# Lifelong Topological Visual Navigation

**Authors:** Rey Reza Wiyatno, Anqi Xu, Liam Paull  
**arXiv:** [2110.08488](https://arxiv.org/abs/2110.08488)  
**Date:** 2021-10-16  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2110.08488v2/

## Key points

- The paper addresses challenges faced by learning-based topological navigation approaches
- Spurious or missing edges in the topological graph often lead to navigation failure
- The authors propose a sampling-based graph building method for sparser graphs with higher navigation performance
- Graph maintenance strategies are introduced to eliminate spurious edges and expand the graph as necessary
- Fine-tuning the model using real-world trajectory images enables successful navigation in real-world environments
- Lifelong graph maintenance strategies result in significant improvements in navigation over time
- The approach adapts to changing environments and ensures reliable long-term navigation performance

## Summary

The paper titled "Lifelong Topological Visual Navigation" by Rey Reza Wiyatno, Anqi Xu, and Liam Paull addresses the challenges faced by learning-based topological navigation approaches. These approaches typically produce a local policy while maintaining loose connectivity of the space through a topological map. However, the presence of spurious or missing edges in the topological graph often leads to navigation failure. To overcome this limitation, the authors propose a sampling-based graph building method that results in sparser graphs with higher navigation performance compared to baseline methods. Additionally, they introduce graph maintenance strategies that eliminate spurious edges and expand the graph as necessary, thereby improving lifelong navigation performance. One notable aspect of their work is that unlike controllers that learn from fixed training environments, their model can be fine-tuned using only a small number of collected trajectory images from a real-world environment where the agent is deployed. This fine-tuning process enables successful navigation in real-world environments. Furthermore, the authors demonstrate significant improvements in navigation over time by applying their lifelong graph maintenance strategies. This highlights the effectiveness of their approach in adapting to changing environments and ensuring reliable long-term navigation performance. Overall, this paper presents a novel approach to topological visual navigation that addresses issues related to spurious or missing edges in the topological graph. The proposed sampling-based graph building method and graph maintenance strategies contribute to improved lifelong navigation performance. The ability to fine-tune the model using real-world trajectory images further enhances its practical applicability.