A survey of visual navigation: From geometry to embodied AI

Abstract

The capacity to extract information and comprehend an unseen environment is critical for mobile robots to navigate. Few surveys has mentioned the combinatorial-non-optimality problem of the traditional visual navigation methods. As computer vision technology has improved in recent years, visual navigation approaches have escalated drastically, particularly after the appearance of the CVPR Embodied AI workshop. However, few studies take these important changes into account. This survey fills this research gap by collecting, analyzing, and summarizing more than 100 recent papers. The majority of them are published within 5 years and are cited over 80 times, which provide more credible results. Based on our thorough comparison, this survey categorizes all visual navigation methods into two styles: geometry style and embodied AI style. This survey examines these two styles from the perspective of input–output. In addition, this survey attempts to provide mathematical formulations for each style. This paper provides a case study to illustrate the methodological paradigm with greatest potential. This methodological paradigm using photo-realistic simulation in the Embodied AI style, which could solve the combinatorial-non-optimality problem. Thereafter, this survey discusses several issues including pros–cons analysis, problem formulation, common framework, task generalization, dynamic environment consideration, sim-to-real, and inspiring approaches, which are all based on the scholars who have cited the method. In the last part, challenges and future trends are summarized. This survey would assist researchers who work on AI-empowered visual navigation systems.