Unitree Robotics Open-Sources UnifoLM-WLA-1.0 Humanoid Robot Foundation Model
Unitree Robotics has fully open-sourced UnifoLM-WLA-1.0, a next-generation foundation model for general-purpose humanoid robots. The single unified model sets new open-source SOTA records across multiple embodied reasoning benchmarks and manages 64 tabletop and mobile manipulation tasks. Fully open-sourcing a competitive embodied foundation model lowers the barrier for global researchers to build versatile humanoid robots. By unifying mobile navigation, spatial reasoning, and manipulation across different end-effectors, UnifoLM-WLA-1.0 accelerates progress toward general physical AI. The model incorporates interaction-centric world modeling combined with large-scale multimodal perception data to enhance spatial understanding and embodied reasoning, rivaling top proprietary models. Real-robot evaluations demonstrated generalized performance across diverse tasks and hardware configurations, including varied end-effectors.
## BACKGROUND
Embodied AI focuses on intelligent agents, such as humanoid robots, that interact with the physical world using sensory inputs and motor outputs. Historically, robotics relied on separate, task-specific controllers for navigation, object recognition, and robotic arm manipulation. Foundation models in robotics aim to consolidate these functions into a single neural network capable of generalizing to novel environments and physical bodies.