独家对话元客视界CTO：揭秘具身智能大模型的“数据飞轮”密码

Core Insights - The development of humanoid robots and embodied intelligence is still in its early stages, akin to "kindergarten level," with current capabilities limited to basic tasks like grabbing and walking, while facing challenges in complex interactions and task execution [1][6][19] - Achieving "general intelligence" requires a complete perception-reasoning-execution chain, supported by a large volume of high-quality data to enhance model capabilities and product performance [1][2][19] Data and Model Training - The performance of embodied intelligence models follows the Scaling Law, indicating that model performance improves proportionally with increased parameters and data, with a threshold of 100 million high-quality behavior trajectory data points identified as critical for significant capability leaps [2][19] - A mixed training approach using 10% real data and 80% simulated data is preferred to enhance model generalization and efficiency, addressing the limitations of both pure real and simulated data [7][19] Data Collection Techniques - Motion capture technology is essential for data collection, with optical and inertial capture being the two main methods, each having its advantages in precision and continuity [8][10] - The company has achieved an 83% utilization rate in data collection, significantly improving efficiency by reducing time lost in adjustments [10][19] Challenges in Implementation - Key challenges include hardware durability, the need for high-quality data, and efficiency in task execution, which currently lags behind human performance [6][19] - The industry faces a "Sim2Real Gap," where simulated environments do not fully replicate real-world complexities, necessitating a blend of real and simulated data for effective training [7][19] Future Directions - The company aims to enhance data collection precision and efficiency through ongoing development of optical and inertial fusion techniques, while also collaborating with large model technology firms to optimize training efficiency [24][25] - A comprehensive evaluation system is being developed to assess robot performance across various metrics, including stability and energy efficiency, which are critical for commercial viability [18][19]