达摩院开源具身大脑基模RynnBrain

Core Insights - Alibaba's Damo Academy has released the RynnBrain model, a foundational model for embodied intelligence, which includes seven models in total, featuring a 30B MoE architecture [1][5] - RynnBrain introduces spatiotemporal memory and physical world reasoning capabilities, significantly enhancing robotic intelligence and setting new records in 16 embodied open-source evaluation benchmarks, surpassing top models like Google's Gemini Robotics ER 1.5 [1][3][4] Model Features - The RynnBrain model creatively integrates spatiotemporal memory and physical reasoning, essential for robots to interact with their environment. This allows robots to locate objects and predict movement trajectories, providing global spatiotemporal recall capabilities [3][7] - RynnBrain employs a training optimization architecture called RynnScale, achieving a twofold acceleration with over 20 million training pairs, resulting in comprehensive capabilities and leading performance in various tasks [3][7] Scalability and Applications - RynnBrain demonstrates excellent scalability, enabling rapid post-training for various embodied models such as navigation and planning, with minimal data requirements for fine-tuning [4][8] - The model's architecture allows it to outperform a 72B model with only 3B activation parameters, enhancing the speed and fluidity of robotic actions [5][8] Evaluation and Industry Impact - Damo Academy has also introduced a new evaluation benchmark, RynnBrain-Bench, to assess spatiotemporal fine-grained embodied tasks, filling a gap in the industry [5][8] - The head of the embodied intelligence lab at Damo Academy stated that RynnBrain marks a significant step towards achieving deep understanding and reliable planning of the physical world, accelerating the transition of AI from the digital realm to real-world applications [5][8]