UnifoLM-WMA-0

UnifoLM-WMA-0, published by Unitree on 15 September 2025, is the company's first open-sourced world-model-action architecture and the first entry in its Uniform Large Models line. Unitree is known for hardware — nineteen of its robots are catalogued here — and this is the software counterpart: the layer that decides what a robot does, released rather than kept in-house. The design rests on a world model, a network trained to predict how a scene changes when a robot acts on it. Unitree uses that prediction in two distinct ways. Run forward on its own, the model works as an interactive simulator that manufactures synthetic training episodes — valuable because collecting real robot data is the expensive bottleneck in this field. Attached to an action head, the same prediction becomes policy enhancement: the robot imagines the next few moments of an interaction before committing to a movement, which measurably improves its decisions. The architecture is deliberately cross-embodiment, demonstrated on the Z1 robotic arm, the G1 humanoid and dual-arm setups, with training code, inference code and weights on GitHub and Hugging Face. One important caveat: the licence is CC BY-NC-SA 4.0, so despite being open weights this is not free for commercial deployment — a narrower grant than the Apache-2.0 releases common elsewhere in the field. Unitree has since demonstrated successors, including the WVLA 2.0 world model and UnifoLM-OminiA-0.3 for home-care tasks; those are separate releases and are not covered by this profile.

▸Videos

No videos yet.

▸Reviews