UnifoLM-ER-1

UnifoLM-ER-1 is a vision-language-action foundation model released by Unitree Robotics on September 11, 2026, as part of the UnifoLM-WLA-1.0 open-source plan for whole-body locomotion and manipulation. The weights are published on the official unitreerobotics Hugging Face organization under the Apache 2.0 license. The model is a 4B-parameter network built on the Qwen3-VL-4B backbone (Qwen3VLForConditionalGeneration architecture). It is trained on more than 5 million multimodal robot samples spanning perception, language grounding, and action generation, and coordinates desktop manipulation and whole-body manipulation tasks on Unitree humanoid platforms. UnifoLM-ER-1 sits at the core of the UnifoLM-WLA ecosystem alongside UnifoLM-ER-Flow, an interaction-centric world-model extension released the same day. Together they form Unitree's open general-purpose humanoid foundation model stack, complemented by the UniBot-V1 Challenge Dataset and forthcoming post-training code releases. Performance figures reported by Unitree are company-reported and have not been independently verified.