Accurate perception and analysis of human motion form the cornerstone of multimodal computing, sports medicine, and rehabilitation engineering. However, existing datasets are often limited to a single sensing modality or specific simple actions, lacking comprehensive data that concurrently integrates kinematic, kinetic, and physiological information. In this work, we propose MM-Motion, a multimodal dataset featuring 16 standardized actions performed by 33 subjects. The dataset comprises 933 IMU data files, 1,866 Kinect video files including over 260,000 RGB-D image pairs, 918 pressure insole data streams, and fundamental physiological metadata. Through hardware synchronization and data preprocessing, the dataset achieved high-precision temporal alignment at the decimillisecond level. All video sequences were evaluated by experts using a multi-dimensional joint-scoring scale to obtain fine-grained action quality scores.MM-Motion provides high-quality data support for personalized assessment and intervention in intelligent sports medicine. By establishing a multi-task benchmark framework encompassing action recognition, quality assessment, and injury risk stratification, this research validates the high fidelity of the MM-Motion dataset. Furthermore, the experimental results demonstrate that the multimodal fusion of kinematics and kinetics significantly enhances the accuracy of ankle sprain risk prediction, underscoring the necessity of collecting diverse sensory data for complex physiological analysis.