Map building and path planning are essential for long-term visual navigation. Methods, such as semi-parametric topological memory (SPTM), use deep learning to build a topological map consisting of nodes sampled from an agent's past observations and edges generated by a neural network (NN) and perform path planning on the map. Such methods require neither accurate sensor nor advanced expertise for map building needed in classical metric-map-based techniques such as visual simultaneous localization and mapping. However, they often plan improper paths including teleportations and detours, even if we have collected enough training data. In this paper, we propose a topological-map-based planning method that includes two modifications to SPTM to address this problem. The first modification is that we observe omnidirectional images and construct an NN on the basis of spherical convolutional NNs, which guarantee rotation invariance on omnidirectional images. The second modification is that we train the NN on a metric learning framework. We conducted experiments to show the effectiveness and applicability to real world of our method for path planning in visual navigation. The results indicate that the first modification prevents detouring and the second one prevents teleportations in path planning.
展开▼
