February 10 News – Alibaba (阿里巴巴) DAMO Academy released RynnBrain, a foundation model for embodied AI "brains," and simultaneously open-sourced seven models from the full series, including a 30B MoE version.
The model introduces spatio-temporal memory and physical spatial reasoning to robotics for the first time, enabling robots to record and backtrack historical states during task execution and restore previous work scenes after interruptions. For example, a robot running RynnBrain that is suddenly interrupted while performing Task A to handle Task B can remember the temporal and spatial status of Task A and resume work after completing Task B.
According to DAMO Academy, in 16 open-source embodied AI benchmarks, the model achieved leading results across multiple tasks such as environmental perception, first-person visual question answering (VQA), and spatial reasoning. Its performance exceeds existing representative models like Google (GOOG) Gemini Robotics ER1.5 and NVIDIA (NVDA) Cosmos Reason2.
Zhao Deli, head of the Embodied Intelligence Lab at DAMO Academy, stated that RynnBrain achieves a deep understanding and reliable planning of the physical world by the "brain," marking a key step in the development of general embodied intelligence. Previously, DAMO Academy has successively open-sourced embodied models such as WorldVLA and RynnEC, and introduced the robot context protocol RynnRCP, continuing to advance the exploration of AI applications in real physical scenarios.