Find out how to keep away from the teleoperation entice in robotics improvement

Flexion is building a reinforcement learning and sim-to-real platform for humanoid robots. Source: Flexion

In the past 18 months, humanoid robotics companies have raised billions of dollars – a majority of which is quietly funding hiring humans to operate robots. This means the robotics industry has a teleoperation and data problem it keeps describing as a labor solution.

Teleoperation and human demonstration at scale have become the dominant method for training physical AI systems, attracting serious capital, recruiting workers across lower-wage economies, and earning enthusiastic coverage as evidence of progress. The assumption underneath all of it is that enough demonstrations will eventually produce robots capable of generalizing across real environments.

I believe that assumption deserves a lot more scrutiny than it’s getting.

Teleoperation hits a structural wall

Language models trained on text can draw from decades of writing, articles, and books. With robots, there’s no archive to draw from – someone has to generate every demonstration, which means the data can only grow as fast as human labor allows.

Teleoperation datasets are over 100,000 times smaller than what’s used to train today’s language and vision models. That gap doesn’t close by hiring more operators, because the real world never stops changing: A shelf moves, a door handle is slightly different, and a new package type shows up on the line. Every variation requires a new demonstration, meaning the problem grows faster than the workforce can.

Data quality is the other issue. Operators can’t feel what they’re touching or judge depth reliably, so they move slowly and overcorrect. This forces the robot into learning from footage of someone struggling with a controller, and that’s what it ends up practicing.

Save the date for RoboBusiness 2026

The human cost of data generation

The industry’s answer to the data problem has been to recruit more people, predominantly workers in lower-wage economies, hired to film household tasks, operate robots remotely, or move through facilities wearing camera rigs.

A whole commercial ecosystem has emerged around it, with startups across China, India, Europe, and the U.S. selling teleoperation data the same way companies once sold labeled text for language models.

The original pitch for humanoid robots is that humans won’t be able to fill these jobs in the future due to demographic shifts, labor shortages, and aging populations. But if what we’re actually building is infrastructure that requires a permanent stream of human demonstrations to function, then we might as well have those humans do the task directly.

A system that can’t handle anything new without fresh human input is essentially just a labor system.

A convenient defense for robotics improvement

The standard response is that teleoperation is a bridge – a way to get started while better robot model training methods catch up. For narrow, repetitive tasks in controlled environments, that’s fair.

But what much of the industry is actually building is infrastructure for generating demonstrations indefinitely, with no clear account of how or when that changes.

The field is tracking what’s easy to count – demonstrations collected, hours of footage logged, tasks completed in controlled settings – none of which tells you whether the robot can handle something it hasn’t seen before, in a place that wasn’t set up for it. Building more of the same infrastructure deepens that dependency on humans rather than resolving it.

Flexion’s full autonomy stack includes a command layer, a motion layer, and a control layer. Source: Flexion

The path that fits the problem

When researchers trained early language models on massive amounts of text, they got systems that could loosely imitate the style of Shakespeare, but produced words that didn’t quite make sense. Impressive on the surface, but not yet capable of reasoning.

The breakthrough came through reinforcement learning in synthetic environments, which produced systems capable of reasoning, coding, and following complex instructions.

The robotics industry is largely stuck in that early moment. Scaling teleoperation data is the equivalent of scaling pre-training text on 100,000x less data. You get robots that somewhat move their arms, sometimes grab something, sometimes don’t. They can vaguely imitate what a human operator showed them, but they can’t reason through a situation they haven’t seen before.

There are approaches that sit between traditional teleoperation and full autonomy – egocentric video capture and devices like the Universal Manipulation Interface (UMI), which lets operators demonstrate tasks more naturally by wearing a handheld gripper rather than controlling a robot remotely.

These methods reduce the burden on operators and produce somewhat more natural motion data. They still require humans in the loop, but they’re less invasive, and for narrow, well-defined tasks, they can be useful stepping stones. That said, they don’t resolve the industry’s full dependency.

Reinforcement learning is what changes this. Rather than imitating what a human operator showed it, a system trained with RL figures things out through trial and error: attempting a task, failing, adjusting, and trying again across millions of iterations, without a human in the loop.

Simulation follows naturally from that; running millions of RL iterations in the real world destroys hardware and takes years. In simulation, you reset instantly, run in parallel, and generate variation at a scale no human workforce could match. And unlike teleoperation, it scales directly with compute; more GPUs mean more environments, more variation, and faster iteration.

The people doing teleoperation work deserve to know if autonomy is the actual goal, and so do the people funding these projects. If you don’t have data showing the dependency on humans reduces over time, teleoperation moves from a stopgap to the permanent method.

About the author

Nikita Rudin is co-founder and CEO of Flexion. Rudin completed his Ph.D. at the Robotic Systems Lab at ETH Zurich while working at NVIDIA, where he focused on large-scale reinforcement learning, control systems, and robotic simulation.

At NVIDIA, Rudin was part of the team behind Isaac Gym and Isaac Lab, simulation tools now widely adopted across the robotics industry. Now he is leading Flexion, which recently raised $50 million from DST/NVentures to build the general-purpose “brain” for humanoid robots.

Source link