Over the past few years, I have shared various incidents on this platform involving robots that appeared to advance toward or make dangerous movements near humans. When we view these cases in sequence, a problem emerges that goes far beyond the notion of a so-called "robot rebellion." One of the most striking incidents took place in February 2025 during a festival in China.

The audience was watching the humanoid robots from behind a barrier when a Unitree H1 robot began moving toward the crowd; suddenly, the machine lurched forward. Footage shows security guards quickly stepping in front of the humanoid and pushing it back. The video went viral, accompanied by claims that a robot had attempted to attack humans.

We featured this case in my posts and raised a question that the footage alone could not answer: did the robot actually try to strike someone, or did it simply lose its balance and make sudden movements while trying to stabilize itself? Local authorities treated the incident as an unexpected anomaly during the demonstration; there was no evidence that the humanoid had made a conscious decision to attack the people.

A few months later, however, another video featuring a Unitree humanoid drew even more attention; this time, the setting wasn't a festival but a testing environment. An H1 model appears suspended from a structure while two men work near the machine. At first, everything seems normal, but then the robot begins to violently thrash its arms and legs. The movements rapidly intensify, dragging the structure along; a monitor falls, and the men are forced to step back as they try to regain control of the machine.

We recently showed that footage here again because, viewed in context, it appears to show a robot in a state of genuine rage. Yet, there is a fundamental difference: this was not an artificial intelligence deciding to attack the engineers. The footage actually shows a control failure during testing. There is an important detail here: even without any malicious intent, a machine weighing tens of kilograms—equipped with powerful motors and joints capable of rapid movement—can pose a real danger when things go wrong. In fact, this issue predates the current generation of humanoids.

In 2023, an accident involving an industrial robot occurred at Tesla’s Giga Factory in Texas; an engineer working near production equipment was pinned and injured by one of the robots. The incident quickly made headlines claiming that "a Tesla robot had attacked an engineer," but again, there was no evidence of intent. It wasn't a rogue AI; it was an accident involving an industrial machine. Perhaps that distinction is becoming increasingly important, given that traditional industrial robots usually operate in controlled areas, separated from people by safety protocols.

Humanoids are being developed for the exact opposite purpose: they need to enter our environment, walk among people, work alongside us, navigate stores, enter hospitals, operate in factories, and—in the future—enter our homes. At the same time, their physical capabilities are advancing rapidly; today, there are already humanoids capable of running, jumping, withstanding impacts, and performing martial arts movements.

Unitree itself—the model apparently serving as the basis for the Russian robot that attacked a person in a store—has already been demonstrated performing punches, kicks, and complex kung fu sequences. A software error on a computer typically generates an on-screen message; however, a software error in a physical machine can trigger movement, and the stronger, faster, and heavier the machine, the more severe the consequences of that error can be.

Now, another issue has arisen: research has revealed that a single malicious command could compromise humanoid robots and, in certain scenarios, cause the attack to spread to other vulnerable machines—adding a new dimension to the problem. We do not need to imagine a robot developing rage or consciousness for a risk to exist; a machine can carry out a dangerous action simply because it received an erroneous command. Was it because a sensor misinterpreted the environment? Did it lose its balance? Was there a programming glitch? Did someone take remote control of it? Or was the system compromised?

And now we return to that shop in Russia; this time, we are not looking at a robot inside an isolated industrial zone, nor suspended in a structure during testing, nor separated from the public during a technology demonstration. We are seeing a physically powerful machine inside a shop, surrounded by ordinary people—and that may well be the most important part of the entire story. These incidents do not show that robots are developing a will to attack human beings, but they do demonstrate something far more concrete: the more these machines leave the laboratory and enter our world, the less room there is for error. After all, a humanoid does not need to *want* to hurt someone; it simply needs to execute the wrong movement at the wrong moment.