From Route Instructions to Landmark Graphs
This addresses the gap in navigation technologies that lack landmark incorporation, potentially improving human-computer interaction in domains like robotics and delivery services.
The paper tackles the problem of generating landmark-based spatial representations from natural language route instructions, proposing a fully end-to-end neural approach. It shows high quality results on the SAIL dataset and real-world delivery instructions, with performance also demonstrated on robotic navigation tasks.
Landmarks are central to how people navigate, but most navigation technologies do not incorporate them into their representations. We propose the landmark graph generation task (creating landmark-based spatial representations from natural language) and introduce a fully end-to-end neural approach to generate these graphs. We evaluate our models on the SAIL route instruction dataset, as well as on a small set of real-world delivery instructions that we collected, and we show that our approach yields high quality results on both our task and the related robotic navigation task.