UnifoLM-WMA-0

0 (Base / Dual)

Unitree's open world-model-action framework: a video-generation world model serving as both simulator and policy enhancer for robotics.

🔬 Research🔬 Research only⚖ Open weightsWorld ModelRobotics foundation modelVision-Language-Action model

Release date

15 September 2025

🏢Unitree RoboticsProducer

Access:DownloadDeployment:💻 Local☁ Cloud

Overview

UnifoLM-WMA-0 is an open (CC BY-NC-SA 4.0) World-Model-Action framework from Unitree Robotics that couples a generative video world model with an action head. Trained on large-scale embodied datasets (incl. Open-X-Embodiment), it operates in two modes: (1) Simulation Engine — predicting future video frames conditioned on robot actions, providing a cheap, photorealistic sim for policy training and evaluation; (2) Policy Enhancement — using the internal world model to 'imagine' the outcomes of actions and improve control policy quality. Released variants include UnifoLM-WMA-0-Base and UnifoLM-WMA-0-Dual.

Classification

World ModelRobotics foundation modelVision-Language-Action model

Applications

Simulation / synthetic data generation 3D world generation Game / VFX environment design Robot policy training Robotic manipulation

Access & deployment

Download

LocalCloud

Weights: Open weights

Key parameters

✓ Fine-tuning

📥 Input: video, image, robot state data, robot sensors

Robotics