On January 29, 2026, Unitree Technology made a groundbreaking announcement: the open-source release of the UnifoLM-VLA-0 large-scale model. This model belongs to the UnifoLM series and is a visual-language-action (VLA) large model, specifically crafted for the versatile operation of humanoid robots. Its primary objective is to break free from the constraints that traditional visual-language models (VLMs) face when it comes to physical interactions.

By undergoing continuous pre-training on a wealth of robot operation data, UnifoLM-VLA-0 has successfully transformed from a model focused on general "image-text understanding" to an "embodied brain" that is equipped with a solid grasp of physical common sense. This evolution has endowed the model with markedly improved spatial reasoning abilities. Moreover, it showcases dependable multimodal perception performance across a diverse array of task scenarios. As a result, it can proficiently complete 12 types of intricate operational tasks with high quality.

In real-world robot validation tests, a single strategy employed by the model is sufficient to carry out these tasks. It can maintain stable execution and demonstrate strong resistance to interference, even when subjected to external perturbations.