How RynnBrain Unifies Perception, Reasoning, and Planning for Embodied AI

RynnBrain is an open‑source unified spatiotemporal foundation model released by Alibaba DAMO Academy, designed to power embodied intelligence by integrating perception, localization, physical reasoning and planning into a single brain.

Model Architecture

The model accepts full‑range visual inputs—including single‑view images, multi‑view images and video—and combines them with natural‑language commands. A shared dense or mixture‑of‑experts decoder produces aligned multimodal outputs such as text, region proposals, trajectories and pointing signals, enabling a unified output space for self‑centered understanding, spatiotemporal localization, physics‑based reasoning and fine‑grained action planning.

Model Variants

RynnBrain is offered in three scales—2 B, 8 B and a 30 B mixture‑of‑experts (MoE)—as well as task‑specific variants (Nav, Plan, VLA, CoP) that are fine‑tuned for navigation, planning, visual‑language‑action and collaborative‑policy tasks.

Evaluation Benchmarks

Extensive testing on more than 20 embodied benchmarks shows that RynnBrain consistently outperforms existing models across perception, reasoning and planning metrics.

https://alibaba-damo-academy.github.io/RynnBrain.github.io/
https://arxiv.org/pdf/2602.14979

Overall, RynnBrain serves as both a physical‑world reasoning engine and a pre‑trained brain that can be efficiently adapted to a wide range of robotic tasks, demonstrating the feasibility of a “unified brain” for real‑time, dynamic environments.