Getting Unitree’s G1 to perform kung fu takes more than an accidental tap on a controller. First, the operator must hold two nonadjacent buttons for three seconds to enter dance mode. Then they must hold a combination of six different buttons for a specified period. The source report doesn’t say how long that second combination takes.
That sequence matters when watching a humanoid robot kick. According to János Körömi, managing director of INFUZE Robotics, these are operator-triggered routines, not a machine independently deciding to attack someone.
INFUZE Robotics, the Central European representative of Chinese manufacturer Unitree Robotics, demonstrated the G1 to Hungarian publication Rakéta in a Budapest parking lot. The robot ran and performed the kung fu movements familiar from online videos, following commands from a button-covered remote resembling a game console controller.
The demonstration followed footage of a G1 kicking a remote out of a tester’s hand. That clip prompted debate about why a robot could perform such combative movements in the first place. But a legged machine’s ability to kick isn’t inherently surprising: humanoids are being trained to execute increasingly complex movements, including actions resembling sports and martial arts.
Reports about AI agents supposedly going rogue, alongside wider concerns about AI safety, can encourage alarming interpretations of these physical demonstrations. Körömi’s explanation was much less cinematic.
“There is no such thing as going rogue, because everything is controlled. Our robots do not go anywhere on their own without having received a command to do so. That is why the expression ‘going rogue’ does not hold up in itself, because the robot does not make decisions on its own. This is a computer with arms and legs. If you think about it, Excel rarely escapes from our computer, a mobile application does not escape either, and neither does a robot,” he told Rakéta.
That is the distributor’s account of how these robots operate, not evidence that a moving machine cannot cause harm. Körömi said the complicated button combinations prevent an operator from accidentally triggering a routine by pressing the wrong control. He described the dances as entertainment movements deliberately called up by a person.
Culture also affects how those movements look to an observer. In Körömi’s view, kung fu may appear aggressive through European eyes, whereas in Asia it is also understood as an art form, including a kind of dance performance. Starting such a routine in a cramped shop, as in another circulating video, is therefore the decision and responsibility of an operator who has misjudged the space, he said. The robot doesn’t initiate those routines by itself.
The models have an emergency-stop button so an operator can intervene when necessary. Someone might start a routine and then realize there isn’t enough room, or an inexperienced operator might discover they have selected a different dance than expected. A robot could also fall and attempt to get up in an unsuitable spot. If it becomes tangled in a bench, for example, that recovery movement could damage the robot or something around it.
What kung fu footage does, and doesn’t, demonstrate
Rakéta’s assessment is that dancing and acrobatic demonstrations primarily reveal progress in balance, coordination and control. They don’t establish the arrival of combat-ready, autonomous android units. A convincing kick shows physical capability; it doesn’t, by itself, demonstrate independent judgment.
Still, increasingly lifelike footage has fueled online speculation that China might be preparing robot soldiers. Apocalyptic stories in popular culture make a Terminator-style scenario easy to picture. The source article argues that these conspicuous, expensive and difficult-to-program machines are an implausible choice for that purpose.
Körömi also pointed to restrictions on their use. “This is also strictly prohibited. Unitree is a civilian robot manufacturer, and its products must not be modified in ways that would make them usable as offensive weapons. All the major robot manufacturers have a declaration on this. If the customer breaks the rules, the manufacturer can revoke their permissions,” he said.
Rakéta makes an economic argument as well: commercially available quadcopters can, it says, be converted into lethal, precise weapons for a few hundred thousand Hungarian forints. By comparison, the publication considers spending several times as much on a much less effective humanoid platform illogical. That is its reasoning against the Terminator scenario, not proof that misuse is impossible.
Some footage fueling those fears isn’t real in the first place. A circulating video depicting humanoids undergoing military training with firearms was subsequently identified as AI-generated, according to Rakéta. The report says numerous synthetic videos feature the G1 because it is already a relatively widespread model internationally.
Körömi said those clips can be particularly difficult to assess for someone who has never seen such a robot in person. Excessively humanlike motion can be a warning sign, as can arms or hands bending at more points than the hardware allows.
He also identified model-specific clues. The G1 shown in the report has a blue-lit head; a different lighting color could be suspicious. According to Körömi, footage showing that robot moving its head immediately gives away the fabrication, because this particular robot cannot do that. These observations concern the machine he described, not a universal test for every humanoid.
Why a T-shirt is harder than a rehearsed kick
The less theatrical limits become clearer when a robot has to interpret its surroundings. The report describes navigation as relying on machine vision to identify shapes, colors and forms. Recognizing something isn’t always straightforward, especially when the object changes shape as it moves.
Textiles are a major example. A T-shirt lying flat looks different from the same shirt lifted into the air. Its changing shape and surface appearance make it a particularly difficult machine-vision problem. Performing a rehearsed movement and understanding an unfamiliar household object are different challenges.
“As we grow up, we acquire a great deal of knowledge about our world through experience, because we live in it. A small child learns that water is wet and fire is hot, so it can burn their hand. This shared knowledge is very difficult to translate into the language of IT and program code,” Körömi explained.
A robot can interpret an unfamiliar environment and carry out some tasks to a certain extent, he said, but that ability remains limited. In his explanation, this is why the robots under discussion are not primarily intended for households. Compared with a production line, a home is a jungle.
On a production line, relatively similar events happen with relatively similar objects. Living spaces contain many different objects and activities, while pets introduce behavior the robot cannot predict. Körömi argues that these limits on understanding the world are also a reason not to expect the machines to become excessively independent anytime soon.
Rakéta’s photographs by Zsuzsanna Dömös show the G1 standing still and its head illuminated blue, alongside a separate image of the Unitree H2 and Unitree R1. Each is a physical machine with specific capabilities and limits. For the G1 demonstration described here, the immediate safety question is concrete: has the operator left enough room for the movement they are about to command?
