Topological Map

A map made of place nodes and connecting edges, recording only what leads where, not exact coordinates.

A topological map represents an environment as a graph: nodes are representative places (a particular room, an intersection, or a keyframe image), and edges mean that two places can be reached directly from each other. Unlike a metric map such as an occupancy grid, it does not require an exact metric coordinate for every location, so it takes far less storage, can scale to very large environments, and is closer to how a person describes a route (“go out the door, turn left down the hallway, then into the kitchen”). During navigation, an algorithm such as A* first searches the graph for a sequence of nodes, and a local controller or policy then moves between adjacent nodes. Recent learned navigation systems make heavy use of this idea: ViNT and NoMaD use images as nodes for long-range navigation, while Mobility VLA builds a topological map from a video of a person walking around giving a tour, then uses a vision-language model to interpret an instruction and find the target node.

ExampleA robot is first walked around an office building while recording video, saving one image as a node every few seconds; later, told “go to the printer,” it finds the matching node in the graph and plans a path of nodes to reach it.

Also called

Topological Graph

Related

Navigation · Occupancy Grid Map · Semantic Map · A* Search · ViNT · Mobility VLA

Sources

Mobility VLA: Multimodal Instruction Navigation with Long-Context VLMs and Topological Graphs (arXiv 2407.07775)
ViNT: A Foundation Model for Visual Navigation (arXiv 2306.14846)