Language to Map: Create Maps from Natural Language

A study conducted by Hideki Deguchi et al., was selected for the presentation at the 2024 IEEE International Conference on Robotics and Automation (ICRA).

A map of the environment is essential for tasks such as wayfinding and robot navigation. Conventional methods create these maps using information obtained from sensors like cameras and LiDAR. However, if the user is aware of the environment, this knowledge can be directly mapped using natural language path instructions.
In this study, we propose a method to generate maps from user natural language path instructions using a system based on a large language model. The proposed method extracts waypoints from user-provided path instructions and identifies the action (straight, turn right, and, turn left) at each waypoint to create a map. The map is represented as a node–edge graph, where nodes represent the extracted information and edges represent the connections between nodes. When using the map, the user provides the destination name, and the map generates path instructions in natural language. The proposed method not only repeats the path instruction stored by the map, but also combines the action–information of each node to generate a path for an unspecified destination. This method is expected to facilitate the easy creation and modification of maps, even in environments where no current map is available.

Title: Language to Map: Topological Map Generation from Natural Language Path Instructions
Authors: Deguchi, H., Shibata, K., Taguchi, S.
Appears in: IEEE International Conference on Robotics and Automation
Presented: May 15, 2024