With semantic reasoning on the partial scene reconstruction, our agent can leverage the
general scene layout priors and predict the possible target locations. As a result, our
agent can navigate to the target object with shorter paths in unseen environments.
Abstract
Summary: This work presnets a modular and training-free system to tackle the object goal
navigation problem.
The system builds a structured scene representation during active exploration, propagates semantics
in the scene graphs to infer the target location, and
introduces those semantics to the geometric frontiers. With semantic frontiers, the agent navigate
to the most promising areas to search for the goal object and avoid detours in unseen environments.
Intuition of this work. ObjectNav can be decomposed into 1) semantic questiosn: inferring the
potential position
of the target object in the scene and 2) geometric question: point-to-point navigation. These two questions can be solved by semantic reasoning and classic planning separately on different representations.
Video
Method
The proposed method builds a structured scene representation on the fly, which consists of a semantic
point cloud, a 2D occupancy map, and a spatial scene graph. Then the scene graph is used to propagate semantics to the geometric frontiers.
With semantic frontiers, the agent can navigate to the most promising areas to search for the goal object and avoid detours in unseen environments.
Junting Chen, Guohao Li, Suryansh Kumar, Bernard Ghanem, Fisher Yu. How To Not Train Your Dragon: Training-free Embodied Object Goal Navigation with Semantic
Frontiers.
In Conference Robotics: Science and Systems, 2023.
(hosted on ArXiv)
@misc{chen2023train,
title={How To Not Train Your Dragon: Training-free Embodied Object Goal Navigation with Semantic Frontiers},
author={Junting Chen and Guohao Li and Suryansh Kumar and Bernard Ghanem and Fisher Yu},
year={2023},
eprint={2305.16925},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
