SIASUN(SZ:300024) Di Yi Cai Jing·2025-07-28 02:07Core Viewpoint - The debate between the value of real-world data versus simulation data in robot training is intensifying, with industry leaders emphasizing the necessity of real data for complex tasks while acknowledging the cost-effectiveness of simulation data for simpler tasks [1][2][4]. Group 1: Importance of Real Data - Sergey Levine, co-founder of Physical Intelligence, argues that real-world data is essential for effective robot training, challenging the reliance on simulation data [1]. - Industry experts, such as Yao Maoqing from Zhiyuan Robotics, support Levine's view, stating that while some tasks can be trained using simulation, most complex tasks require real data [1][3]. - The CEO of Qingtong Intelligent, Li Tong, emphasizes that robots must be deployed in real environments to accumulate valuable training data, suggesting that a deployment scale of tens of thousands is necessary for effective data collection [3]. Group 2: Simulation Data Advantages - Companies like Galaxy General advocate for simulation data, claiming it allows for faster learning and lower costs, even enabling training without real data [2]. - The COO of Self-Variable Robotics, Yang Qian, acknowledges the role of simulation in training lower-body movements but stresses that real-world data is crucial for tasks involving complex interactions [10][12]. - The industry faces a dilemma in balancing the use of simulation and real data, with some companies using a 7:3 ratio of simulation to real data, while others prefer a 3:1 ratio favoring real data [9][10]. Group 3: Challenges and Future Directions - The industry is grappling with the technical challenge of integrating simulation and real data effectively, as highlighted by Chen Yuanpei from Lingchu Intelligent, who notes that data from different sources must be weighted differently [9]. - The consensus is that while simulation data is beneficial for initial training phases, real data is indispensable for achieving advanced capabilities in robots [10][12]. - Companies are increasingly focusing on building extensive real-world data sets to enhance their models, with Zhiyuan Robotics aiming to create a comprehensive dataset to support embodied intelligence [10][12].