Scaling Robotic Navigation from 2D Planes to 3D Volumes

152 concepts149 papers2 proposals

Most autonomous navigation algorithms were born on a flat surface. Whether it is a vacuum cleaner mapping a living room or a self-driving car navigating a highway, the underlying mathematics typically relies on a two-dimensional grid or plane. However, when we transition these algorithms to the sky (UAVs), underwater (AUVs), or multi-level indoor structures (automated warehouses), we encounter a fundamental planar dimensionality constraint.

Moving from 2D to 3D is not a simple matter of adding a $z$-coordinate to an existing equation. It fundamentally alters the topology of space, the geometry of obstacles, and the physics of movement. Understanding why 2D navigation models break down in 3D requires looking at how dimensionality scales across three core axes: geometric complexity, combinatorial explosion, and unconstrained physical dynamics.

The Volumetric Complexity of 3D Obstacles

In a 2D plane, an obstacle is a boundary that completely partitions a path; an agent can only choose to go left or right around it. In a 3D volume, obstacles introduce complex geometries like overhangs, hollow structures, and vertical tunnels.

As highlighted in AerialVLN: Vision-and-Language Navigation for UAVs, Shubo Liu et al. note that aerial agents must navigate vertical and overhanging structures (such as roofs or bridges) where ground-based 2D assumptions fail, causing agents to get trapped under vertical barriers.

2D Obstacle (Wall):          3D Obstacle (Overhang/Bridge):
   [Agent] -> |                 [Agent] ->  ___ (Roof)
              |                            /   \
        (Blocked)                         |     | <- Can fly under or over!

To handle this, safety boundaries must scale from flat buffers to volumetric safety zones. For instance, DUAN Shuyong et al. in Improved A-star Algorithm for Safety Insured Optimal Path with Smoothed Corner Turns point out that traditional grid-based safety distance mechanisms are restricted to 2D planes, failing to account for the multi-axis clearance required by aerial vehicles or robotic arms.

To maintain computational efficiency while scaling up, some approaches decompose 3D safety constraints rather than computing full volumetric intersections. For example, Manan Tayal et al. in Polygonal Cone Control Barrier Functions (PolyC2BF) for safe navigation in cluttered environments project polygonal collision cones onto both horizontal and vertical planes. This decomposition allows controllers to manage 3D safety boundaries without losing the speed of 2D geometric calculations.

The Combinatorial Explosion of Pathfinding and Topology

When you add a third dimension, the search space doesn't just grow—it explodes combinatorially. In multi-agent pathfinding (MAPF), this expansion drastically changes the density of conflicts and the variety of available routes.

In MAPF in 3D Warehouses: Dataset and Analysis, Rishi Veerapaneni et al. demonstrate that while some 2D pathfinding heuristics can generalize, the sheer volume of a 3D warehouse makes specific optimizations like Safe Interval Path Planning (SIPP) exponentially more critical due to the vast number of potential agent interactions across multiple elevations.

Furthermore, the topological math changes. In 2D, obstacles cleanly partition space into distinct "homotopy classes" (e.g., going left vs. right around a pillar). In Topologically distinct trajectory predictions for probabilistic pursuit, Florian Shkurti et al. explain that while 2D tracking can easily group target trajectories using Voronoi diagrams, extending these topological methods to 3D space requires advanced mathematical frameworks like persistent homology to define distinct trajectory groups. Without this, an agent cannot reliably predict or intercept targets moving through a 3D volume because paths can wrap around obstacles in an infinite variety of vertical and horizontal combinations.

Unconstrained Dynamics and Action Spaces

Ground-based robots are constrained by gravity to a surface, meaning their action spaces are inherently planar. Aerial and underwater vehicles, however, operate with expanded degrees of freedom.

In GNM: A General Navigation Model to Drive Any Robot, Ajay Sridhar et al. observe that navigation models trained primarily on ground trajectories struggle with true 3D movement because they rely on fixed-altitude assumptions and 2D action spaces. True autonomy requires models that natively understand verticality and complex spatial orientations.

Operating in a 3D volume also exposes the agent to unconstrained physical forces. As noted in the framework presented in Autonomous Advanced Aerial Mobility - An End-to-end Autonomy Framework for UAVs and Beyond, navigating a 3D airspace requires managing twice the degrees of freedom while responding to unpredictable forces like wind in milliseconds.

Similarly, in underwater domains, Tom O’Reilly et al. in Dynamic robotic tracking of underwater targets using reinforcement learning emphasize that tracking algorithms must move past the simplifying assumption of a known target depth, expanding into full 3D path planning to handle ocean currents and signal dropouts effectively.

The Multi-Level Discontinuity

Even when navigation occurs within human-built structures, treating multi-story environments as stacked 2D planes creates severe blind spots.

In Status and Problems in Location Model Research for Guidance Signage in Comprehensive Passenger Transportation Hubs, the authors argue that traditional indoor navigation models fail because they treat multi-story transit hubs as flat layouts, ignoring how vertical transitions like escalators and stairs constrain visibility and crowd movement.

This discontinuity is echoed in behavioral modeling; as discussed in Imitation of Human Behavior in 3D-Shooter game, autonomous agents often fail to maintain realism because they cannot effectively identify threats or navigate across multiple elevations simultaneously.

Ultimately, overcoming the planar dimensionality constraint requires a paradigm shift: treating the third dimension not as an optional parameter, but as a fundamental property of the environment's topology and physics.

Key Foundational Work

Intelligent Agents in Computer Games (1999) — contributed 2 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_3d_spatial_navigation_8805d371,
  title = {3D spatial navigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/3d-spatial-navigation/rdr_8805d371570c40d48a8b263643b9e0d6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_3d_spatial_navigation_8805d371_v1,
  title = {3D spatial navigation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/3d-spatial-navigation/rdr_8805d371570c40d48a8b263643b9e0d6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake