Alibaba DAMO Academy Releases Open-Source Embodied AI Model RynnBrain
Alibaba, through its research institute DAMO Academy, has released the open-source embodied foundation model RynnBrain. The model aims to provide robots with physical reality grounding capabilities, enabling them to shift from passive observation to actively performing complex real-world tasks. RynnBrain is based on Alibaba’s Qwen3-VL vision-language model and is now available for developers on GitHub and the Hugging Face platform.
Core Model Functions
RynnBrain focuses on physical-aware reasoning, capable of processing egocentric scene observations, grounding language in physical space and time, and planning for real-world tasks. The model features spatial understanding, episodic memory, and object recognition capabilities, supporting skills like counting, spatial awareness, and motion trajectory prediction. Alibaba states that the model achieves active physical-aware reasoning through systematic upgrades.
Demonstration Task Showcase
In the released demo video “RynnBrain’s Housework Diary,” a robot equipped with the model demonstrated several household management abilities. For instance, the robot can organize tableware around a sink based on commands, identify and select three oranges from a variety of fruits to place in a bowl, retrieve a milk bottle from the refrigerator, and tidy up items in a cluttered front room. These tasks highlight the model’s practical application in object recognition, target localization, and motion planning.
Performance Benchmark Comparison
RynnBrain has shown excellent performance in multiple benchmarks, comparable to leading embodied models like Google DeepMind’s Robotics-ER 1.5 and Nvidia’s Cosmos Reason2. The model is offered in several variants, including 2B and 8B parameter versions, as well as a Mixture-of-Experts variant, to suit different computational needs. Alibaba has also released the RynnBrain-Bench evaluation suite on GitHub for comprehensive testing of embodied models’ cognitive and localization capabilities.
Open-Source Release Details
The RynnBrain series of models is fully open-sourced. Developers can access them in the GitHub repository alibaba-damo-academy/RynnBrain and the Alibaba-DAMO-Academy collection on Hugging Face. Related resources include model weights, evaluation benchmarks, and application cookbooks, supporting functions like spatial understanding and grasping pose prediction. The open-sourcing of this model promotes development and research in the field of embodied intelligence.
