Analyzing and Improving Object Navigation in Indoor Environments

Abstract

Object navigation is the embodied artificial intelligence task of an agent navigating to a target object within an environment. While there have been significant advances in this task, models still struggle to reach human performance and generalize to novel environments. One high-performing model, SemExp, combines a semantically-aware long-term planning module with a simultaneous localization and mapping (SLAM) system. We analyze this model’s components and failure modes, finding that failures often occur due to semantic errors especially near the end of task episodes. We then use the insights from the analysis to guide the development of improvements to the model.