RynnBrain - Alibaba DAMO Academy's Open-Source Embodied Intelligence Brain Foundation Model

What is RynnBrain

RynnBrain is an open-source embodied intelligence brain foundation model launched by Alibaba DAMO Academy, giving robots spatiotemporal memory and physical spatial reasoning capabilities for the first time. The model is based onQwen3-VLtrained with the self-developed RynnScale architecture, with over 20 million pairs of training data. The model set new records (SOTA) in 16 embodied open-source benchmarks, surpassing top models such as Google Gemini Robotics ER 1.5. Alibaba open-sourced a full series of 7 models at once, including the industry's first 30B MoE architecture embodied model, which requires only 3B activated parameters to achieve efficient inference, making robot actions faster and smoother, providing "brain"-level support for complex mobile manipulation and long-horizon task planning.

Main Features of RynnBrain

• Spatiotemporal Memory: The robot can locate objects and target areas within complete historical memory, predict motion trajectories, and achieve global spatiotemporal backtracking capability.
• Physical Spatial Reasoning: The model adopts a reasoning strategy interleaving text and spatial localization, ensuring the reasoning process is closely grounded in the physical environment and greatly reducing hallucination issues.
• Task Continuity: After a robot is interrupted while executing task A to switch to task B, it can remember the temporal and spatial state of task A and automatically resume it upon completion.
• Multi-dimensional Perception: The model supports 16 embodied capabilities including environment perception, object reasoning, first-person visual question answering, spatial reasoning, and trajectory prediction.
• Rapid Expansion: Based on the foundation model, fine-tuning with only a few hundred data samples can post-train various embodied specialized models such as navigation, planning, and action.

Technical Principles of RynnBrain

• Multimodal Fusion Architecture: RynnBrain adopts an end-to-end multimodal architecture that uniformly encodes first-person video, language instructions, and physical environment information, achieves vision-language alignment through cross-modal attention, and incorporates a physical constraint module so that the reasoning process balances statistical patterns with physical laws such as object affordance and geometric relations, ultimately outputting executable action sequences.
• Egocentric Cognition: For dynamic first-person viewpoints, the model uses a video Transformer to process long temporal sequences, establishes cross-frame consistency, and constructs queryable spatial memory through 3D scene graphs or neural radiance fields, supporting backtracking reasoning over historical observations and accurately answering questions involving spatiotemporal relations.
• Spatiotemporal Localization: RynnBrain encodes historical videos into episodic memory vectors, supports natural language-based temporal retrieval and keyframe localization, combines visual grounding techniques to map text to image regions, outputs 3D positions through spatial coordinate regression, and can predict motion trajectories to support predictive planning.
• Physics-Aware Planning: Unlike VLA models that directly predict actions, RynnBrain explicitly introduces a physical world model: extracting object affordances, verifying action feasibility, and using a hierarchical planner to decompose high-level intentions into low-level actions with conditional branches, ensuring the success rate of complex long-horizon tasks.
• Model Variants: Provides two types of architectures: Dense (2B/8B) and MoE (30B-A3B): Dense with full parameter activation is suitable for real-time edge deployment; MoE with sparse activation balances large capacity and low inference cost, and provides specialized optimized versions for planning, navigation, and spatial reasoning tasks.

RynnBrain Project Address

• Project Website：https://alibaba-damo-academy.github.io/RynnBrain.github.io/
• GitHub Repository：https://github.com/alibaba-damo-academy/RynnBrain
• HuggingFace Model Library：https://huggingface.co/collections/Alibaba-DAMO-Academy/rynnbrain

Application scenarios of RynnBrain

• Industrial manufacturing: The model supports multi-step operations on complex assembly lines, and can automatically resume tasks after interruptions, improving production line flexibility.
• Home services: An intelligent butler can remember the placement positions of items, complete cross-room retrieval commands, and predict people's trajectories to proactively avoid them.
• Medical care: Assistive robots remember the positions and statuses of patients, can resume care procedures after emergency interruptions, and precisely deliver instruments.
• Logistics and warehousing: Forklifts and AGVs update environmental memory in real time among dynamic shelves, and predict motion trajectories to avoid collisions.
• Scientific research and exploration: Field survey robots record spatiotemporal information of sampling points, supporting long-term scientific expedition tasks and disaster rescue searches.