Abstract: A model based on the integration of bidirectional temporal convolutional network (BiTCN) and attention mechanism is proposed to address the challenges in state perception of robotic footstep interaction with wet granular terrain, and meet the real-time and high-precision requirements for robot walk control in scenarios such as post-disaster rescue and planetary exploration. Firstly, the designed model leverages the advantages of BiTCN in processing time series data, combined with the ability of the attention mechanism to capture key information, achieving accurate perception of the robot foot shape, horizontal speed, and penetration depth. Furthermore, the Bayesian optimization (BO) method is introduced to further enhance the model performance, and automatically adjust key hyperparameters of the model, including learning rate, dropout rate, attention dimension, and number of network layers. Compared with the pre-optimized model, a reduction of about 77.8% and 12.7% is achieved after optimization in mean squared error for speed and depth, respectively. In the experiment section, the design of the mechanical parameter acquisition system, experimental methods, and data acquisition and processing procedures are elaborated. Orthogonal experimental design is employed to conduct experiments and analyze data. The model training and validation results indicate that the model can effectively improve the accuracy and robustness of state perception, providing strong support for robot walk control in complex environments.