RynnUnify: Towards Unified Embodied Brain

RynnUnify. Towards Unified Embodied Brain. By Minghao Zhu and twenty-two others, from DAMO Academy at Alibaba Group and Hupan Lab. Published on the twenty-third of October. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

An embodied system has to do more than understand what it sees. It has to anticipate how the physical world will evolve across a run of successive subtasks, and then act in that world. Long tasks involve states that have not yet been observed, and interactions whose outcomes are still to come, and a model has to reason about both. Understanding, visual planning and robot control are three capabilities, and the question this work takes up is whether one model can carry all three.

RynnUnify is a sixteen billion parameter unified embodied model. It combines strong embodied understanding with long-horizon visual planning, through reasoning in which video and language are interleaved. Its capabilities span spatial reasoning, physical prediction and grounding, and it connects an understanding of the physical world with visual anticipation of how that world evolves from one subtask to the next. Planning, in this design, is visual: the model anticipates how the scene will look as each subtask is carried out.

Two extensions build on that base. An agentic extension is developed through reinforcement learning to selectively use video predictions, reasoning about states it has not observed and about the outcomes of interactions. The word selectively carries weight: the extension learns when to draw on a video prediction, rather than producing one at every step. A world-action extension of the full model translates its understanding and prediction capabilities into robot control. To evaluate long-horizon visual planning and anticipatory reasoning, the authors introduce two benchmarks, Rynn-VP and Rynn-GEU.

Among the unified multimodal models evaluated, RynnUnify achieves the highest average embodied understanding score. On Rynn-VP it improves the overall score by nine point seven five points over Cosmos 3 Nano. Its agentic extension improves the Rynn-GEU average by six points over direct reasoning. Rynn-VP and Rynn-GEU are the two benchmarks introduced alongside the model, so these two gains are measured on the authors' own evaluations.

The world-action extension achieves fifty-one point five percent success on twenty selected composite RoboCasa365 tasks. Across three real-robot tasks, its average success is forty-six point seven percent.

Fifty-one point five percent on RoboCasa365 is the figure that connects the model to action, but the case RynnUnify makes is breadth. One sixteen billion parameter model serves as the shared foundation for embodied understanding, for visual planning, and for robot control. Its authors present it as a step toward a unified embodied brain, rather than as the arrival of one.