In recent years, the embodied AI data industry has witnessed exponential growth, largely fueled by the extensive training needs of embodied AI robots. From robot "trainees" tirelessly operating in training arenas, to human data collectors outfitted with sensors in motion capture facilities, and further to simulated training environments within virtual worlds—where a single data point can spawn multiple derivatives—the sector is drawing a multitude of participants into the physical AI revolution. Real-world machine data, intangible collection data, and simulated data represent the three fundamental data types crucial for embodied AI training, each characterized by unique production methodologies and inherent constraints. As the industry swiftly scales up its data volume, experts underscore the importance of not neglecting data quality and evaluation criteria amidst this rapid expansion.