UnifoLM-X2-1.0 Sends Unitree’s G1 Into a Fight With No One on the Sticks

Hangzhou’s Unitree Robotics posted a clip last week that looks familiar until you watch who is missing. A G1 humanoid in red boxing gloves squares off against a padded human trainer inside a ring. Stance changes, slips, straight punches, and body kicks arrive on their own. Unitree calls the software UnifoLM-X2-1.0 and labels the demo the first real-time world-model-driven fully autonomous humanoid combat. No VR headset operator appears, and no gamepad rests just off frame.

Standing 1.32 meters tall and weighing approximately 35 kilos, Base G1 is powered by 23 joints and a 90 Newton-meter knee motor, which is sufficient to keep the thing upright even after being hit hard. When the trainer gets too close, you can see it resetting its feet, bending out of the direction of incoming blows, throwing a counter punch or two, and even balancing on one leg to launch a strike. After collision, it wobbles briefly before regaining balance and continuing to monitor the opponent. The overlay visuals indicate the projected opponent path a split second before the motors move.

UnifoLM-X2-1.0 simulates the physical situation ahead of time, anticipating what will happen. It uses the present poses of both bodies to forecast what will happen next, such as where a punch will land, how the g1’s mass will shift if it plants a foot, and so on. These forecasts are subsequently turned into movement commands in this tight cycle of watch – forecast – choose – move. This entire stack is built on top of previous UnifoLM work, which began with factory manipulation and ultimately included the open-source world-action stuff dubbed WMA-0. The robot was put through a severe test by being placed in a fighting setting, where hits come quickly, contact is dirty, and a biped must be able to stay on its feet while its opponent changes its tactics.

Previous humanoid fights required a human on the opposite end of the system, and replacing that person was a significant challenge. In May 2025, G1 units boxed live on TV in Hangzhou via a remote control. However, further showcase bouts in China and the United States relied on VR setups and gamepads due to the raw delay between camera and motor, leaving the robot playing catch-up to a moving opponent. Unitree’s new clip is the first public test to break that tether, albeit in a contact-heavy environment.

The text viewable on the console reveals what is really going on behind the scenes, including lines like PolicyServer OBS/replan & env.step, which show what data is being passed from the robot to the computer performing the heavy prediction work, and then back to the robot. Unitree has not disclosed any latency numbers or a technical report, and no outside lab has been able to replicate the test either. The trainer has been really helpful by just giving the robot with an easy target and not attempting to break it, though the company does advise that people keep a couple of meters between themselves and the robot.

The robot is getting to the point where it can pick when and where to punch and step. If the same prediction loop can handle the chaos of a real-world contact battle while remaining balanced, Unitree believes it might enable humanoids blend into a factory floor where people are present. At the present, the move set is rather simple, and the heavy math is likely to remain on the server, but the next task is to transfer that same level of foresight to the robot’s own processors so it can continue fighting after being cut away from the connection.