Development and comparative analysis of two improved models for indoor embodied navigation

Service robots usually need to navigate in a complex indoor environment, and sometimes robots need to perform target search tasks autonomously without a prebuilt map. The existing navigation algorithms have problems such as slow response and long time-consuming model training. Based on SemExp, we propose enhanced learning models that integrate image recognition with PPO algorithms. By refining the neural network architecture, modifying the feature map sampling, and redesigning the linear layer dimensions, our method achieves rapid convergence within short-term training. Simulation experiments show that the algorithm significantly improves navigation performance, and the training rounds and time required to obtain the best model are also greatly reduced. We also get a model that can overfit more slowly, which performs better under higher frames training.