2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map

Jiwon Park; Dong-Uk Seo; Hyun Myung

cs.ROOct 8, 2026

2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map

Authors: Jiwon Park, Dong-Uk Seo, Hyun Myung

Abstract

Gaussian splatting provides an explicit and efficiently rasterizable scene representation for robot navigation. However, individual Gaussian primitives may not reliably represent obstacles as they are jointly optimized through alpha-composited rendering from a finite set of reconstruction views. We propose 2DGS-Planner, 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. During offline roadmap construction, multi-view attribution converts rendered normal dispersion into structural scores for non-ground disks supported by the reconstruction views. These scores guide adaptive node sampling on the ground. Path-aligned orthographic queries validate candidate edges, while cylindrical queries estimate local clearance fields that are cached on the edges. During online planning, graph search initializes a route, and path refinement reuses the cached fields while accounting for the robot's dimensions and ground constraints. Experiments demonstrate improved roadmap connectivity, more accurate clearance estimation, and higher planning success compared with the tested baselines. These results support rasterization as an effective geometric query interface for planning directly on Gaussian maps. Code and data are available at https://2dgs-planner.github.io/

Figures & tables

Fig. 1 : 2DGS-Planner uses rasterization to read surface structure and local clearance directly from a 2D Gaussian splatting map. Multi-view normal dispersion guides structure-aware roadmap construction, while path-aligned rasterization queries validate candidate motions and provide local clearance for path refinement. The resulting roadmap and refined path are shown in the reconstructed scene.

Fig. 2 : Geometric discrepancy between individual 2DGS primitives and the alpha-composited surface. (a) RGB rendering. (b) Underlying Gaussian disks visualized at their $1\sigma$ extent. (c) Disks with large depth discrepancies from the rendered surface, highlighted in red. Visually consistent rendering can therefore coexist with inaccurate primitive-level geometry.

Fig. 3 : Overview of 2DGS-Planner. Roadmap construction (top): Given a 2DGS map and its reconstruction views, (a) semantic preprocessing identifies supporting ground and aligns the map with gravity. (b)–(c) Multi-view attribution aggregates contribution-weighted per-pixel normal dispersion across reconstruction views to obtain structural scores $\Sigma_{i}$ for observation-supported non-ground disks. (d) The scores guide adaptive node placement. (e) Orthographic queries screen candidate edges, and (f) cylindrical queries estimate surrounding clearance and cache the resulting fields on retained edges. Online planning (bottom): Roadmap search using $A^{*}$ initializes a route, which is refined as a B-spline using surface samples back-projected from the cached clearance fields while accounting for robot dimensions and ground constraints. The final panel shows the refined path and its corresponding gravity-aligned body poses, with headings following the path tangents.

Fig. 4 : Structure-aware roadmap sampling. (a) RGB rendering. (b) Structural score produced by multi-view attribution. (c) Uniform sampling with its resulting roadmap. (d) Adaptive sampling using the structural scores, which concentrates nodes near locally complex structures while reducing sampling in open regions under the same node budget.

Fig. 5 : Local cylindrical clearance (LCC) representation. (a) For an edge $e=(\mathbf{x}_{1},\mathbf{x}_{2})$ , radial rasterization queries are cast from axial position $s$ in the edge-aligned frame $\{\mathbf{t}_{e},\mathbf{b}_{1},\mathbf{b}_{2}\}$ . The resulting distance $\rho_{e}(s,\theta)$ measures radial clearance to the observed surface at azimuth $\theta$ . (b) Unrolling the measurements over $(s,\theta)$ forms the LCC field, which is cached on the edge and reused during path refinement. Colors encode radial clearance.

Fig. 6 : Correlation and discrimination performance of structural scores with respect to reference-mesh normal dispersion on Stonehenge (exact mesh, left) and Tokyo (proxy mesh, right). The top and bottom rows report Spearman correlation

\rho_{s}

and ROC-AUC, respectively. Disk-level results compare rendered attribution with geometric neighborhood scores, while sampler-input results evaluate the aggregated node score

u(\mathbf{q})

over reference radii

r

–

5r

. For Tokyo , “Bldg.” denotes the building subset. Dotted horizontal lines indicate chance-level ROC-AUC.

Fig. 7 : Clearance fidelity and representation cost on Stonehenge . LCC fields and ESDFs are constructed from the same rendered 2DGS surface samples and evaluated on 5,716 near-boundary queries satisfying

\lvert c_{\mathrm{GT}}-r\rvert<5\,\mathrm{cm}

along 30 paths. Clearance error is summarized by the mean (diamond) and median–95th percentile interval (bar); the remaining panels report false-safe (FS) and false-blocked (FB) rates, memory usage, and construction time. Error, memory, and time are shown on logarithmic scales.

	Stonehenge			Tokyo
Method	$\varepsilon{=}\varepsilon_{\mathrm{ref}}\uparrow$	$\varepsilon{=}0\uparrow$	SPL $\uparrow$	$\varepsilon{=}\varepsilon_{\mathrm{ref}}\uparrow$	$\varepsilon{=}0\uparrow$	SPL $\uparrow$
Disk PRM ( $1$ k)	21/50	15/50	0.335	6/50	6/50	0.091
Disk PRM ( $10$ k)	43/50	29/50	0.746	10/50	10/50	0.176
Disk PRM ( $60$ k)	45/50	27/50	0.810	10/50	9/50	0.168
RRT ∗	35/50	24/50	0.663	10/50	8/50	0.180
Informed RRT ∗	33/50	24/50	0.650	13/50	10/50	0.248

TABLE I : Reference-geometry evaluation on

50

start–goal pairs per scene. Collision pass reports paths satisfying the reference check at tolerance

\varepsilon

, while SPL evaluates successes at

\varepsilon=\varepsilon_{\mathrm{ref}}

using

\min(1,L_{\mathrm{ref}}/L)

, where

L_{\mathrm{ref}}

and

L

are the reference and returned path lengths. We use

\varepsilon_{\mathrm{ref}}=0.005

for Stonehenge and

0.10\,\mathrm{m}

for Tokyo .

Fig. 8 : Qualitative planning results on Stonehenge (top) and Tokyo (bottom) for four representative start–goal pairs from Table I . Each panel overlays the evaluation reference path in yellow with the paths returned by the compared planners; missing trajectories indicate planning failures. Circles and stars denote the start and goal, respectively. For both scenes, the camera symbol indicates the secondary viewpoint shown in the inset. Across these examples, 2DGS-Planner more consistently returns feasible paths, and its trajectories tend to follow the reference path more closely than those of the compared methods.

Fig. 9 : Effect of roadmap node budget on construction cost and planning performance. Roadmap build time (left) and SPL (right) are shown for uniform Disk PRM with 1k, 10k, and 60k nodes and for 2DGS-Planner. Stonehenge and Tokyo are indicated by solid and hatched bars, respectively. Increasing the uniform sampling budget substantially increases construction time while providing diminishing or saturated SPL gains.

Fig. 10 : Per-request online planning latency for Stonehenge and Tokyo , averaged over

50

requests including failures and reported separately from offline roadmap construction. LCC-based path refinement increases online computation, while the full 2DGS-Planner remains faster than RRT ∗ on both scenes. Colors follow the method palette.

Fig. 11 : Robot-size-aware planning through footprint-specific edge validation. (a) Initial graph route (red), refined path (green), and robot body with lateral half-width $b=0.3\,\mathrm{m}$ ; the narrower passage is infeasible, leading to the wider route. (b) With $b=0.1\,\mathrm{m}$ , the narrower passage becomes feasible and yields a shorter path. Candidate edges and the refined path are validated using the corresponding orthographic robot footprint.

Fig. 12 : Real-world deployment of 2DGS-Planner on a humanoid robot. (a) Initial graph route (red), refined path (green), and corresponding robot body poses in the reconstructed 2DGS map. (b) Autonomous execution of the refined path in the corresponding real-world environment.

Explore similar work

Date pendingcs.RO

ATLAS Navigator: Active Task-driven LAnguage-embedded Gaussian Splatting

We address the challenge of task-oriented navigation in unstructured and unknown environments, where robots must incrementally build and reason on informative metric-semantic maps in real time. Since tasks may require clarification or re-specification, it is necessary for the information in the map to be rich enough to enable generalization across a wide range of tasks. To effectively execute tasks specified in natural language, we propose a hierarchical representation built on language-embedded Gaussian splatting that enables both sparse semantic planning that lends itself to online operation and dense geometric representation for collision-free navigation. We validate the effectiveness of our method through real-world robot experiments on a ground robot equipped with only vision-based sensing. The experiments are conducted in both cluttered indoor and kilometer-scale outdoor environments, covering more than 2 km and 47,000 square meters of area. We achieve a competitive ratio of about 59% relative to the shortest possible path, demonstrating efficient exploration and semantic navigation in challenging real-world environments. Experiment videos and more details can be found on our project page: https://ongdexter.github.io/atlasnav.

Dexter Ong, Yuezhan Tao, Varun Murali +3

GRASP Laboratory, University of Pennsylvania, Philadelphia, PA 19104 USA · Texas A&M University, College Station, TX 77840 USA · University of California Riverside, Riverside, CA 92521 USA

3D Gaussian Splatting Language-Embedded 3DGS Robot Navigation Vision-Language Navigation Semantic Mapping

Ask PDF

Sep 28, 2026cs.RO

CollisionSplatting: Collision-Aware Motion Planning in 3DGS Scenes with Image-Conditioned Objectives and Adjustable Conservatism

Incorporating dense visual information into motion planning remains challenging, as geometric planners rely on abstracted scene representations that discard visual richness, while learned visual models often lack geometric interpretability and computational efficiency. This paper introduces CollisionSplatting, a simple, modular, GPU-accelerated, probability-inspired distance metric with tunable conservatism that operates directly on standard 3D Gaussian Splatting (3DGS) scenes. When combined with learned image-conditioned reward functions, this metric enables joint geometric and visual planning by unifying collision-aware costs with image-space objectives. We integrate the metric into GPU-accelerated Model Predictive Path Integral (MPPI) and Rapidly-Exploring Random Tree (RRT) planners, and show on-par or better collision-classification performance compared to representative baselines while achieving substantially higher collision-checking throughput and significantly lower VRAM usage. Finally, we demonstrate the effectiveness of our metric in real-world vision-guided navigation and manipulation tasks, highlighting 3DGS as a practical bridge between rich perception and real-time motion planning.

R. Khorrambakht, Joaquim Ortiz-Haro, Stephan Weiss +1

Machines in Motion Lab, Tandon School of Engineering, New York University · Department of Smart Systems Technologies, University of Klagenfurt, Austria. · Artificial and Natural Intelligence Toulouse Institute (ANITI)

3D Gaussian Splatting Model-Based Planning Safe Motion Planning Motion Planning Sampling-Based Motion Planning

Ask PDF

Jun 12, 2026cs.RO

SplatlessDF: Continuous Distance Field Mapping with Non-Splatting Gaussians

Recent Gaussian splatting (GS) methods have shown that scenes can be represented efficiently with optimisable Gaussians for high-quality reconstruction and rendering. In this paper, building on this principle, we introduce SplatlessDF, a continuous distance field (DF) mapping framework that uses anisotropic Gaussian elements from a spatial rather than photometric perspective. SplatlessDF directly parameterises the Gaussians and optimises to recover a differentiable DF, enabling distances and gradients to be queried in the spatial domain for downstream robotic tasks such as navigation. Furthermore, SplatlessDF can be coupled with 2D Gaussian splatting (2DGS), providing a unified framework based solely on Gaussian primitives that can learn continuous DF and surface models and supports photometric rendering. We consider two settings: a standalone DF-only formulation and a joint DF-rendering formulation coupled with 2DGS. Experiments show that the standalone formulation provides efficient and accurate distance and gradient queries, while the joint formulation improves rendering geometry and simultaneously models a continuous DF. These results highlight the potential of GS-style representations not only for surface modelling and rendering but also for mapping representations suited to robotic navigation.

Monisha Mushtary Uttsha, Lan Wu, Teresa Vidal-Calleja

UTS Robotics Institute, Faculty of Engineering and IT, University of Technology Sydney, Sydney, NSW 2007, Australia · School of Engineering, University of Western Australia, Perth, WA 6009, Australia

3D Gaussian Splatting 3D Reconstruction 3D Scene Representation

Ask PDF