Text-To-Motion Generation for Diverse Human Body-Motion Simulation

For practical deployment of autonomous robots in human-existed environment, it is essential to train robot policies with simulation environments that reflect the diversity of human body features and motion behaviors. However, existing datasets often overlook this need, relying on simplified skeleton models that ignore variations in age, height, or body shape. Moreover, realistic scenarios representing such diversity are largely missing due to the high cost and complexity of data collection. In this work, we address these limitations by constructing a human motion dataset that captures a wide range of body types and age groups, using accurate and characterized body models. These detailed representations allow robots to better learn how physical attributes influence movement, thereby enhancing their responsiveness and safety during interaction. To further expand the dataset efficiently, we also explore data generation techniques that create diverse motion samples from limited inputs. Our approach enables the scalable construction of simulation environments that reflect human variability, offering a valuable resource for future robot policy learning.