Unitree open-sources UnifoLM-WLA and releases the UnifoLM-ER models
SUMMARY
Unitree published the UnifoLM-WLA repository on GitHub on 11 September 2026, with the weights of two models, UnifoLM-ER-1 and UnifoLM-ER-Flow, on Hugging Face the same day; the model modules and code for training action experts followed on 20 September. Unitree describes UnifoLM-WLA-1.0 as a 6-billion-parameter humanoid foundation model, trained on about 2,500 hours of real-robot data, that coordinates 64 tasks across tabletop and whole-body manipulation with two-finger grippers and five-finger hands. UnifoLM-ER-1, built on Qwen3-VL-4B, targets spatial understanding and reasoning in embodied settings. The code is released under the Apache License 2.0.
Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.
WHY IT MATTERS
EDITORIALThe maker publishes weights and the code to train action experts on top of them, so developers can fine-tune on their own robot data instead of only reading benchmark claims; the repository also specifies the action and state format its models expect.