RoboBrain

An open-source series of embodied 'brain' models from BAAI in Beijing, responsible for task planning and spatial understanding.

RoboBrain is a series of open-source embodied 'brain' models from the Beijing Academy of Artificial Intelligence (BAAI). Given an image and an instruction, it outputs intermediate results such as task decomposition, affordances (where an object can be grasped or pressed), and end-effector trajectories, which are then handed to a lower-level controller for execution; it does not directly output motor commands itself. Version 1.0 was released in February 2025 and accepted at CVPR 2025, together with the ShareRobot annotated dataset; version 2.0's technical report followed in July 2025, with 7B and 32B sizes, adding capabilities such as spatial referring and long-horizon planning for multi-robot collaboration; version 2.5 began releasing 8B and 4B weights from January 2026 onward, upgrading from predicting 2D image coordinates to predicting 3D spatial trajectories with depth, and adding a general-purpose reward model that estimates task progress. RoboBrain is one of the more complete open-source 'brain' models in the domestic 'brain-cerebellum' hierarchical approach.

ExampleGiven a kitchen photo and the instruction 'put the bowl in the microwave,' RoboBrain first breaks the task into sub-steps, then marks on the image where the bowl should be grasped and the hand's movement trajectory, for a lower-level controller to execute.

Also called

RoboBrain 1.0, RoboBrain 2.0, RoboBrain 2.5, BAAI embodied brain model

Related

Embodied Foundation Model · Brain–Cerebellum Architecture · Affordance · Spatial Reasoning · RoboOS · Beijing Academy of Artificial Intelligence

Sources

arXiv 2502.21257: RoboBrain (CVPR 2025)
arXiv 2507.02029: RoboBrain 2.0 Technical Report
GitHub: FlagOpen/RoboBrain2.5

As of

2026-03