Data Acquisition for Vision-Language-Action Models: From Real Robots to Simulation and Video Data

With the rapid development of artificial intelligence, embodied intelligence has become a prominent research frontier. In the paradigm of large embodied models, Vision-Language-Action (VLA) models have emerged as a powerful approach, endowing agents with generalizable multimodal perception, decision-making, and action execution capabilities. The form and quality of training data play a critical role in determining training cost, task success rate, and generalization performance. This paper presents a systematic survey of data acquisition methods and representative models for VLA training. First, real-robot data collection strategies based on teleoperation and kinesthetic teaching are introduced. Second, synthetic data generation approaches driven by physics engines and world models are elaborated. Third, strategies leveraging large-scale human demonstration videos are analyzed. Finally, the advantages and limitations of these three paradigms are comprehensively compared, and promising future directions for VLA data acquisition are discussed.