Unitree UnifoLM-ER-1 and UnifoLM-ER-Flow

2026-09-11 · Publisher : Unitree · Domain : Robotics

Unitree releases under Apache 2.0 UnifoLM-ER-1, a 4-billion-parameter embodied reasoning model built on Qwen3-VL-4B, and UnifoLM-ER-Flow, which adds interaction-centric world modeling.

### Background On September 11, 2026, Unitree released on Hugging Face the weights of two embodied reasoning (ER) models under the Apache 2.0 license: UnifoLM-ER-1 and UnifoLM-ER-Flow. They form the perception foundation of its upcoming humanoid foundation model, UnifoLM-WLA-1.0. ### Specifications - **UnifoLM-ER-1**: about 4 billion parameters, built on Qwen3-VL-4B, trained on 5M+ samples (image pointing, object detection, multi-image reasoning, 2D trajectory prediction, 3D detection, spatial QA), mixed with general image-text data. - **Results**: across 16 multimodal perception and understanding benchmarks, ER-1 leads open models on 7 and approaches leading proprietary models (scores published by the company). - **UnifoLM-ER-Flow**: extends ER-1 with interaction-centric world modeling and discrete action learning. Using optical flow, it predicts as tokens the scene regions that will change, conditioned on the image and instruction. - **Demos**: table cleaning, towel folding, phone installation. - **License**: Apache 2.0, unlike the non-commercial license of the first UnifoLM model. ### Why it matters By building on an open Alibaba VLM and releasing under a permissive license, Unitree makes its building blocks commercially usable. Predicting change regions is a frugal way to bring a world model into a VLM. ### Key takeaway ER-1 and ER-Flow paved the way for UnifoLM-WLA-1.0, released seventeen days later.

Tags : Raisonnement incarné, Unitree, Open Source, VLM