Unitree open-sources UnifoLM-WLA-1.0: one model for 64 tasks, and household tasks are among them
Unitree has released UnifoLM-WLA-1.0 as open source: 6 billion parameters, about 2,500 hours of data collected on real robots, one model for 64 tasks. The declared tasks include laundry, bed and trash. But the company's founder places the real leap 2-10 years from now.
⚠️ How to read this piece. The technical figures below have been verified on the model's official repository. The performance and the list of tasks remain claims by the company, not verified by independent third parties: where that is the case, we say so line by line.
In brief
Unitree has made public the weights and training recipes of UnifoLM-WLA-1.0, a model that translates what the robot sees and what it is asked into movements of the body. (source: official repository unitreerobotics/unifolm-wla)
What the repository declares, and what we have read directly:
- 6 billion parameters;
- trained on about 2,500 hours of high-quality data collected on real robots — not in simulation;
- a single model coordinates 64 tasks, spanning tabletop manipulation and whole-body manipulation;
- it works both with the two-finger gripper and with multiple different five-finger hands. This is the most technically notable point: as a rule, changing the robot's hand throws the model into crisis and forces data to be collected again.
The household tasks
According to the trade press that broke the news, the 64 tasks include taking out the trash, putting clothes in the washing machine, making the bed and folding towels, with a split of ten whole-body tasks and fifty-four tabletop tasks. (source: Humanoids Daily, September 10, 2026 — this split and the list of tasks do not appear in the official description of the repository, which only declares the 64 tasks overall)
Our reading. This is the first time Unitree has shown its G1 on a list like this instead of in dances or fighting demonstrations. For a site that looks at the home, the difference is not one of degree but of kind: folding towels is a household task, a backflip is not.
The counterpoint comes from Unitree itself
The founder and chief executive Wang Xingxing has stated that humanoids are not ready to scale, not even in factories: physical contact accumulates millimeter errors, and success rates collapse as soon as the environment departs from that of the training data. The real leap, in his view, will come when a humanoid enters a room it has never seen and completes 80% of the tasks asked of it by voice — a goal he places 2-10 years from now. (statement reported by Humanoids Daily; we have not verified the exact date and venue of the remarks)
It is worth noting where this caution comes from: not from an outside critic, but from the company that was announcing the model that day.
Why it matters to us
It is open source, and that lowers the barrier. Anyone in Italy doing research or integration can download the model and try it without depending on a closed service. (To be verified before drawing practical conclusions: exactly which license, and whether it allows commercial use.)
The bottleneck is the software, and the people selling the hardware admit it. A company that gives away the algorithmic layer is saying where it thinks the value lies: in the body and in the volumes, not in the code. This is our reading, not a statement by the company.
None of these tasks is a product yet. There is no home G1 that can be bought in Italy, nor a price, nor a date. What there is is a downloadable model and a list of things that, according to those who trained it, it can do.
Sources
- Unitree Robotics (official repository) primary source
- Humanoids Daily press
- Humanoids Daily press
Next to each source we state what kind of source it is: “primary source” means the manufacturer, an official document or the author of the data; “press” a news report; “retailer” a distributor or a store, reliable about its own offer — price and availability in Italy — but not about the product’s features; “analysis” a professional firm or a research center that signs its own reading; “Chinese state media” an outlet controlled by the state, whose content should be weighed as such; “blog/forum” a non-editorial source; “unclassified” a source we have not assessed yet: the link stays, the judgment does not.