UnifoLM

Unitree's open-source series of robot foundation models, spanning a world model, a VLA, and a general-purpose humanoid model.

UnifoLM is a series of open-source robot foundation models from Unitree Robotics, with three main releases so far. UnifoLM-WMA-0 (September 2025) uses a world-model-plus-action-head architecture: the world model predicts the future images that result from the robot interacting with the environment, which can either assist an action head with decision-making or serve as an interactive simulator to generate training data; it supports the Z1 arm and the G1 humanoid. UnifoLM-VLA-0 (January 2026) is based on Qwen2.5-VL, continues pretraining on manipulation data, and uses a single policy to perform 12 categories of manipulation tasks on the G1. UnifoLM-WLA-1.0 (September 2026) is a 6-billion-parameter general-purpose humanoid foundation model that the company says was trained on about 2,500 hours of real-robot data, covering 64 tasks across tabletop and whole-body manipulation, supporting both two-finger grippers and several five-fingered dexterous hands, open-sourced under Apache 2.0. This series reflects Unitree's expansion from selling hardware into open-source models as well.

ExampleUnifoLM-WMA-0's world model can predict the upcoming images as a Unitree Z1 arm stacks boxes, which can both help choose the next action and generate synthetic training data.

Also called

UnifoLM-WMA-0, UnifoLM-VLA-0, UnifoLM-WLA-1.0

Related

Unitree Robotics · Unitree G1 · World Action Model · Vision-Language-Action Model · Unitree Z1 Robotic Arm · Open-weight Model

Sources

unitreerobotics/unifolm-world-model-action (GitHub)
unitreerobotics/unifolm-vla (GitHub)
unitreerobotics/unifolm-wla (GitHub)

As of

2026-09