Unitree Robotics released UnifoLM-ER-Flow on Hugging Face
Unitree Robotics released UnifoLM-ER-Flow on Hugging Face under Apache-2.0. The model extends UnifoLM-ER-1 with interaction-centric world modeling and discrete action learning, jointly aligning visual observations, language conditions, predicted future dynamic regions, and robot actions in a single vision-language model.
China context
- Original name
- 宇树科技
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download and use UnifoLM-ER-Flow under Apache-2.0, enabling integration into robot learning pipelines without licensing restrictions.
- For investors
- Unitree Robotics' open-source release of UnifoLM-ER-Flow signals its commitment to building an ecosystem around embodied AI models, which can influence competitive dynamics in the robotics software market.
Unitree Robotics released UnifoLM-ER-Flow on Hugging Face under Apache-2.0. The model extends UnifoLM-ER-1 with interaction-centric world modeling and discrete action learning. It uses optical flow to extract dynamic regions, trains a VQ-VAE to encode them into discrete tokens, and predicts mask tokens for future dynamic regions conditioned on the current image and task description or action. The action space is partitioned into end-effector poses, end-effector joints, and lower-body joints, each encoded with a separate residual vector quantization model. The resulting token sequences share timesteps and are fed synchronously into the VLM, jointly aligning visual, language, and action representations.
UnifoLM-ER-Flow introduces discrete action tokens and mask tokens for future dynamic regions, enabling a single VLM to predict interaction-induced scene changes and robot actions. The action encoding pipeline uses separate RVQ models for end-effector poses, end-effector joints, and lower-body joints, with synchronized token sequences fed into the VLM. This design extends UnifoLM-ER-1, which was built on Qwen3-VL-4B and trained on over 5 million samples across multimodal perception and understanding benchmarks.
Robot developers outside China can now access an Apache-2.0 licensed model that unifies visual, language, and action representations for embodied AI, reducing the cost and complexity of building interaction-aware robot systems. Unitree Robotics' release of UnifoLM-ER-Flow on Hugging Face makes its interaction-centric world modeling approach available for commercial use without licensing fees.
The Apache-2.0 license allows commercial use, modification, and distribution, enabling robotics companies and researchers to integrate UnifoLM-ER-Flow into their products without licensing costs. The model's focus on interaction-centric world modeling and discrete action learning addresses a key challenge in robot manipulation and can accelerate development of more capable embodied AI systems.
Observable next signals include whether Unitree Robotics releases additional models in the UnifoLM-ER series, whether the model is adopted in downstream robotics applications, and whether independent benchmarks validate its performance on interaction-centric tasks. The release of UnifoLM-ER-1 on the same day suggests a coordinated open-source push by Unitree Robotics.