专访星海图赵行：热闹的Demo不等于泛化能力，具身智能胜负仍在数据量

Core Insights - The demonstration of bed-making by the robot at the 2025 WRC highlights the complexity of seemingly simple tasks, showcasing the robot's capabilities in flexible object manipulation and full-body control [1][2][4] - The newly released G0 model by the company aims to enhance generalization capabilities in embodied intelligence, moving beyond previous smaller models that struggled with scalability [2][4][11] - The company emphasizes the importance of high-quality data collection and engineering processes to support the development of robust models, with a focus on real-world data [4][19][28] Group 1: Technology and Model Development - The G0 model utilizes a three-stage training framework that has shown a 20% improvement over the previous PI 0 model in average metrics [9][10] - The company plans to open-source a dataset of 500 hours of real-world data to establish a high-quality benchmark for the industry, facilitating comparisons and algorithm validation [5][30] - The focus on data collection involves training personnel and addressing various challenges in real-time data acquisition, which is considered foundational for model training [19][22][24] Group 2: Industry Context and Future Directions - The company believes that the scaling laws observed in large language models can also apply to embodied intelligence, suggesting a potential for significant advancements in the field [14][16] - The VLA paradigm is seen as a primary industrial path, with ongoing exploration of additional technologies such as tactile sensing and world modeling for future applications [32][39] - The collaboration between academia and industry is viewed as beneficial, with the potential for academic insights to drive industrial advancements and vice versa [45][46]