8位具身智能顶流聊起「非共识」：数据、世界模型、花钱之道

Core Viewpoint - The article discusses the emerging industry revolution driven by embodied intelligence in the AI era, highlighting the diverse perspectives of top practitioners in the field regarding the allocation of significant funding for its development [5][6]. Group 1: Funding Allocation and Perspectives - During a roundtable forum, participants were asked how they would allocate 10 billion yuan to advance embodied intelligence, revealing varying strategies and priorities among industry leaders [5][6]. - Some participants emphasized the need for collaboration and building data ecosystems, while others focused on addressing data bottlenecks and creating self-evolving data systems [7][68]. Group 2: Data Challenges and Solutions - A significant discussion point was the "data scarcity" issue, with differing opinions on the importance of real-world data versus synthetic data for training models [9][10]. - Participants highlighted the necessity of high-quality, diverse data collected from real-world scenarios to enhance model performance, with some advocating for a combination of real and synthetic data [43][44][50]. Group 3: World Models and Embodied Intelligence - The concept of world models was debated, with some experts agreeing on their importance for embodied intelligence, while others suggested that they are not a mandatory foundation [14][17]. - The need for predictive capabilities in robots was emphasized, suggesting that training data must come from the robots' own experiences to be effective [16][18]. Group 4: Future Model Architectures - There was a consensus that embodied intelligence requires a unique model architecture distinct from existing large language models, with some advocating for a vision-first or action-first approach [19][20][21]. - The idea of a unified model that integrates various elements such as vision, action, and language was discussed, with the potential for a closed-loop system that allows for real-time feedback and adjustment [22][24][25]. Group 5: Long-term Vision and Data Collection - Participants expressed that the development of a powerful embodied intelligence model would depend on accumulating vast amounts of real-world data through practical applications and interactions [27][60]. - The importance of creating a "data flywheel" through the deployment of robots in real environments was highlighted as a means to gather diverse and extensive data [50][51][56].