The founder of a robotics company recently told Silicon Valley Arrival a startling industry figure: “In fact, 90% of robots are remotely controlled.”

Tesla’s Optimus robot has repeatedly faced questions over “remote control,” and its share price fell in response; at the end of 2025, 1X released its NEO home robot, only for people to find that all of its complex movements were being remotely controlled by humans.

Since then, doubts about remote-controlled robots have not stopped.

We thought we were seeing the future. The reality is that humans are hiding behind machines, using remote controls to prop up a grand technological illusion...

Have We Been Fooled?

Tesla’s Optimus alone has been caught up in three separate “remote-control” controversies.

The first began in January 2024, when Elon Musk made a high-profile release of a video showing Optimus folding clothes. The robotics industry erupted.

A robot doing a backflip is not especially impressive; that can be programmed. But a robot that can adapt to endlessly changing soft objects, such as folding clothes, is genuinely impressive.

That is why the industry has long treated “whether a robot can fold clothes” as a core benchmark of high-level manipulation ability.

Musk had barely had time to celebrate before internet users noticed what looked like the shadow of a human hand in the lower right corner, moving in sync with the robot.

Tesla eventually admitted it had relied on remote operation. At that point, questions about robot teleoperation began entering public view.

In October of the same year, after the Cybercab launch event, tech bloggers confirmed that scenes showing Optimus pouring beer and handing out popcorn were remotely operated.

The Wall Street Journal and other outlets then began aggressively following the story. It briefly weighed on Tesla’s stock, which fell 8.78% that day.

The episode did not end there. In December 2025, at an event in Miami, Optimus fell while handing out bottled water. Before the fall, it made a movement resembling a human removing a VR headset, triggering an uproar.

That contradicted Musk’s earlier, high-profile claim that Optimus was “fully autonomous.” The Wall Street Journal and other media outlets reported on the issue again, Tesla’s stock fell 3.39%, and public skepticism reached a peak.

1X’s release of its NEO home robot last year also caused a sensation. But after The Wall Street Journal tested it, reporters found that its complex tasks all depended on an “expert mode.”

In other words, real people in the back end were remotely controlling it through VR, with only a small number of actions labeled “automatic.” Buying a robot with a cloud-based nanny? The industry was stunned.

After that, several robotics companies in China were also reported to have been involved in remote-control incidents.

For a time, the questions kept coming, and public opinion boiled over: “Have we been fooled? Are so-called robots just large, expensive remote-controlled toys?”

The Brutal Truth: 90% Are Remotely Controlled

Luo Hao, the founder of a robotics company, told Silicon Valley Arrival a striking figure: at present, fully “true” autonomous robots on the market may account for less than 10%.

And even those are limited to simple walking, fixed waving gestures, and highly standardized repetitive motions on production lines.

Semi-remote operation, known in the industry as Human-in-the-Loop, accounts for roughly 60% to 70%.

Examples include Tesla folding clothes, Figure making coffee, and 1X doing housework. These difficult actions still largely require human intervention, or completion with human assistance.

“Pure teleoperation, often used to push up valuations, is also common, accounting for about 20%,” Luo said. In other words, more than 90% of the robotics industry involves some form of remote control.

He once saw a peer use a remotely controlled robot for a demo and “raise 5 million yuan in angel funding.”

Some robotics companies have even used remote-control tricks to fraudulently obtain robotics subsidies at the national level.

The industry is now filled with an extremely impatient mood: those who make videos do better than those who build technology, and those who act do better than those who write code.

In reality, at the current stage of humanoid robot development, “remote control” is not necessarily meant to deceive. It is an unavoidable “apprenticeship.”

Today’s robots are like newborn infants, with blank brains. To teach them how to tighten screws or move boxes, the most efficient method is to have humans put on motion-capture equipment and guide them through the task step by step.

Every act of “remote control” is, in essence, feeding the AI brain the most valuable kind of “physical-world data.”

An operator puts on a motion-capture suit and remotely controls a robot, almost like acting, to fold clothes 10,000 times. Only after the AI records those 10,000 rounds of data might it be able to perform the task autonomously on the 10,001st try.

Yet the industry has been unwilling to speak openly about this perfectly reasonable evolutionary process.

“That is because investors and the public do not have the patience to wait for AI to grow up slowly. They want results now,” Luo said. To maintain high valuations, remote control has become a “magic weapon.”

Luo himself admits that at some trade shows, to avoid mistakes, he also uses remote control to “assist” the robot.

The industry is also seeing bad money drive out good. “Many platforms have secured investor funding and government subsidies by relying on remote control, raising expectations across the industry.

If you do not perform, you cannot get money, and you cannot get public attention.” That is the harshest truth.

When the fig leaf is pulled away, the public feels the anger of having been “played.”

Expectations for robots are too high. The reality is that robots will awaken, but absolutely not that quickly.

But this is not necessarily a bad thing. Every technology matures at the cost of being demystified.

The Way Out

Now that we know the biggest constraint holding back the robotics industry, is there no way to solve it?

There is. To cross the chasm of “data training” and free robots from remote controls, the tech world has developed three main schools of thought. The first is the ascetic path of end-to-end imitation learning.

Tesla and Figure are currently in this camp.

As long as there are enough remote-control demonstrations, AI can learn directly through neural networks.

OpenAI and Tesla are both betting on a mathematical pattern: when data volume and computing power become large enough, intelligence will suddenly “emerge.”

ChatGPT is one example. No one taught it translation, and no one taught it to write code. Simply because it read enough books, it suddenly “got it.”

Musk believes robots are the same. As long as he feeds them enough “action videos,” they will eventually develop “common sense about the physical world.”

At that point, you will not need to teach a robot how to cut a cucumber. It will look at the knife and the cucumber and know what to do.

The downside is that we do not know when machines will awaken, or how long it will take. Factory production lines cannot wait, and neither can bosses’ patience.

The Matrix-Style Sim-to-Real Training

For now, the companies trying to do this are mostly tech giants, including NVIDIA and Tencent.

Start with Jensen Huang’s Isaac. He wants to use the computing power of super chips to simulate a virtual world, incorporating every data dimension of the physical world.

In this way, robots can train virtually inside that world, and time can be sped up by a factor of 1,000.

Tencent’s product is called “Tairos,” with the Chinese nickname “Titanium Screw.”

Tencent is not directly competing with NVIDIA. It has built a super-connection system whose underlying layer can even plug into NVIDIA’s system.

The core competitiveness of “Titanium Screw” is that it can allow millions of robots to train simultaneously in the cloud, much like Honor of Kings supports hundreds of millions of people online at the same time.

But the biggest problem with this model is the “reality gap.”

Simulators have no dust or oil stains, and physical friction is perfect. But in the real world, the floor may be dirty or wet, tires may be worn down, and friction can change at any time.

Perfectly simulating the real world may still be a long way off.

That gave rise to the third school: Lego-style “modular skill orchestration.”

Leju Robotics’ Taskor and Flowstate, a platform launched by Google-owned Intrinsic, are among the key players.

Before “machine awakening,” and before the “reality gap” is closed, the industry needs a bridge.

They are that bridge. They do not expect robots to achieve sudden enlightenment in a virtual world. Instead, through standardized packaging, they directly reuse the most stable actions that have already been validated in the real world.

In a factory, you may only need a porter, a loading and unloading worker, or an inspector.

So pragmatic players such as Taskor and Flowstate only need to put together “building blocks” such as moving, grasping, and placing, and assemble them into a workflow.

The difference is that Taskor focuses on humanoid robots;

Google’s Flowstate focuses on industrial robotic arms.

In one sentence: Musk wants to build a “robot” that can be used in any scenario.

Leju and Google, by contrast, believe large-scale robot deployment can begin with “working in factories.”

At this stage, the goal is not to make robots think like humans, but to make them finish the job in a specific work environment with less trouble.

Looking back at the controversy over “remote control,” what angered the public was not the technology itself, but the sense of being fooled.

The real future will not be faked with remote controls. It will grow slowly through patience, data, and time.

Before robots truly learn to walk, we should at least learn to be honest...