Lack of Training Data Hinders Humanoid Robot Development

The shortage of training data required for AI models has emerged as a critical challenge in the development of humanoid robots. Unlike large language models that have evolved rapidly by learning vast amounts of text, images, and videos available on the internet, robots cannot apply this approach directly. Physical movements and interactions with the environment require real-world motion data that digital content alone cannot provide.

One of the key driving forces supporting the advancement of AI models is the massive amount of data accumulated on the internet. Digital content such as text, images, and videos has served as a "teacher" enabling AI to understand human language, recognize images, and generate text. However, the physical actions that robots must acquire—such as grasping objects, stepping over obstacles, and handling tools—exist in forms of data that are rarely found on the internet. How to move in real physical environments cannot be taught through digital data alone.

Against this backdrop, collecting training data for robot development requires a fundamentally different approach than text-based AI. Methods such as operating actual robots to record motion data or simulating movements in virtual environments are employed, but both are time-consuming and costly. It is difficult for the quality and quantity of data to match the scale that language models have benefited from.

Industry interest in humanoid robots has grown in recent years, with applications envisioned in manufacturing, logistics, and care services. However, to accelerate development, it is not just software and computational power that matter—the key lies in how effectively and efficiently robots can gather large-scale data from the real world to "learn through their bodies." This challenge underscores the structural fact that AI's evolution has been supported by the abundance of data.

The difficulty of data collection directly affects the pace of robot development. While language AI achieved dramatic performance improvements in just a few years, the learning process in robot body control has been more gradual, in part due to this asymmetry in training data. To develop robots that can adapt to diverse real-world environments and unpredictable situations, accumulation of vast and varied motion data is essential.

In future development, attention is focused on the advancement of simulation technology and the adoption of "teleoperation"—a method in which humans remotely operate robots while collecting data. Additionally, efforts by multiple companies and research institutions to share data could help address challenges across the industry. Whether the industry can overcome the data barrier represents a critical watershed for the practical deployment of humanoid robots.