GoLF can serve as the map a planner uses directly. In the paper's experiments, that shift changes the scale of the planning problem: planners search over a compact spatial structure instead of a dense image grid. The result is much faster planning across several kinds of environments, with only small route length tradeoffs in the reported benchmarks. The practical value is not just speed. Faster planning makes it easier to run larger sweeps, test policies across many floor plans, and study navigation behavior without letting map resolution dominate the question.

GoLF can also support pedestrian tracking, where the important question is not only where each person is now, but how motion unfolds through the surrounding space. The teaser below shows this idea in a shared indoor scene: people move through a structured environment while the representation keeps their motion tied to the places, passages, and nearby activity that give it context. This is useful for robotics because a robot that shares space with people needs more than isolated detections. It needs a compact view of how people are moving, where interactions may form, and which parts of the scene deserve attention as the situation changes.

Shared spaces are not only geometric. Doorways, hallways, and bottlenecks carry social meaning because people move through them, pause in them, and negotiate right of way around them. GoLF gives a learning-based navigation system a structured way to relate human motion to the surrounding map context. In simulated and real-world demonstrations, the GoLF-based navigation setup supports behavior that is more than collision avoidance. The robot can follow a route when the way is clear, slow down near a constrained passage, yield to a passing person, and continue after the interaction becomes appropriate again.

GoLF can also help describe manipulation, where a robot has to reason about objects, hands, surfaces, and contact at the same time. In the teaser below, the representation follows a robot arm, gripper, table, and blocks as separate but connected parts of the scene, with links appearing when objects come close enough to interact. This matters because manipulation is not only about where objects are; it is about how relationships between objects change as a robot acts. A compact, structured view of those relationships could make it easier to inspect robot behavior, compare manipulation attempts, and eventually build systems that understand when an action is starting to matter physically. This is early, unpublished work, so we present it here as an application teaser rather than a technical result.

Although this project page focuses on robot navigation, the premise behind GoLF is broader. Many physical quantities are uneven: some regions are smooth and predictable, while others change sharply and deserve more attention. GoLF is a way to represent that unevenness without forcing every part of the field into the same resolution. That makes the idea relevant beyond floor plans. The paper frames GoLF as a representation for physical fields such as occupancy, motion, temperature, airflow, and related spatial quantities, with navigation as the central case study.