Lifelong Topological Visual Navigation

Commonly, learning-based topological navigation approaches produce a local policy while preserving some loose connectivity of the space through map. Nevertheless, spurious or missing edges in graph often lead to failure. In this work, we propose sampling-based building method, which results sparser graphs yet with higher performance compared baseline methods. We also maintenance strategies that eliminate and expand as needed, improves lifelong performance. Unlike controllers learn from fixed training environments, show our model can be fine-tuned using only small number collected trajectory images real-world environment where agent is deployed. demonstrate successful after fine-tuning on notably significant improvements over time by applying strategies.