机器人能跑能跳能搏击，为何仍陷“成长烦恼”？

Core Viewpoint - The article discusses the current state and challenges of the humanoid robotics industry, highlighting the need for better integration between embodied manufacturers and AI model companies, as well as the importance of data quality and quantity for advancing technology [3][6][16]. Industry Growth and Challenges - The humanoid robot market in China is experiencing strong growth, with commercial sales expected to rise from approximately 2,000 units in 2024 to nearly 60,000 units by 2030, reflecting a compound annual growth rate of 95.3% [6]. - Despite the growth, the industry faces significant challenges, including high costs, inconsistent hardware quality, data quality and scale gaps, immature software and algorithms, and a lack of supply chain standards [7][8]. Technological Development - The industry is currently divided into three main technological routes: end-to-end VLA (Vision-Language-Action), layered architectures, and brain-like models that simulate biological neurons [12]. - There is a consensus on the importance of the "small brain" technology for motion control, but the "big brain" aspect, which involves decision-making and learning, has not yet reached an ideal state [11][12]. Data Issues - A significant barrier to progress in the humanoid robotics sector is the insufficient amount of quality data available for training models. The industry currently relies on a limited dataset, with the largest public dataset containing only 1 million entries [16][17]. - Companies are exploring various methods to generate data, including using computer graphics to simulate real-world physics and creating large interactive object libraries [17]. Integration and Collaboration - The development of humanoid robots is seen as a systemic engineering challenge, where the integration of various technological routes and data sources is crucial for creating effective systems [13]. - The industry requires collaboration among multiple companies to address the full chain of model development, data acquisition, and robot functionality [18].