RynnUnify: Towards Unified Embodied Brain
RynnUnify is a 16B parameter unified embodied model designed to integrate embodied understanding with long-horizon visual planning through interleaved video-language reasoning. By incorporating reinforcement learning to predict unobserved states and a world-action mechanism for direct robot control, the model bridges physical reasoning with motor execution. Benchmarked against Rynn-VP and Rynn-GEU, it achieved a 51.5% success rate on composite RoboCasa365 tasks and outperformed existing models like Cosmos 3 Nano, positioning itself as a foundational architecture for integrated robotic agency.
Why it's worth reading
This model introduces a consolidated architecture for long-horizon visual planning, physical prediction, and motor control, providing a concrete performance baseline for future embodied AI development.