.

A shelf moves. A door handle is slightly different. A new box shape arrives on a factory floor. Each tiny change means someone has to film the task again from scratch. The real world never stops changing, and the human workforce generating demonstrations can never keep up.

The quality of that data is shaky too. Operators controlling a robot remotely can't feel what the robot is touching or judge distance accurately, so they move slowly and overcorrect. The robot then learns from footage of someone struggling with a controller, and struggling is exactly what it practises.

Will more human operators actually solve this?

No, and that is the core of Rudin's argument. He draws a sharp comparison to the early days of large language models, the AI technology behind chatbots such as ChatGPT. Those early models were trained on enormous amounts of text and could loosely copy the style of Shakespeare, but couldn't actually reason. The real leap came through reinforcement learning, a training method where software learns by trying, failing, adjusting, and trying again across millions of attempts, with no human telling it what to do at each step.

Rudin argues humanoid robotics is stuck at that earlier, imitation-only stage. More teleoperation data just means more imitation at massive cost.

The alternative he describes is reinforcement learning inside computer simulations. A robot's software can fail millions of times inside a virtual world, reset instantly, and try again, running in parallel across many computers at once. More computing power directly means faster progress. That is a very different equation from hiring more workers in lower-wage countries to film themselves doing laundry.

For the people doing that teleoperation work today, and for investors funding these companies, Rudin's challenge is direct: if robot autonomy is genuinely the goal, the data should show that human involvement is decreasing over time. If it isn't decreasing, teleoperation has stopped being a bridge and quietly become the permanent method.