Hierarchical Multi-Resolution Spatial Mapping

85 concepts53 papers1 proposal

When an autonomous agent navigates a large, complex environment, it faces a fundamental scalability paradox. To avoid obstacles and execute precise motor commands, the agent needs high-resolution geometric details of its immediate surroundings. However, maintaining a uniform, high-resolution metric map (such as a dense 3D occupancy grid) across an entire building or city causes an exponential explosion in memory and computational overhead. Long-horizon path planning over such uniform grids becomes prohibitively expensive, as information must propagate linearly through millions of individual cells.

To resolve this tension, researchers have converged on hierarchical multi-resolution spatial mapping. Instead of forcing a single map to handle all tasks, this approach splits the representation into multiple layers operating at different spatial and temporal scales. By nesting high-resolution local maps within sparse, low-resolution global frameworks, autonomous systems can achieve real-time local precision alongside highly efficient global planning.

The Core Mechanism: Agent-Centric Non-Uniform Resolution

The foundational intuition behind multi-resolution mapping is that an agent does not need to know the exact millimeter-level geometry of a distant room to decide which direction to walk down a current hallway. Precision is only critical in the immediate vicinity of the agent.

   [ Coarse Global Topology ]  <-- Long-range strategic planning
              |
              v
   [ Egocentric Local Grid ]   <-- High-resolution obstacle avoidance
              |
              v
     ( Immediate Action )

One elegant way to formalize this intuition is through an agent-centric non-uniform representation. By anchoring the highest spatial resolution to the agent's current position and progressively coarsening the resolution for more distant areas, the state space scales logarithmically rather than linearly with the size of the environment.

Daniel Schleich demonstrates this in Value Iteration Networks on Multiple Levels of Abstraction, showing that anchoring fine-grained details to the agent prioritizes immediate motor control while maintaining a computationally lightweight global perspective. In reinforcement learning contexts, this logarithmic state-space scaling enables agents to learn non-myopic paths across massive search areas without becoming overwhelmed by dimensionality, as shown by Sandeep Manjanna et al. in Reinforcement Learning with Non-uniform State Representations for Adaptive Search.

This non-uniform focus can also be implemented geometrically. For example, Hyun Soo Park et al. introduce an "EgoRetinal map" in Egocentric Future Localization, utilizing log-polar coordinates to map immediate obstacles at high resolution while preserving low-resolution context for distant goals within a single unified structure.

Why It Works: Logarithmic Planning Complexity

The primary computational benefit of a multi-resolution hierarchy shows up during path planning. In a standard flat grid, finding a path via algorithms like value iteration requires updating every cell sequentially, scaling linearly ($O(N)$) with the distance to the goal.

Multi-scale hierarchical planning bypasses this limitation by performing value iteration across different map resolutions simultaneously. The system initiates planning on a highly compressed, coarse-grained global map to find a macro-level trajectory, and then refines that trajectory at higher resolutions only within the local neighborhood.

Flat Planning:   [Start] . . . . . . . . . . . . . . . . . . . . . . . . . [Goal]  (Linear Cost)

Hierarchical:   [Start] ----------------- Coarse Path -----------------> [Goal]  (Logarithmic Cost)
                           |
                           v
                        [Local High-Res Refinement]

As demonstrated by Saurabh Gupta et al. in Cognitive Mapping and Planning for Visual Navigation and its subsequent expansion in Cognitive Mapping and Planning for Visual Navigation, this coarse-to-fine refinement reduces the computational complexity of long-range pathfinding to logarithmic time complexity relative to the distance to the goal. Information from a distant target can propagate back to the agent almost instantly through the coarse layer, biasing local actions without requiring a dense sweep of the entire environment.

This mechanism mirrors biological navigation. Dane Corneil explores this in Attractor Network Dynamics Enable Preplay and Rapid Path Planning in Maze-like Environments and Rapid path planning and preplay in maze{like environments using attractor networks, showing that the brain's hippocampus utilizes a hierarchical division of labor. Large-scale place fields act as a low-resolution bridge, allowing top-down goal signals to influence local neural activity across significant distances, while small-scale fields provide the high-precision guidance needed for final goal reaching.

Dual-Scale Architectures in Practice

In robotic implementations, this hierarchy is typically operationalized as a dual-scale memory or mapping architecture. This structure separates high-frequency perception tasks (like avoiding a sudden obstacle) from low-frequency global state tracking (like navigating to a different room).

Several distinct frameworks utilize this dual-scale split:

Scaling to 3D Volumetric Mapping and SLAM

The utility of hierarchical representations extends directly to 3D reconstruction and Simultaneous Localization and Mapping (SLAM). Managing large-scale 3D environments requires data structures that can be dynamically queried at multiple resolutions depending on the immediate task.

Daniel Duberg demonstrates in UFOMap: An Efficient Probabilistic 3D Mapping Framework That Embraces the Unknown that a single volumetric map can support diverse robotic tasks simultaneously by allowing high-speed path planning to query coarse data levels while high-precision reconstruction tasks query the maximum resolution level of the same underlying structure. To organize messy 3D point clouds, Zhenpeng He et al. (Hierarchical Topometric Representation of 3D Robotic Maps) structure space into a three-layer hierarchy—storey, region, and volume—allowing a robot to select a 2D region graph for global pathfinding or 3D volumetric data for local manipulation.

Ultimately, this multi-scale approach provides a robust framework for managing long-term drift in SLAM. As reviewed by Georges Younes et al. in Keyframe-based monocular SLAM: design, survey, and future directions, combining locally metric representations with global topological graphs allows systems to scale to expansive environments without excessive memory growth, ensuring that local tracking remains precise while global loop closures can be executed efficiently.

Key Foundational Work

Integrating Grid-Based and Topological Maps for Mobile Robot Navigation (1996) — contributed 1 concept elements to this direction.
A Multi-Resolution Pyramid for Outdoor Robot Terrain Perception (2004) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hierarchical_spatial_representation_760984a6,
  title = {hierarchical spatial representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-spatial-representation/rdr_760984a6c4ca4f6a85f0b788f46ccfa6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_hierarchical_spatial_representation_760984a6_v1,
  title = {hierarchical spatial representation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hierarchical-spatial-representation/rdr_760984a6c4ca4f6a85f0b788f46ccfa6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake