Smart data gloves capable of monitoring finger activities and inferring hand gestures are of significance to human–machine interfaces, robotics, healthcare, and Metaverse. Yet, most current smart data gloves present unstable mechanical contacts, limited sensitivity, as well as offline training and updating of machine learning models, leading to uncomfortable wear and suboptimal performance during practical applications. Herein, highly sensitive and mechanically stable textile sensors are developed through the construction of loose MXene-modified textile interface structures and a thermal transfer printing method with the melting-infiltration-solidification adhesion procedure. Then, a smart data glove with adaptive gesture recognition is reported, based on the integration of 10-channel MXene textile bending sensors and a near-sensor adaptive machine learning model. The near-sensor adaptive machine learning model achieves a 99.5% accuracy using the proposed post-processing algorithm for 14 gestures. Also, the model features the ability to locally update model parameters when gesture types change, without additional computation on any external device. A high accuracy of 98.1% is still preserved when further expanding the dataset to 20 gestures, where the accuracy is recovered by 27.6% after implementing the model updates locally. Lastly, an auto-recognition and control system for wireless robotic sorting operations with locally trained hand gestures is demonstrated, showing the great potential of the smart data glove in robotics and human–machine interactions.