Unitree UnifoLM-VLA-0

2026-01-29 · Publisher : Unitree · Domain : Robotics

Unitree releases the code and weights of UnifoLM-VLA-0, a humanoid manipulation VLA that completes 12 categories of complex tasks with a single policy, per the company.

### Background On January 29, 2026, Unitree released training and inference code plus weights for UnifoLM-VLA-0, a large vision-language-action model in its UnifoLM family, designed for generalist humanoid manipulation. ### Specifications - **Principle**: a vision-language model extended by continued pretraining on robot manipulation data, evolving from "vision-language understanding" into an "embodied brain" with physical common sense, per Unitree. - **Spatial perception**: deep integration of text instructions with 2D and 3D spatial information, improving geometric understanding. - **Generalization**: in real-robot validation, 12 categories of complex manipulation tasks completed with a single policy (results published by the company). - **Released models**: UnifoLM-VLM-Base, UnifoLM-VLA-Base (fine-tuned on Unitree's open datasets) and a variant fine-tuned on the LIBERO benchmark. - **Tooling**: builds on Hugging Face's LeRobot library. ### Why it matters Unitree complements its September 2025 world-action model with a classic VLA. The maker keeps adding open building blocks for its G1 and H1 humanoids, already widespread in labs. ### Key takeaway UnifoLM-VLA-0 paved the way for UnifoLM-ER-1 and WLA-1.0, released in September 2026.

Tags : VLA, Humanoïdes, Unitree, Poids ouverts