This work studies the problem of object goal navigation which involves navigating to an instance of the given object category in unseen environments. End-to-end learning-based navigation methods struggle at this task as they are ineffective at exploration and long-term planning. We propose a modular system called, `Goal-Oriented Semantic Exploration' which builds an episodic semantic map and uses it to explore the environment efficiently based on the goal object category. Empirical results in visually realistic simulation environments show that the proposed model outperforms a wide range of baselines including end-to-end learning-based methods as well as modular map-based methods and led to the winning entry of the CVPR-2020 Habitat ObjectNav Challenge. Ablation analysis indicates that the proposed model learns semantic priors of the relative arrangement of objects in a scene, and uses them to explore efficiently. Domain-agnostic module design allows us to transfer our model to a mobile robot platform and achieve similar performance for object goal navigation in the real-world.

This work studies the problem of object goal navigation which involves navigating to an instance of the given object category in unseen environments. End-to-end learning-based navigation methods struggle at this task as they are ineffective at exploration and long-term planning. We propose a modular system called, Goal-Oriented Semantic Exploration' which builds an episodic semantic map and uses it to explore the environment efficiently based on the goal object category. Empirical results in visually realistic simulation environments show that the proposed model outperforms a wide range of baselines including end-to-end learning-based methods as well as modular map-based methods and led to the winning entry of the CVPR-2020 Habitat ObjectNav Challenge. Ablation analysis indicates that the proposed model learns semantic priors of the relative arrangement of objects in a scene, and uses them to explore efficiently.

Summary and Contributions: This paper presents an extension to recent work on Active Neural SLAM [1], where semantic information about object categories is explicitly incorporated into the model. The extensions in the model architecture provide explicit semantic information about the various objects of the scene in the generated 2D map, that allows an agent to navigate in its environment and find a specified goal object much efficiently compared to baselines. Some of these baselines use - and others do not - semantic information. The comparison was performed using Gibson [2] and Matterport3D (MP3D) [3], which include 3D reconstructions of real environments. Training was performed on 86 scenes and testing on 16.

This paper proposes to train an ObjectNav policy that generalises to unseen environments by using a modular system that classifies objects and builds an episodic semantic map, which it is uses to explore the environment based on the object category, building upon the hierarchical method in "Learning to explore using Active Neural SLAM". The method achieved SOTA performance on the 2020 CVPR Object Goal Navigation Habitat Challenge. Interestingly, the policy, trained on Gibson and MP3D, has been transferred and deployed in a real robot, with some success. While the initial reviews were mixed (9, 7, 4, 5), the reviewers converged on (8, 7, 6, 6), agreeing during discussion that the paper deserved to be accepted. Based on the reviews, I recommend this paper for acceptance as a spotlight or poster presentation.

This work studies the problem of object goal navigation which involves navigating to an instance of the given object category in unseen environments. End-to-end learning-based navigation methods struggle at this task as they are ineffective at exploration and long-term planning. We propose a modular system called, Goal-Oriented Semantic Exploration' which builds an episodic semantic map and uses it to explore the environment efficiently based on the goal object category. Empirical results in visually realistic simulation environments show that the proposed model outperforms a wide range of baselines including end-to-end learning-based methods as well as modular map-based methods and led to the winning entry of the CVPR-2020 Habitat ObjectNav Challenge. Ablation analysis indicates that the proposed model learns semantic priors of the relative arrangement of objects in a scene, and uses them to explore efficiently.