Unitree Just Open-Sourced Its Best Shot at a General-Purpose Humanoid Brain
On September 11, 2026, Unitree did something no leading humanoid maker has done this year: it fully open-sourced its top-tier embodied foundation model. The model is called UnifoLM-WLA-1.0, and per Unitree it sets new state-of-the-art results across multiple global open-source embodied-model benchmarks — while running both tabletop manipulation and whole-body mobile manipulation from a single model.
For an industry that treats its AI stacks like trade secrets, this is a genuine line in the sand. And it lands five weeks after the first humanoid pure-play IPO — so it’s worth asking what the newly public company that owns the hardware market is doing giving away the software crown.
What WLA-1.0 actually is
Strip the naming and UnifoLM-WLA-1.0 is Unitree’s bid for the “one brain, many bodies” prize. The architecture combines the company’s UnifoLM-ER-Flow multimodal backbone with an MMDiT action-expert module, integrating embodied reasoning, future dynamic-region prediction, and discrete action learning into one framework. It was trained on roughly 2,500 hours of high-quality real-robot data — including Unitree’s own previously open-sourced datasets and public resources such as BitRobot-HIW-500 — and it builds a unified action space with cross-robot-body transfer priors.
In plain terms: the model isn’t tuned for one robot or one task family. In real-robot tests Unitree says a single model completed 64 distinct tasks, spanning fine tabletop manipulation to full-body mobile manipulation, with cross-task and cross-end-effector generalization. The company frames it as “one-model-driven, whole-body coordination.”
Why this matters beyond the benchmarks
The open-source humanoid model race has been quietly heating up since January 2026, when Unitree released its UnifoLM-VLA-0 research model. Since then NVIDIA’s Isaac GR00T line has become the default for labs that can afford NVIDIA’s ecosystem, and every serious humanoid vendor has an in-house model it keeps firmly closed. Unitree’s play is different: make the software the standard, because the hardware is already yours.
If WLA-1.0 becomes the reference stack for researchers and startups, every paper, every sim-to-real pipeline, and every trained policy becomes a de facto advertisement for the machines it runs best on — the Unitree G1, the H2, and the R1 line. Openness as distribution strategy is not new in software. In humanoids, Unitree is the first with the scale to pull it off.
The open-source caveat nobody prints
There is one awkward footnote for American labs. Under the FCC’s Covered List ruling of July 28, 2026, foreign-produced advanced robotic devices are blocked from new US authorizations. The code is free; the bodies it drives are getting harder to buy stateside — with the notable exception of the grandfathered Unitree fleet (G1, R1, H2, Go2, B2, A2), which keeps its existing FCC authorizations. For US researchers, the open model plus a grandfathered robot is suddenly the only fully legal, fully open combination on the market. Unitree could not have planned the timing better.
The bottom line
WLA-1.0 is the strongest open-source embodied model release of 2026 by any maker that also ships hardware at scale. Whether the benchmark claims hold up will be settled by the community in weeks, not months — that’s the point of open-sourcing it. Either way, the strategic message is delivered: the company that already sells the most humanoids wants to own the software layer too, and it’s betting openness gets there faster than secrecy.
One week before this release, the same model family produced something stranger: a humanoid sparring autonomously with a human. We covered that separately — the world-model era is arriving faster than the spec sheets suggested.