阿里发布RynnBrain大模型，让机器人具备“思考大脑”！

Core Insights - Alibaba's Damo Academy has released the RynnBrain foundational model for embodied intelligence, which includes seven models, notably the 30B MoE model, all of which are open-sourced [1][6] - RynnBrain significantly enhances robots' capabilities by introducing spatiotemporal memory and physical world reasoning, achieving state-of-the-art (SOTA) results in 16 embodied open-source evaluation benchmarks, surpassing leading models like Google's Gemini Robotics ER 1.5 [1][4] Model Capabilities - RynnBrain enables robots to locate objects and predict motion trajectories using its spatiotemporal memory, providing global recall capabilities [2] - The model employs a reasoning strategy that intertwines text and spatial positioning, effectively reducing hallucination issues during reasoning processes [2] - RynnBrain possesses multiple capabilities, including cognition, localization, reasoning, and planning [3] Training and Performance - Trained on the Qwen3-VL foundation using the proprietary RynnScale architecture, RynnBrain's training data exceeds 20 million pairs, resulting in a performance boost of two times with equivalent resources [4] - RynnBrain has set new records in various benchmarks, including environmental perception, object reasoning, first-person visual question answering, spatial reasoning, and trajectory prediction [4][5] Benchmark Results - RynnBrain outperforms other top models in multiple benchmarks, achieving scores such as 74.5 in VSI-Bench and 78.9 in EgoTaskQA, demonstrating its superior performance across various tasks [5] Scalability and Applications - RynnBrain exhibits excellent scalability, allowing for rapid fine-tuning to create various embodied models, such as navigation and planning models, with minimal data [6] - The model's architecture allows it to exceed the performance of Gemini 3 Pro with just a few hundred data points for fine-tuning [6] Open Source and Industry Impact - The open-sourcing of RynnBrain's complete reasoning and training code, including the first MoE architecture in the industry, positions it as a foundational model for the embodied intelligence sector [6][8] - Damo Academy aims to accelerate the transition of AI from digital to real-world applications, contributing to the development of deployable, scalable, and evolvable embodied intelligence systems [8]