Robotic tactile perception has become a core capability for realizing robotic embodied intelligence, yet faces challenges such as poor dynamic adaptation, overlapping perceptual states, and stringent real-time requirements. To address this, this paper employs a spatial-temporal decoupling-based deep learning framework, and proposes a Transformer-Mamba-based classification method for robotic tactile perception. Firstly, a multi-scale feature enhancement module based on an efficient channel attention mechanism is designed to strengthen the spatial-temporal multi-scale features of tactile perception data. Secondly, a feature extraction module based on Transformer-Mamba is constructed to extract spatial and temporal features from tactile perception data. Herein, Transformer serves as the core for extracting spatial features, while Mamba serves as the core for extracting temporal features, thereby reducing the algorithmic inference time. Furthermore, we design a feature classifier based on gated fusion, which reduces feature loss and improves the algorithm’s stability by fusing cross-stage features. Finally, we conduct experiments to validate the proposed algorithm using three public datasets: Tactile-event-corl, Multimodality slip detection, and Dataset for SCT-CNN. The results demonstrate that the proposed algorithm achieves an accuracy of 98.72%, precision of 98.87%, recall of 98.85%, and F1 score of 98.85%, representing improvements of 1.67%, 1.41%, 1.70% and 1.65% respectively over existing algorithms, while reducing the average inference time by 90%.