Navigating to Objects Specified by Images

Team:

Suciu Alexandru

Stan Claudiu Mihnea

Ursu Andrei

 

Introduction:The advancement of navigation systems has become an essential aspect of modern technology. In this paper, the authors present a significant breakthrough in the field of instance-based Image Goal Navigation. They introduce a modular system that addresses the challenges of navigating towards specific goals in image-based scenarios. The system is composed of four distinct components: Exploration, Instance Re-Identification, Goal Localization, and Local Navigation. Through rigorous testing, the proposed method achieves a remarkable 56% success rate in the HM3D InstanceImageNav benchmark and an impressive 88% in real-world tests. This paper offers a significant contribution to the field of navigation systems and opens up new possibilities for image-based navigation.

Related work: In the realm of navigation systems, Image Goal Navigation is a critical form of embodied navigation that relies on visual descriptions provided by an image. However, previous methods utilizing deep reinforcement learning (DRL) have faced significant challenges, particularly in the areas of visual scene understanding, semantic exploration, and long-term memory. In this paper, the authors propose a novel modular approach to address these challenges and solve the InstanceImageNav task without the need for fine-tuning. This approach represents a significant advancement in the field and offers a promising solution for future image-based navigation systems.

            Method: The authors of the paper(https://arxiv.org/abs/2304.01192) propose breaking down the InstanceImageNav task into sub-tasks, including Exploration, Goal Instance Re-Identification, Goal Localization, and Local Navigation. Each sub-task contributes to solving the overall problem by exploring the space, re-identifying the object instance, localizing the goal, and navigating through the environment.

            Experiments: The researchers evaluated their modular method for InstanceImageNav in simulation and real-world scenarios, comparing it to prior art and alternate sub-task modules. The proposed Modular InstanceImageNav (Mod-IIN) method outperforms the baseline model by 6.8x and the state-of-the-art ImageNav model (OVRL-v2) by 2.3x.

            Conclusion: The study presents a modular system for InstanceImageNav that does not require fine-tuning. The system outperforms end-to-end learned policies and successfully transfers to real-world execution. Despite some limitations, the system serves as a solid baseline for evaluating trained navigation policies and developing modular semantic navigators.