2DGS-Planner: How Rasterization Is Rewriting the Rules of Path Planning on Gaussian Maps

Main Story

Gaussian splatting has rapidly established itself as one of the most compelling scene representations for robotic navigation, valued for its explicit, GPU-accelerated, and efficiently rasterizable structure. Yet the same training mechanics that make it powerful introduce a subtle planning hazard: individual Gaussian primitives may not reliably represent obstacles, as they are jointly optimised through alpha-composited rendering from a finite set of reconstruction views. The consequence is that obstacle geometry inferred primitive-by-primitive can mislead a planner — a wall may appear passable simply because the training camera never looked at it head-on.

A team from KAIST's Urban Robotics Lab — Jiwon Park, Dong-Uk Seo, and Professor Hyun Myung — has tackled this problem head-on with a system called 2DGS-Planner. The framework is a path planner for ground robots that reads planning-relevant geometry from a 2D Gaussian splatting (2DGS) map through rasterization, rather than treating individual Gaussian primitives as obstacles. The key insight is architectural: instead of extracting an intermediate point cloud or mesh from the Gaussian representation, the planner interrogates the composited surface directly through the same rasterization pipeline used for rendering.

The reasoning is grounded in how 2DGS training works. Primitives are jointly optimised through their contributions to alpha-composited renderings, so the training objective constrains the composited geometry rather than any individual primitive — motivating querying the composited geometry instead of treating primitives as independent obstacles. Furthermore, this supervision comes from only a finite set of posed training views, meaning geometry outside their observation support is less constrained by the training data. 2DGS-Planner's design explicitly accounts for both limitations.

The system operates in two phases. During offline roadmap construction, multi-view attribution converts rendered normal dispersion into structural scores for non-ground disks supported by the reconstruction views, and these scores guide adaptive node sampling on the ground. The normal-dispersion signal is available because 2DGS training aligns each disk with the local surface through a normal consistency term, so that surface normals become a rendered channel alongside depth and alpha. This makes normals a reliable proxy for surface structure without requiring any separate geometry extraction.

Edge validation and clearance estimation are handled by two complementary query types. Path-aligned orthographic queries validate candidate edges, while cylindrical queries estimate local clearance fields that are cached on the edges. Caching the clearance fields is a deliberate efficiency choice: it front-loads the heavier rasterization work at map-build time so that online replanning remains fast.

At runtime, the full pipeline — preprocessing and multi-view attribution leading to adaptive sampling, orthographic edge screening, and cylindrical clearance caching — feeds an online A* search that finds an initial route, which is then refined as a B-spline using the cached clearance and revalidated. The refinement step accounts for the physical footprint of the robot throughout.

Experiments demonstrate improved roadmap connectivity, more accurate clearance estimation, and higher planning success compared with the tested baselines, supporting rasterization as an effective geometric query interface for planning directly on Gaussian maps. Experiments were conducted on both public and self-collected outdoor scenes. The team's hardware setup comprised an RTX 4090, Core i9-10900K, and 94 GB RAM, with each edge caching a 64 × 128 clearance field and path refinement using 12 B-spline control points.

The broader context is significant. Research into constraining Gaussian splatting representations for resource-limited deployment — such as the Constrained Dynamic Gaussian Splatting (CDGS) framework (arXiv:2602.03538), which addresses the dilemma of unconstrained densification leading to excessive memory consumption incompatible with edge devices by formulating dynamic scene reconstruction as a budget-constrained optimisation problem — points to a field actively working to make Gaussian representations practical on real robotic hardware. 2DGS-Planner contributes to that trajectory from the planning side.

The paper's contributions include a ground-robot path planner that queries 2DGS maps through rasterization without an intermediate geometric representation, alongside the multi-view attribution scoring and the cylindrical clearance-caching pipeline. Code and data are publicly available at the project page (https://2dgs-planner.github.io/).

Technical Breakdown

Map representation: 2D Gaussian Splatting (2DGS) — flat disk primitives aligned to local surfaces via a normal-consistency regularisation term, rendering depth, normals, and alpha as differentiable channels.

Platform scope: Ground robots with an ellipsoidal body model operating on ground-supported paths; the clearance pipeline explicitly accounts for robot dimensions.

Offline pipeline:

• Multi-view attribution — rendered normal dispersion across reconstruction views is converted into per-disk structural scores that distinguish obstacle surfaces from ground.
• Adaptive node sampling — structural scores guide where roadmap nodes are placed on the ground plane.
• Orthographic edge screening — path-aligned top-down rasterization validates whether candidate roadmap edges are collision-free.
• Cylindrical clearance caching — a 64 × 128 clearance field is rasterized and cached per edge for use during online refinement.

Online pipeline:

• A* graph search initialises a route over the prebuilt roadmap.
• B-spline refinement (12 control points) smooths the route using cached clearance fields while satisfying robot-body and ground constraints.

Autonomy level: Full offline map-build + online replanning; no intermediate mesh or point-cloud conversion required.

Experimental hardware: NVIDIA RTX 4090, Intel Core i9-10900K, 94 GB RAM.

Availability: Open-source code and datasets at https://2dgs-planner.github.io/

Industry Impact

For robotics platform developers and integrators: 2DGS-Planner removes the need for a separate geometry-extraction step (meshing, voxelisation, or point-cloud projection) between a Gaussian reconstruction and a planning module. This tightens the sensor-to-plan pipeline and reduces engineering surface area for ground-robot systems that already use photogrammetric reconstruction workflows — particularly relevant for inspection, logistics, and outdoor survey robots.

For autonomy stack designers: The rasterization-as-query-interface paradigm is architecture-agnostic with respect to the upstream reconstruction tool. Any system that produces a 2DGS map — whether from a drone survey, a ground-based LiDAR-camera rig, or a handheld scanner — can feed directly into the planner. This has meaningful implications for multi-sensor autonomy stacks where map format standardisation is an ongoing challenge.

For the Gaussian splatting research community: The paper validates that planning-relevant signals (surface normals, clearance, structural occupancy) are latent in the rendered outputs of 2DGS maps and can be extracted without retraining or post-processing. This positions rasterization as a general geometric query interface, not just a rendering tool, opening a design space for other perception-and-planning modules built on the same foundation.

For regulators and certification bodies: As Gaussian-splatting-based navigation matures toward deployment, questions of map fidelity, obstacle representation reliability, and planning auditability will become central. The multi-view attribution scoring mechanism in 2DGS-Planner offers a tractable proxy for reconstruction confidence — a property that could inform future requirements around map quality assurance in autonomous ground vehicle standards.

For investors and commercial operators: The open-source release lowers the barrier to adoption and benchmarking. The demonstrated gap between primitive-based obstacle models and composited-surface queries underscores that naive integration of Gaussian splatting into existing planners carries real performance risk — creating differentiated value for teams that invest in representation-aware planning architectures.