What Unitree released
China's Unitree, one of the largest humanoid makers, opened UnifoLM-WLA-1.0, a 6-billion-parameter model it says drives the G1 humanoid through 64 tasks. Weights are on Hugging Face under Apache 2.0 (commercial use allowed), a checkpoint of about 12.5 GB, with fine-tuning code on GitHub.
Per the project README, it shipped in steps:
- September 11: weights for embodied-reasoning models UnifoLM-ER-1 and ER-Flow.
- September 20: model modules and action-expert training code.
- September 28: UnifoLM-WLA-1.0-Base and full plus LoRA fine-tuning code.
It drew attention with Robotica Guide's analysis on October 10.
How it works
- One model for hands and body: of 64 tasks, 54 are tabletop manipulation and 10 whole-body, with two-finger grippers and several five-finger hands.
- Vision-language-action (VLA) design: a vision-language model reads cameras and instructions; a diffusion-based "action expert" (MMDiT) outputs joint movements.
- The model also learns to predict what will change in the scene (optical flow), understanding the effect of its own actions.
- Trained on about 2,500 hours of real-robot data plus 5M+ reasoning samples. For comparison, Physical Intelligence's pi0 used over 10,000 hours.
What you can run
The training guide is concrete: default fine-tuning freezes the vision-language part and trains only the action head, targeting a single 24 GB GPU. A script compares predicted vs. recorded actions before going near hardware.
But the docs admit a gap: the published model server only covers the two-finger-gripper format. Whole-body control would require extending the protocol.
What's missing: numbers
Unitree has published no per-task success rate, failure counts or comparison with other open models. Videos are company-labeled as autonomous (tabletop clips at 2x speed). ER-1 has benchmarks, but they measure scene understanding, not how often the robot completes a task. It's also unclear how much training data is public.
Why it matters
Open weights from a company that has shipped thousands of humanoids shift the question from "what does Unitree say?" to "what can anyone reproduce?". With the G1 at about $13,500 in the US, independent tests are within reach of labs and startups.
For companies eyeing robotics
- Don't decide from demo videos; ask for success rates on your use case.
- Start with tabletop manipulation, where the public path is ready.
- Validate in simulation and replay before hardware, as the guide suggests.
- Review third-party licenses noted in the README before commercial use.
Sources
- GitHub, unitreerobotics/unifolm-wla: https://github.com/unitreerobotics/unifolm-wla
- Robotica Guide (2026-10-10): https://roboticaguide.com/unitree-unifolm-wla-1-0-open-source-humanoid-model/
Want to evaluate robots and physical-AI models for your operations? KGA IT offers a free consultation: kga-it.com.