PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Jiwon Park, Dong-Uk Seo, Hyun Myung

arXiv:2610.11752v1cs.RO

TL;DR

Individual Gaussian primitives may not reliably represent obstacles because training jointly optimizes alpha-composited renderings from a finite set of reconstruction views. 2DGS-Planner queries rasterized 2DGS geometry to construct and refine ground-robot paths. It achieves the highest SPL on both scenes, while its clearance fields yield lower error and fewer false-blocked configurations than the tested ESDFs.

  • Problem

    Individual Gaussian primitives may not reliably represent scene geometry because optimization constrains alpha-composited renderings from only finite reconstruction views.

    Source: §I · p. 1

  • Method

    2DGS-Planner uses multi-view normal dispersion to guide roadmap sampling and path-aligned rasterization queries to validate edges and cache clearance for refinement.

    Source: §I · pp. 1–2, D. Roadmap Construction · p. 4, E. Path-Aligned Motion Queries · p. 4

  • Results

    2DGS-Planner achieves the highest SPL on both scenes, and LCC yields lower clearance error and fewer false-blocked configurations than the tested ESDFs.

    Source: F. End-to-End Planning · p. 6, E. Clearance Query Fidelity · p. 6

  • Takeaways & Limitations

    The results support rasterization as a practical geometric query interface for planning directly on Gaussian scene representations.

    Source: §VI · pp. 7–8

    The current formulation approximates the robot body with an ellipsoid and does not explicitly account for robot dynamics.

    Source: §VI · p. 8

Abstract

from arXiv · show

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/

I. INTRODUCTION

Individual Gaussian primitives may not reliably represent scene geometry because training constrains alpha-composited renderings and covers only finite reconstruction views. 2DGS-Planner instead queries rasterized map geometry for structure-aware roadmap construction, edge validation, and clearance-aware path refinement.

I. INTRODUCTION

  • Individual primitives may misalign with the alpha-composited surface because training jointly optimizes their contributions and uses only finite reconstruction views.The planner therefore considers both rendered surface geometry and how well it was observed during training.Source: §I · p. 1
  • 2DGS-Planner uses rasterization to query planning-relevant geometry directly from 2DGS maps, without an intermediate geometric representation.Its pipeline includes multi-view attribution, path-aligned motion queries, roadmap construction, and path refinement.Source: §I · pp. 1–2
  • Multi-view attribution converts rendered normal dispersion into structural scores that guide adaptive node sampling on the ground.Sampling places more nodes near corners and cluttered areas and fewer in open regions.Source: §I · p. 2
  • The framework evaluates structure, clearance estimation, end-to-end planning, and real-world robot experiments.Orthographic queries validate candidate connections, while cylindrical queries estimate local clearance fields cached for refinement.Source: §I · p. 2

II. RELATED WORK

Prior work uses Gaussian splatting for rasterized geometry and for motion planning, but commonly derives collision information from primitives or auxiliary geometric representations. This work uses rasterization as a geometric query interface for planning on a fixed 2DGS map.

A. Rasterization-based Geometry in Gaussian Splatting

  • Rasterization methods have used rendered quantities for reconstruction, while 2DGS-Planner uses them as geometric queries for path planning on a fixed map.Related examples render plane distances and normals, derive depth and normals, or support secondary rays.Source: A. Rasterization-based Geometry in… · p. 2

B. Planning on Gaussian Splatting

  • Prior Gaussian-splatting planners derive collision information from primitives, including safe corridors, barrier functions, collision-probability bounds, and robot-scene overlap.The cited methods include Splat-Nav, SAFER-Splat [16], SPLANNING, and FOCI.Source: B. Planning on Gaussian Splatting · p. 2
  • Other planners use auxiliary distance representations or rendered depth for reactive collision avoidance, unlike this planner’s path-organized queries and cached clearance.Examples include ESDF, TSDF–GS, and continuous distance-field approaches; ReMoSPLAT [21] targets reactive avoidance.Source: B. Planning on Gaussian Splatting · p. 2

III. PRELIMINARIES

A 2DGS map is queried by casting rays against planar Gaussian disks and alpha-compositing their contributions. The overview connects rasterized geometry to planning queries, while treating unobserved rays as lacking geometric evidence rather than as free space.

III. PRELIMINARIES

  • A 2DGS map contains oriented planar Gaussian disks reconstructed from posed images, and its rasterization evaluates ray–disk intersections in world space.The intersections provide Gaussian responses and hit distances, whose contributions are alpha-composited along each ray.Source: §III · p. 2
  • Front-to-back alpha composition determines each disk’s contribution and the accumulated opacity for a query sample.For positive accumulated opacity, the disk contribution can be normalized.Source: §III · p. 2
  • Unobserved rays provide no geometric evidence and are not treated as free space.The overview places rasterized geometry in a pipeline of preprocessing, attribution, edge queries, clearance caching, and route refinement.Source: §III · p. 3, Figure 3 · p. 3

IV. 2DGS-PLANNER

The planner seeks a ground-supported path for an ellipsoidal robot on a static 2DGS map. It preprocesses the map to identify supporting ground and align planning with gravity, then builds and refines a route using rasterized queries.

A. Problem Formulation and Overview

  • The planning problem is to connect start and goal locations on observed ground while maintaining robot-body clearance and ground support.The robot is modeled as a gravity-aligned ellipsoid with prescribed center height above the supporting ground.Source: A. Problem Formulation and Overview · p. 3
  • An offline roadmap uses multi-view structural cues for sampling and path-aligned queries to validate edges and cache clearance for online route refinement.A* initializes the graph route, which is refined under body-clearance and ground constraints.Source: A. Problem Formulation and Overview · p. 3

B. Map Preprocessing

  • Preprocessing transfers semantic predictions from reconstruction views to disks, identifies ground regions, and aligns subsequent planning to gravity.Ground-labeled disks meeting orientation and height criteria define the support surface for roadmap sampling and path refinement.Source: B. Map Preprocessing · p. 3

C. Multi-View Attribution

Multi-view attribution turns variation in rendered surface normals into structural scores for non-ground disks, which then guide sampling density on nearby ground candidates.

C. Multi-View Attribution

  • Dispersion rises where multiple surface orientations contribute to one pixel, including corners, thin structures, and cluttered regions.The contribution-weighted mean normal is not renormalized, so its magnitude reflects agreement among contributing normals.Source: C. Multi-View Attribution · p. 3
  • Rendered normal dispersion is aggregated across reconstruction views using each disk’s alpha-compositing contribution.Views where a disk contributes more strongly have greater influence on its attributed dispersion.Source: C. Multi-View Attribution · p. 3
  • Eligible non-ground disks receive normalized structural scores based on their empirical dispersion percentile; ineligible disks score zero.Eligibility requires sufficient rendering contribution and support from a minimum number of reconstruction views.Source: C. Multi-View Attribution · p. 4
  • Ground disks are excluded from the ranking, and the resulting scores guide sampling density for nearby ground candidates.Excluding ground disks prevents their numerous low-dispersion values from biasing non-ground scores upward.Source: C. Multi-View Attribution · pp. 3–4

D. Roadmap Construction

The roadmap is built from ground-supported candidate positions, with adaptive sampling guided by local structural complexity and path-aligned queries used to screen and characterize candidate edges.

D. Roadmap Construction

  • Candidate positions sampled from ground-disk centers are lifted to the prescribed ride height and assigned local complexity scores.The structural scores determine denser sampling in complex regions and sparser sampling in open regions.Source: D. Roadmap Construction · p. 4, Figure 4 · p. 4
  • Percentile-ranked complexity scores set target node spacing between ℓmax and ℓmin; candidates without nearby non-ground disks use ℓmax.Adaptive Poisson-disk thinning produces the final node set.Source: D. Roadmap Construction · p. 4
  • A Gabriel graph proposes connections, which are retained only after orthographic screening and cylindrical clearance estimation.Retained edges store their clearance fields for online path refinement.Source: D. Roadmap Construction · p. 4

E. Path-Aligned Motion Queries

  • Path-aligned orthographic queries screen translation between nodes, but do not cover longitudinal overhang or in-place rotations.The footprint test rejects an edge if a sampled ray observes a surface before reaching the endpoint.Source: E. Path-Aligned Motion Queries · p. 4
  • For edges passing the orthographic screen, cylindrical queries record local clearance by axial position and azimuth; slices with inadequate observed clearance reject the edge.The clearance field is stored on each retained edge for later queries during path refinement.Source: E. Path-Aligned Motion Queries · p. 4

F. Path Optimization

Path optimization initializes a route from the roadmap, reuses cached local clearance fields during refinement, and revalidates the refined path against the map.

F. Path Optimization

  • A* connects the start and goal to nearby roadmap nodes and supplies a graph route for B-spline initialization, with endpoints fixed during refinement.The start and goal connections use the same path-aligned queries.Source: F. Path Optimization · p. 4
  • Cached local cylindrical fields are back-projected into surface samples and reused during refinement without additional rendering.Clearance is evaluated at each current refined-path sample against surface samples associated with its route edge.Source: F. Path Optimization · p. 4, Figure 5 · p. 5
  • Required clearance increases with local path curvature to account for the additional space an elongated body needs while turning.The refined-path clearance is evaluated at the current path position, rather than radially from the original edge axis.Source: F. Path Optimization · pp. 4–5
  • Refinement balances path regularization against clearance penalties while constraining control-point heights to the prescribed ride height.The path is revalidated with the orthographic query, and failed spans fall back to validated graph segments.Source: F. Path Optimization · p. 5

V. EXPERIMENTS

The experiments assess roadmap construction, query fidelity, and end-to-end planning across two reconstructed scenes, using sampling-based and ESDF-based comparisons.

V. EXPERIMENTS

  • The evaluation covers structural attribution and adaptive sampling, cached-clearance fidelity, end-to-end planning, robot-dimension effects, and a physical humanoid demonstration.End-to-end planning experiments include ablations of adaptive sampling and LCC-based refinement.Source: §V · p. 5

A. Evaluation Environments

  • Stonehenge supplies an exact synthetic reference mesh, while Tokyo supplies an aligned proxy mesh for a partially observed real-world reconstruction.Both scenes are used for the attribution, clearance, and end-to-end planning experiments.Source: A. Evaluation Environments · p. 5

B. Baselines

  • Comparisons include Disk PRM, RRT*, Informed RRT*, and a ground-constrained ESDF planner operating on the same 2DGS reconstruction.Uniform Disk PRM roadmaps use budgets of 1k, 10k, and 60k nodes; the ESDF baseline plans through an intermediate volumetric representation.Source: B. Baselines · p. 5

C. Implementation Details

The experiments specify a shared implementation setup, evaluate rendered attribution as a robot-scale sampling signal, and compare cached LCC clearance fields with ESDFs. Attribution performs best for node placement at tested local scales, while LCC reduces clearance error and false-blocked cases but does not guarantee collision freedom.

C. Implementation Details

  • In the experiments, the robot is modeled as a sphere with radius r = 0.05 in the nonmetric Stonehenge map and r = 0.50 m in the metric Tokyo map; the roadmap uses a 64×128 cylindrical clearance field per edge.Structural attribution retains disks with support τS = 10−3 from at least 3 views. Path refinement uses a B-spline with 12 control points and revalidates the resulting path using the orthographic query.Source: C. Implementation Details · p. 5

D. Structure-Aware Roadmap Construction

  • Rendered attribution achieves the highest ρs and ROC-AUC for aggregated node-placement scores at Rref = r and 2r on both scenes, despite lower disk-level ranking on Stonehenge.The relevant sampling quantity is the aggregated score u(q), with R = 3r; rankings become less consistent at larger scales.Source: D. Structure-Aware Roadmap Construction · p. 6

E. Clearance Query Fidelity

  • LCC yields lower clearance error and fewer false-blocked configurations than tested ESDFs, and its construction is faster than that of finer ESDF grids.The comparison uses identical rendered surface samples and focuses on near-boundary configurations; false-safe cases remain, so clearance estimates do not guarantee collision freedom.Source: E. Clearance Query Fidelity · p. 6, Figure 7 · p. 6

F. End-to-End Planning

Across 50 start–goal pairs per scene, 2DGS-Planner achieves the highest SPL on Stonehenge and Tokyo, with stronger Tokyo results against the tested baselines. Its refinement improves SPL but adds online computation, while larger uniform roadmaps incur greater construction cost with diminishing gains.

F. End-to-End Planning

  • 2DGS-Planner achieves the highest SPL on both scenes; on Tokyo it improves collision-pass counts and SPL over primitive-based and ground-constrained ESDF baselines.On Stonehenge, collision-pass counts are comparable to the largest Disk PRM roadmap while SPL is higher; Fig. 8 shows representative paths.Source: F. End-to-End Planning · p. 6, Figure 8 · p. 7
  • Larger uniform Disk PRM roadmaps increase construction cost while SPL gains diminish or saturate, and the full planner achieves higher SPL with lower construction time than the largest tested roadmap on both scenes.Additional uniform nodes do not consistently improve strict collision-pass rates within the tested range; Fig. 9 compares budgets of 1k, 10k, and 60k nodes.Source: F. End-to-End Planning · pp. 6–7, Figure 9 · p. 7
  • LCC-based refinement improves SPL on both scenes, increasing collision-pass counts on Stonehenge and maintaining comparable counts on Tokyo without modifying the roadmap.The ablation uses the same roadmap and edge validation but omits path refinement.Source: F. End-to-End Planning · p. 7
  • Path refinement increases online computation, but the full planner remains faster than RRT* on both scenes; on Tokyo it is slower than Disk PRM and ESDF.Latency is reported per request, including failures, separately from offline roadmap construction.Source: F. End-to-End Planning · p. 7

G. Robot-Size-Aware Planning

Robot-size-aware validation changes which passages are feasible, and the paper also demonstrates physical execution of a refined path by a humanoid robot. The conclusion notes that the current robot model and planning formulation have limits in representing detailed geometry and dynamics.

G. Robot-Size-Aware Planning

  • With lateral half-width b = 0.3 m, the planner selects the wider passage; at b = 0.1 m, the narrower passage becomes feasible and yields a shorter path.Candidate edges and the refined path are validated using the corresponding robot-specific orthographic footprint query.Source: G. Robot-Size-Aware Planning · p. 7, Figure 11 · p. 7

H. Real-World Humanoid Deployment

  • A humanoid robot successfully executes the refined path in the real-world environment after the reconstructed 2DGS map is registered with a LiDAR map.The robot tracks the path using a pure-pursuit controller; Fig. 12 shows the map path and physical execution.Source: H. Real-World Humanoid Deployment · p. 7, Figure 12 · p. 8

VI. CONCLUSION

  • The current formulation approximates the robot body with an ellipsoid, limiting detailed and articulated geometry, and does not explicitly account for robot dynamics.The conclusion identifies more detailed body models and kinematic and dynamic feasibility as future work.Source: §VI · p. 8
Loading 2610.11752v1…