Grid Cell-Inspired Fragmentation and Recall for Efficient Map Building

TL;DR

This work introduces FARMap, a grid cell–inspired framework that fragments space online based on surprisal and recalls previously stored local maps from long-term memory to enable efficient, scalable map-building. By maintaining local maps in STM and linking fragments via an LTM connectivity graph, FARMap generates intrinsic subgoals that guide both local exploration and global coordination, reducing memory and computational costs. Empirical results across procedurally generated environments, dynamic scenes, robot simulations, and Neural SLAM integrations show FARMap outperforms frontier-based exploration in wall-clock time and memory usage while achieving comparable or better map coverage. The approach offers a biologically motivated, memory-efficient alternative for large-scale spatial mapping with potential applicability beyond SLAM to broader streaming and memory-reuse tasks.

Abstract

Animals and robots navigate through environments by building and refining maps of space. These maps enable functions including navigation back to home, planning, search and foraging. Here, we use observations from neuroscience, specifically the observed fragmentation of grid cell map in compartmentalized spaces, to propose and apply the concept of Fragmentation-and-Recall (FARMap) in the mapping of large spaces. Agents solve the mapping problem by building local maps via a surprisal-based clustering of space, which they use to set subgoals for spatial exploration. Agents build and use a local map to predict their observations; high surprisal leads to a "fragmentation event" that truncates the local map. At these events, the recent local map is placed into long-term memory (LTM) and a different local map is initialized. If observations at a fracture point match observations in one of the stored local maps, that map is recalled (and thus reused) from LTM. The fragmentation points induce a natural online clustering of the larger space, forming a set of intrinsic potential subgoals that are stored in LTM as a topological graph. Agents choose their next subgoal from the set of near and far potential subgoals from within the current local map or LTM, respectively. Thus, local maps guide exploration locally, while LTM promotes global exploration. We demonstrate that FARMap replicates the fragmentation points observed in animal studies. We evaluate FARMap on complex procedurally-generated spatial environments and realistic simulations to demonstrate that this mapping strategy much more rapidly covers the environment (number of agent steps and wall clock time) and is more efficient in active memory usage, without loss of performance. https://jd730.github.io/projects/FARMap/

Figures (12)

• Figure 1: (a) Firing fields of grid cells in various environments from derdikman2009fragmentation (top) and carpenter2015grid (bottom). The firing pattern changes at the boundary between two regions (fragmentation). (b) Overview of our approach. Given an observation from the environment, the FARMap agent decides whether to fragment the space based on how well it can predict the observation. If fragmentation occurs, the current map (or model) fragment is stored in long-term memory (LTM); the agent then initializes a new map (or model) fragment. Conversely, if the current observation closely matches the observations stored in LTM, the agent loads an existing map (or model) fragment from there (recall). Based on the current fragment, the agent selects an action to explore the environment.
• Figure 2: Illustration of the FARMap framework. Navigation (black arrow): Given the current observation which is an egocentric top-down view with a restricted field of view and previous action, the agent updates its short-term memory (STM) and selects a subgoal from the current local map in STM or the local map connectivity graph stored in LTM. The planner generates a sequence of actions for the shortest path to the subgoal. Recall (dashed arrow): If the agent arrives at a fracture point (circle in the map), a corresponding local map is recalled from LTM and the current local map stored in LTM is updated. Fragmentation (gray arrow): If the current surprisal is higher than a threshold, the current local map is stored in LTM and a new local map is initialized. $o'_t$ is a spatially transformed observation with the same size as the current local map to update the map.
• Figure 3: Schematic illustrations of how the local map is updated. In this figure, we only consider the visibility of each cell ignoring occupancy and color for simplification. (a) We first rotate the current observation $o_{t,C}$ based on the head direction of the agent in the local map. Then, the observation is zero-padded to match the same size as the local map. Finally, the local map is updated by adding the transformed observation $o'_{t,C}$. (b) If the current observation does not fit within the local map due to the agent's location, we add zero-padding (gray) to both the observation and the local map. Hence, the size of the local map increases ($H$ changes).
• Figure 4: Environments. Empty cells (that can be occupied by the agent) are black; walls are randomly colored. (a) Top-down visualization of the agent's local field of view (FOV) (agent: red triangle; shaded region: observation) within an environment (b). The agent has only a locally restricted egocentric view. The right side is occluded by a wall. (b) Top-down view of one environment. The red box marks the region shown in (a). (c), (d) Examples of medium and large environments.
• Figure 5: Comparison between remapping locations of grid cells in neuroscience experiments derdikman2009fragmentationcarpenter2015grid and fracture points of FARMap in simulation. The red rectangles and emerald circles denote the actual remapping locations and fracture points, and the red triangle is the start location of the simulation. The fracture points are well aligned with the actual remapping locations.