推动具身智能跨入ChatGPT-3时代，上海布局全球最大数据集

Core Insights - The launch of the "Longyue" (MindLoongGPT) generative humanoid robot motion model marks a significant advancement in China's capabilities in the field of embodied intelligence and motion control, positioning the country as a global leader in this area [1][3][4] Group 1: Model Development and Features - The Longyue model utilizes a "natural language-driven" approach, creating a complete loop from multimodal input to high-fidelity motion generation, thus revolutionizing traditional motion control paradigms [3] - The model allows users to generate actions simply by inputting commands or reference videos, breaking the reliance on specialized parameter adjustments and making it accessible to a broader audience [4] - The model aims to generate simulation-level actions to enhance the dataset for embodied intelligence, especially in scenarios where data collection is challenging [3][4] Group 2: Data Challenges and Solutions - The development of embodied intelligence faces significant challenges in ensuring both the quantity and quality of data, with the complexity of environments for humanoid robots being much higher than that for autonomous vehicles [5][6] - The China Academy of Information and Communications Technology (CAICT) and the National and Local Joint Innovation Center are working together to establish quality standards for embodied intelligence data, focusing on the entire data lifecycle from collection to usage [6] - The goal is to create a comprehensive data ecosystem with seven quality evaluation indicators, including completeness, consistency, diversity, authenticity, usability, practicality, and scalability [6] Group 3: Future Directions and Infrastructure - The National and Local Joint Innovation Center aims to enhance the operational accuracy of their humanoid intelligence model from approximately 70%-80% to 90% by the end of the year, targeting a data volume in the petabyte range [8] - The "Ge Wu-Zhi Zhi" general embodied intelligence development platform integrates core software frameworks to facilitate rapid development and promote an ecosystem approach to humanoid robot development [8] - The establishment of a one-stop development infrastructure at the Qilin training ground supports data collection from over 100 heterogeneous robots, providing robust data support for future model training [8][9]