The Limits of 2.5D Spatial Representations for Navigation

215 concepts173 papers3 proposals

Autonomous navigation has long relied on a convenient fiction: that the world is flat. By projecting a three-dimensional environment onto a two-dimensional Bird's-Eye-View (BEV) grid or a 2.5D heightmap, robots can plan paths with minimal computational overhead. However, this dimensionality reduction introduces severe geometric constraints that cause standard planning algorithms to fail in complex, real-world environments.

The Geometry of the 2.5D Assumption

A 2.5D representation, such as an elevation map, simplifies a 3D scene by assigning a single height value to each coordinate on a horizontal plane. Mathematically, the environment is treated as a function $z = f(x, y)$.

As noted by James F. O’Brien et al., this functional relation makes it impossible to represent structures where material exists at multiple vertical levels simultaneously. While this heuristic is highly efficient for flat warehouses or simple terrains, it introduces critical blind spots when an autonomous agent encounters vertical complexity, such as overhanging obstacles, multi-story buildings, or stacked objects.

Failure Mode 1: Overhangs and Tunnels

Consider a robot attempting to walk under a table, enter a cave, or pass through a tunnel. In a standard 2.5D elevation map, the sensor data from the overhanging structure conflicts directly with the ground data beneath it. Because each grid cell can only retain one elevation value, the map must choose between representing the floor or the ceiling.

As Takahiro Miki et al. demonstrate, this lack of vertical complexity causes planners to misinterpret clear paths under obstacles as impassable vertical walls. Similarly, Jonas Frey et al. point out that traditional elevation maps compress overhead structures into solid blocks, preventing robots from exploiting crawlspaces or navigating safely around common furniture. When a robot's perception is restricted to these height fields, it remains entirely blind to what hangs directly over its path—a limitation highlighted by Takahiro Miki in the context of confined-space locomotion.

Failure Mode 2: Multi-Story Disconnectivity

Multi-story navigation exposes another fundamental flaw: the inability to reason about vertical connectivity. In a 2D top-down projection, a staircase or a ramp does not look like a continuous path leading to a new plane; instead, it appears as a dense cluster of obstacles or a sudden, unnavigable cliff.

Work by Albert J. Zhai et al. shows that navigation strategies relying on 2D semantic maps frequently fail because they misinterpret vertical transitions like staircases as impassable barriers rather than gateways to other floors. Even when global 3D maps are available to an agent, Zibo Zhou et al. observe that local path planners often struggle to execute floor-to-floor transitions because they lack the fine-grained 3D understanding required to coordinate stepping or climbing maneuvers. Because 2.5D models assume a single elevation value per coordinate, Jens-Steffen Gutmann et al. argue that they inherently restrict robots from planning paths through vertically stacked traversable areas, rendering multi-floor structures completely unnavigable.

Failure Mode 3: Erasing Fine Geometry and Vehicle Dynamics

Beyond large structural failures like stairs and ceilings, planar mapping constraints erase the fine-grained geometric details necessary for safe interaction and stable locomotion.

Top-down 2D representations struggle when objects are stacked or placed at varying heights. Xinyu Xu et al. note that this prevents manipulation agents from interacting with items on high shelves or within multi-layered storage units. Furthermore, flattening the environment ignores critical 3D hazards that do not intersect a robot's primary scanning plane. Guo Junfu et al. show that 2D occupancy grids fail to capture complex 3D obstacles like overhanging wires or thin furniture legs, leading to severe structural navigation hazards.

For ground vehicles operating on rugged terrain, Long Xu et al. explain that 2D planning spaces fail to account for the geometric complexities of uneven environments. By ignoring how slopes impact a vehicle's pitch, roll, and center of gravity, planar planners produce trajectories that do not accurately reflect the robot's actual physical pose, risking tip-overs or traction loss.

Breaking Free from the Plane

To overcome these limitations, the field of robotic perception is shifting toward representations that preserve full 3D spatial relationships. As Hardik Shah et al. emphasize, moving beyond 2D bird's-eye-view projections is essential for capturing the structural nuances of complex architectures. By transitioning to full 3D volumetric representations, implicit surfaces, or hierarchical spatial memories, modern frameworks aim to bridge this dimensionality gap, enabling agents to reason about what lies above, below, and beyond the immediate plane of motion.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_planar_mapping_constraints_25bbd6e5,
  title = {planar mapping constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/planar-mapping-constraints/rdr_25bbd6e503cd44deb4f2810c4cab84a6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_planar_mapping_constraints_25bbd6e5_v1,
  title = {planar mapping constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/planar-mapping-constraints/rdr_25bbd6e503cd44deb4f2810c4cab84a6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake