Unitree on Sept. 7, 2026, said its UnifoLM-X2-1.0 world-action model drove a G1 humanoid through sparring against a padded human trainer with no teleoperation and no scripted choreography. The claim matters because closed-loop world-model control under high-dynamic contact is rarely shown as a company primary; Unitree is arguing that real-time prediction and planning can replace the remote pilot that earlier humanoid fight demos usually needed.
Fully autonomous here is Unitree's label for removing a visible teleoperator and preset routine from the bout. It does not, on the published record, prove that all inference ran only on the G1. Interesting Engineering noted console logs such as OBS/replan and env.step in the footage, which can indicate sensor data going to a separate computer that returns motion commands. UnifoLM-X2-1.0 is also not the open UnifoLM-WMA-0 code-and-weights release already covered on the archive; Humanoids Daily maps X2 to the sparring demo and says it has not verified an official X2 code-and-weights drop.
Unitree framed the clip as the first real-time world-model-driven fully autonomous humanoid robot combat and said UnifoLM-X2-1.0 breaks bottlenecks in instant planning, decision-making, and dynamic interactive execution. The company said the demonstration validates the fundamental feasibility of large-scale deployment of world model-driven humanoid robots. Official surfaces for that claim include Unitree's LinkedIn post and the Unitree Robotics YouTube video of the same bout.
Independent coverage restated the company framing with harder edges. Interesting Engineering described predictive overlays in the clip and the offboard-compute caution above. AlphaSignal said no paper, latency figures, or benchmarks were released and that open-sourcing of X2-1.0 remains unconfirmed, while noting the lineage from open UnifoLM-WMA-0. Humanoids Daily treated X2 as a fast-decision branch of the UniFoLM family and separated it from WMA-0 and VLA-0 public releases.
What remains open is whether outside labs will instrument latency and success rates, whether the stack runs fully onboard under contact, and whether the same predict-then-act loop transfers beyond a controlled sparring partner. Company video claims and third-party caveats should be weighed together until a paper, benchmarks, or a verified X2 release lands.