To address the low accuracy of pose correction and poor performance of environmental mapping in existing inspection robots, this study proposes a novel technique integrating multi-sensor data and deep learning to enhance positioning accuracy and operational efficiency.Pose data from cameras, wheel encoders, and gyroscopes were fused using a CNN-RNN composite model (3×3 convolution kernels, 20 convolutional layers, 5 RNN hidden layers). Environmental images were processed via SIFT feature extraction, transformation matrix registration, and BiGEMAP-based planar grid map construction. Dynamic probabilistic models and improved support vector machines (SVM) were applied for passable region classification.Ablation experiments showed that the CNN-RNN model achieved 99.5% pose correction accuracy after 26 iterations. The minimum overlapping ratio of stitched images was 1.02%, with a passability misjudgment rate of 1%. The constructed map achieved an SSIM value of 1.0 compared to ground truth.The integration of CNN-RNN and multi-sensor data significantly improves pose correction precision and mapping reliability, offering a robust solution for autonomous navigation of inspection robots in complex environments.