达摩院开源具身大脑基模RynnBrain，登顶16项榜单，超越Gemini

Core Insights - Alibaba's Damo Academy has released the RynnBrain model, which significantly enhances robotic intelligence by introducing spatiotemporal memory and spatial reasoning capabilities, surpassing leading models like Google's Gemini Robotics ER 1.5 in 16 benchmarks [1][4][6] Model Features - RynnBrain incorporates spatiotemporal memory, allowing robots to locate objects and predict movement trajectories based on their historical memory, thus providing global temporal recall capabilities [2][7] - The model employs a novel reasoning strategy that intertwines text and spatial positioning, reducing hallucination issues and ensuring that reasoning is grounded in the physical environment [2][7] Performance Metrics - RynnBrain has achieved state-of-the-art (SOTA) results across 16 embodied evaluation benchmarks, including environmental perception, object reasoning, and trajectory prediction, outperforming models from Google and NVIDIA [4][6] - The model was trained on over 20 million pairs of data using the proprietary RynnScale architecture, which accelerates training by two times with the same resources [4][7] Scalability and Applications - RynnBrain demonstrates excellent scalability, enabling rapid fine-tuning for various embodied models such as navigation and planning with minimal data, achieving superior performance compared to Gemini 3 Pro [7] - The model is open-sourced, including a complete set of reasoning and training codes, and introduces the RynnBrain-Bench for evaluating fine-grained spatiotemporal tasks, filling a gap in the industry [7][9] Industry Impact - The introduction of RynnBrain marks a significant advancement in embodied intelligence, facilitating the transition of AI from digital environments to real-world applications [9]