In the footage, a G1 humanoid wearing red gloves faces off against a human trainer wearing protective gear; the man approaches, moving his arms and attempting to land a few blows, while the robot shifts its stance, adjusts its footing, dodges, and counters with punches and kicks. However, the difference compared to previous encounters lies not in the force of the blows, but in who decides the next move. According to Unitree, the G1 was controlled by Unifold LMX2 1.0, a new AI model developed to interpret the environment and plan physical actions in real-time—without teleoperation or a pre-programmed sequence of movements.

The system employs an approach known as a "world model"; rather than simply receiving a command to raise an arm or take a step, it analyzes unfolding events and attempts to predict how the situation might change in the immediate future. If the opponent advances, for instance, the model must consider their trajectory, assess its own body position, and select a response—all without delay. While a standard computer program might take seconds to formulate a response, a system engaged in physical movement must react before the opponent completes their strike.

Furthermore, every action alters the situation the system has just analyzed; when the G1 lifts a leg to kick, its center of gravity shifts. If the trainer moves during that interval, the robot must reconsider its intended action. It is this cycle—observing, predicting, acting, and observing again—that Unitree aims to demonstrate with its new model; however, there is an important distinction to be made. The demonstration was released by the manufacturer itself; there is no independent evaluation yet to verify the system's performance under various conditions, and the footage does not fully clarify how much processing took place within the robot versus how much may have relied on external computers. Therefore, we are not dealing with a machine that dominates any type of combat or is capable of operating safely in any environment.

What Unitree showcased was a demonstration of autonomous planning in highly dynamic physical scenarios—ranging from human-robot interactions to robot-versus-robot combat. The G1 is not the company's only fighting model; as the footage reveals, the firm is already developing a larger, more sophisticated version: the G2. This is perhaps where the technology’s most interesting aspect lies, as learning to react to an adversary involves more than just learning to fight. A robot capable of tracking movements, adjusting its trajectory, and regaining its balance is developing capabilities that are equally valuable for navigating crowds, dodging obstacles, or handling objects in constantly changing environments.

However, there is a vast difference between facing a trainer wearing protective gear and sharing space with people who are not part of an experiment. It is outside the arena that this new generation of humanoids faces its next challenge; in September, eight Unitree robots shared the stage with human dancers on one of the United States' biggest talent shows.

Meanwhile, another Chinese company has begun introducing humanoids designed to work in hotels, retail stores, and car dealerships. The question is no longer just whether these machines can move independently, but rather: what happens when they start moving among us?