Embodied intelligence—the capacity for AI to interact physically with the world—has long faced a significant "brain-body" gap. Most AI models reason through text or static images but lack a fundamental sense of physical space and time. To bridge this, DAMO Academy has released RynnBrain, an open-source spatiotemporal foundation model designed to serve as a unified cognitive core for real-world agents. By integrating perception, reasoning, and planning into a single architecture, RynnBrain allows robots to not just see, but to understand the physical consequences of their actions within a 3D environment.

The RynnBrain family includes three scales, highlighted by a 30B parameter Mixture-of-Experts (MoE) variant that activates only specific sub-networks to handle tasks efficiently. Unlike standard vision models, RynnBrain is physically grounded, meaning it connects linguistic concepts directly to spatial coordinates and temporal sequences. This grounding allows agents to remember object locations over time and generate executable plans that account for affordances—the specific ways an object can be interacted with, such as a handle being for pulling rather than pushing.

To support the research community, the developers also introduced RynnBrain-Bench to test long-horizon tasks requiring episodic memory, or the ability to recall specific past events within a sequence. This release marks a significant move toward Generalization in robotics, providing tools for agents to move beyond pre-programmed routines toward adaptable autonomy. By open-sourcing the models and the training framework, RynnBrain offers a powerful, transparent alternative to proprietary robotics systems.

Embodied intelligence—the capacity for AI to interact physically with the world—has long faced a significant "brain-body" gap. Most AI models reason through text or static images but lack a fundamental sense of physical space and time. To bridge this, DAMO Academy has released RynnBrain, an open-source spatiotemporal foundation model designed to serve as a unified cognitive core for real-world agents. By integrating perception, reasoning, and planning into a single architecture, RynnBrain allows robots to not just see, but to understand the physical consequences of their actions within a 3D environment.

The RynnBrain family includes three scales, highlighted by a 30B parameter Mixture-of-Experts (MoE) variant that activates only specific sub-networks to handle tasks efficiently. Unlike standard vision models, RynnBrain is physically grounded, meaning it connects linguistic concepts directly to spatial coordinates and temporal sequences. This grounding allows agents to remember object locations over time and generate executable plans that account for affordances—the specific ways an object can be interacted with, such as a handle being for pulling rather than pushing.

To support the research community, the developers also introduced RynnBrain-Bench to test long-horizon tasks requiring episodic memory, or the ability to recall specific past events within a sequence. This release marks a significant move toward Generalization in robotics, providing tools for agents to move beyond pre-programmed routines toward adaptable autonomy. By open-sourcing the models and the training framework, RynnBrain offers a powerful, transparent alternative to proprietary robotics systems.