Unitree UnifoLM-WLA-1.0

2026-09-28 · Publisher : Unitree · Domain : Robotics

Unitree releases UnifoLM-WLA-1.0-Base, a 6-billion-parameter humanoid foundation model trained on about 2,500 hours of real robot data, coordinating 64 tasks with one model, under Apache 2.0.

### Background On September 28, 2026, Unitree released UnifoLM-WLA-1.0-Base and its fine-tuning code on Hugging Face. The Chinese maker presents this 6-billion-parameter model as the fully upgraded next generation of its foundation model for general-purpose humanoid robots. ### Specifications - **Size**: 6 billion parameters. - **Foundations**: large-scale multimodal perception and understanding data plus interaction-centric world modeling, inherited from the UnifoLM-ER-1 and ER-Flow models released on September 11. - **Robot data**: about 2,500 hours of high-quality real-robot data. - **Scope**: a single model coordinates 64 tasks, from tabletop to whole-body manipulation. - **Grasping**: supports two-finger grippers and several five-finger dexterous hands, with strong cross-task and cross-effector generalization per Unitree. - **Reasoning**: Unitree reports leading results on several embodied reasoning benchmarks (scores published by the company). - **Openness**: weights and fine-tuning code (including LoRA) under Apache 2.0, with several datasets released. ### Why it matters Unitree delivers a complete, open, commercially usable humanoid foundation model, while Figure or 1X keep theirs closed. For labs and integrators using Unitree robots, it is a ready-to-tune base. ### Key takeaway In one year, Unitree went from a non-commercial world-action model to an open 6B humanoid foundation model.

Tags : Modèle de fondation, Humanoïdes, Unitree, Open Source