Alibaba-Affiliated Team Releases Embodied Foundation Model RynnBrain 1.1, Adding Contact-Point Prediction Across Three Sizes From 2B to 122B

An Alibaba-affiliated research team (alibaba-damo-academy.github.io) released the embodied foundation model RynnBrain 1.1. The model comes in three sizes—2B, 9B and 122B-A10B—and newly adds contact-point prediction and 3D grounding.

RynnBrain 1.1 indicates that competition over foundation models powering robot intelligence is shifting from hardware to software. Its versatility, evaluated across multiple robots, could make it an industry standard and influence how robots are developed in the future.

RynnBrain 1.1, released by an Alibaba-affiliated research team, is an update to an embodied foundation model. A notable feature is its three model sizes: 2B, 9B and 122B-A10B. The introduction of contact-point prediction and 3D grounding is particularly noteworthy. Contact-point prediction forecasts where a robot should touch an object and the orientation it should use to grasp it, which is important for improving manipulation accuracy in the real world. 3D grounding also enables smaller models, including the 2B and 9B versions, to convert language instructions into coordinates in physical space. This appears intended to balance model efficiency with practical use. In previous coverage, we have reported on Unitree’s new humanoid “Superman,” which is more than 2 meters tall and Boston Dynamics’ modularization of Atlas. These developments show that the importance of software foundations is growing alongside advances in robot hardware. RynnBrain 1.1 was evaluated on several different robots—Unitree G1, Astribot-S1 and Tianji-Wuji—with an emphasis on versatility that does not depend on specific hardware. This could accelerate a shift toward using a shared foundation model rather than having each robot maker develop its own AI stack. In terms of industry structure, if foundation models such as RynnBrain 1.1 become widespread, robot intelligence could become standardized, making hardware differentiation more important. Contact-point prediction and 3D grounding, in particular, could have direct benefits for tasks such as assembly and picking with industrial robots. However, evaluations have been limited to specific robots, leaving performance on other hardware unknown. Details of RynnBrain 1.1’s training data and methods have also not been disclosed, making reproducibility and fair comparisons difficult; this remains a challenge. Key questions going forward include whether RynnBrain 1.1 will be commercialized and whether other robot makers will adopt the foundation model. In particular, whether Chinese robot makers such as Unitree and Astribot prioritize their own AI development or RynnBrain 1.1 could shape the industry’s direction. See the original source for details.