A LIGHTWEIGHT DEEP LEARNING MODEL BASED ON MOBILENETV2 FOR THE NAVIGATION FOR A DISCRETE MOBILE ROBOT

This paper presents the design of a lightweight, deep learning model, based on MobileNetV2, to classify navigation decisions in a discrete robot path-planning environment. The model was created with the help of a structured dataset containing 8,000 balanced samples based on a 5×5 grid-based navigation system. Each sample is also characterized by spatial, directional and obstacle-related characteristics like robot position, goal position, obstacle distances and derived metrics such as the Euclidean distance and angle to the goal. The transfer learning was implemented with pre-trained ImageNet weights to enhance the efficiency of feature extraction and minimise the complexity of training. This model was trained and implemented in Google Colab with TensorFlow and Keras libraries. The model proposed divides the states of navigation into four actions, i.e., forward, left, right, and stop (obstacle detection). The experimental results indicate that the model had a total accuracy of 95.6, and the values of precision, recall, and F1-score of the model were uniformly high across all classes. The best performance was achieved on obstacle detection and few misclassifications were observed between left and right action since the grid environment was symmetrical. These findings indicate the usefulness of MobileNetV2 to learn spatial decision patterns to solve navigation problems. Summing up, the research establishes that MobileNetV2 with transfer learning offers a highly efficient and accurate way of classifying navigation decisions. The model is lightweight, computationally efficient and is applicable to real-time intelligent systems. This renders it a good prospective candidate to be integrated in future autonomous robotic navigation systems and other integrative AI applications.