Success Classification for Object Navigation

Abstract

Object navigation is the embodied task of navigating to an instance of a specified object in unseen environments. Previous work has made impressive progress on the problem, but there remains much room for improvement with current state-of-the-art methods reaching a success rate of less than one in three. In this work, we evaluate a state-of-the-art approach, identifying false positives in object detection as the main point of failure. We propose introducing a new module to verify success when the agent attempts to stop. We introduce a learning-based classifier that learns and compares embeddings for visual observations and object categories and find that it works well at predicting success, outperforming both naive baselines and a heuristic­-based classifier. We also find no improvement when using a ensemble model for semantic segmentation, although we believe there is more to be tested before arriving at a conclusive judgement.