Source-linked AI summary
GS-PRM: Gaussian Visual Roadmap for Image-Goal Navigation
Ashraf Siddique, Kyung Min Han, Young J. Kim
TL;DR
Image-goal navigation needs a global route connecting visually specified endpoints while providing observations a local policy can follow. GS-PRM constructs a roadmap and its visual references directly from 3DGS renderings, with render-based criteria for configurations and paths. The paper gives conditional soundness and completeness guarantees and reports a validation experiment with no false-negative RC-free classifications.
Problem
Image-goal navigation needs persistent topology connecting localized image queries into reachable visual subgoals, together with map-registered observations for repeated pose correction.
Source: §I · p. 1Method
GS-PRM builds a coverage-driven visual roadmap from 3DGS renderings and defines RC-free configurations and paths using conservative depth evidence.
Source: §I · p. 1Results
The authors state conditional soundness and sufficient conditions for deterministic query completeness, and report no false-negative RC-free classifications in their validation experiment.
Source: C. Collision-free Conditions · p. 3, D. Completeness Guarantees · p. 5, 1) Validation of RC-free Condition · p. 6Takeaways & Limitations
The roadmap supports image-defined queries and rendered visual subgoals, with query completeness once certified coverage and connectivity hold.
Source: D. Completeness Guarantees · p. 5, §VII · p. 8
Abstract
from arXiv · showhide
Image-goal navigation requires a global planner that connects visually specified endpoints while providing observations that a local policy can follow. Three-dimensional Gaussian Splatting (3DGS) provides map-registered RGB, opacity, and depth renderings, but does not explicitly encode navigable topology. We introduce GS-PRM, a Gaussian visual roadmap whose milestones, local connections, and visual references are constructed directly through 3DGS rendering without deriving an occupancy grid or signed distance field. We define render-based collision-free (RC-free) configurations and paths. Panoramic depth supplies conservative clearance for milestone validation and adaptive coverage, while edge-aligned orthographic rendering tests the robot's swept volume. We establish conditional soundness and sufficient conditions for deterministic query completeness and probabilistic construction completeness of GS-PRM. To validate the performance of GS-PRM, we conduct experiments in both simulated and real-world environments, demonstrating long-horizon navigation using goal-conditioned policies guided by GS-PRM. Compared with a state-of-the-art 3DGS-based navigation system, our GS-PRM-based navigation system achieves up to a twice higher average success rate and reduces travel time by up to 57.2% when both systems complete the navigation task.
I. INTRODUCTION
Image-goal navigation needs persistent topology connecting image-localized endpoints and map-registered observations for repeated pose correction. GS-PRM builds a coverage-driven visual roadmap using 3DGS rendering, and experiments report improved long-horizon navigation outcomes.
I. INTRODUCTION
- Image-goal navigation needs persistent topology connecting image-localized queries into reachable visual subgoals, plus map-registered observations for pose correction against odometric drift.The goal image conveys appearance but neither a goal pose nor a traversable route, and the destination is usually outside the current camera view.Source: §I · p. 1
- Conventional PRMs assume workspace geometry for collision checking, which can require costly tests and may be invalidated by reconstruction or localization errors; they also lack visual feedback observations.3DGS offers map-registered appearance and fast RGB, opacity, and depth rendering, but its raw primitives do not directly represent physical obstacles.Source: §I · p. 1
- GS-PRM uses one 3DGS rendering interface for roadmap validity, image-defined endpoints, and visual route references, without deriving an occupancy or distance-field planning map.Its clearance-adaptive coverage reduces redundant milestones in open areas while retaining finer resolution near constrained regions.Source: §I · p. 1
- RC-free milestones and local connections use conservative, thresholded depth, with stated assumptions linking accepted configurations and paths to physical free space.The work also establishes sufficient conditions for probabilistic construction completeness and deterministic query completeness of a completed roadmap.Source: §I · p. 1
- GS-PRM with straight-line planning achieves up to twice the success rate and reduces travel time by 57.2% versus SplatNav when both complete navigation.DevGRU with GS-PRM reaches 50% success, versus 0% for standalone DevGRU.Source: §I · p. 2
A. Probabilistic Roadmaps for Navigation
Prior navigation approaches use geometric maps, Gaussian primitives, or rendered visual goals, while GS-PRM defines a reusable roadmap through rendering. Its roadmap connects image queries and supplies observations as localization references and policy subgoals.
A. Probabilistic Roadmaps for Navigation
- Classical PRMs build reusable graphs from sampled collision-free milestones and locally feasible paths, while navigation variants alter sampling, account for dynamic availability, or use learned local policies.These approaches generally rely on a geometric or state representation for collision checking.Source: A. Probabilistic Roadmaps for Navigation · p. 2
B. 3D Gaussian Splatting for Robot Navigation
- Gaussian-map planners either derive auxiliary planning geometry, as in GaussNav and SplatBlox, or check Gaussian confidence ellipsoids, as in SplatNav.3DGS mapping and localization works provide map-registered visual representations but not the persistent traversability graph needed for multi-query planning.Source: B. 3D Gaussian Splatting for Robot… · p. 2
- GS-PRM defines milestone and connection tests, image queries, and visual route references through rendering, unlike prior systems that render views or subgoals for visual policies.Rendered depth is not automatically a conservative collision oracle, so its link to physical collision freedom is treated as an explicit assumption and evaluated separately from connectivity.Source: B. 3D Gaussian Splatting for Robot… · p. 2
C. Goal-Conditioned Visual Navigation
- Goal-conditioned visual policies remain local and need a reachable, informative goal; GS-PRM supplies pose-indexed observations as subgoals and localization references without requiring a densely recorded image graph.The query consists of start and goal images whose configurations must be localized before estimating a physically feasible path.Source: C. Goal-Conditioned Visual Navigation · p. 2, A. Image-Goal Navigation Query · p. 2
B. RC-free Configuration and Path
GS-PRM uses 3DGS renderings to assess configurations and local paths. Panoramic rays test configuration clearance, while edge-aligned orthographic rays cover the safety-inflated swept volume of a connection.
B. RC-free Configuration and Path
- The rendering function produces RGB, depth, and accumulated opacity from a calibrated camera pose and camera intrinsics.Source: B. RC-free Configuration and Path · p. 2
- A panoramic render tests whether a configuration's safety-inflated robot body has reconstruction support and sufficient clearance along robot-relevant rays.Unsupported rays are treated as unknown rather than free; the clearance includes robot radius and a safety margin.Source: B. RC-free Configuration and Path · p. 3
- For a local path, edge-aligned orthographic rendering covers the safety-inflated swept cross-section, and the path is RC-free only with sufficient support and clearance.The clearance test includes a rendering and execution margin.Source: B. RC-free Configuration and Path · p. 3
C. Collision-free Conditions
The section relates render-based collision-free reasoning to physical free space under explicit reconstruction and ray-coverage assumptions. It also describes clearance-adaptive coverage regions used by GS-PRM to represent locally certified free space.
C. Collision-free Conditions
- The method assumes the 3DGS scene is adequately reconstructed to provide reliable depth and opacity for its rendering-based checks.The stated opaque-surface assumption requires obstacle-intersecting rays to accumulate threshold opacity before the true first-surface depth.Source: C. Collision-free Conditions · p. 3
- Under opaque-surface depth and conservative ray-coverage assumptions, the section frames the relationship between RC-free configurations and paths and physical free space as conditional.Insufficiently supported regions are treated as unknown, not free space.Source: C. Collision-free Conditions · p. 3
A. GS-PRM
- For circular robots, thresholded panoramic depth sets a clearance-adaptive translational coverage radius, capped by a maximum allowed radius.The coverage region is used both to reduce redundant sampling in open areas and to represent a locally certified free region.Source: A. GS-PRM · p. 3
- A GS-PRM roadmap stores RC-free milestones with certified local coverage and map-registered visual attributes, while edges represent locally validated paths.Images and descriptors may be stored offline or rendered during online navigation.Source: A. GS-PRM · p. 4
B. Coverage-Guided Sampling
GS-PRM uses coverage to retain samples from the workspace while preserving uniform proposal support. A bounded relocation heuristic can adjust accepted proposals without changing their RC-free acceptance test.
B. Coverage-Guided Sampling
- A low-cost panoramic RGB-and-normal back-face prefilter rejects likely interior viewpoints before the RC-free test.Proposals with insufficient supported pixels are rejected as unknown; under the opaque, orientable-surface condition, the prefilter effectively determines RC-free configurations.Source: B. Coverage-Guided Sampling · p. 4
- Uniform workspace proposals retain support over the workspace because coverage rejects redundant proposals without removing the uniform proposal process.Only uncovered RC-free proposals proceed to relocation and graph connection.Source: B. Coverage-Guided Sampling · p. 4
- Bounded local relocation selects an admissible iterate with greatest minimum directional clearance while retaining the original uniform proposal when coverage would exclude it.Each update is rerendered, and relocation does not change the RC-free acceptance test.Source: B. Coverage-Guided Sampling · p. 4
C. Graph Construction
Graph construction proposes nearby milestones using coverage overlap and a maximum connection distance, then retains only validated local connections. A new milestone joins the roadmap only when it has an RC-valid edge to the existing graph.
C. Graph Construction
- Neighbor candidates are proposed when milestone distance is within the smaller of the summed coverage radii and the maximum connection distance.Additional k-nearest candidates are tested for graph redundancy, and every proposed local path is evaluated for RC validity.Source: C. Graph Construction · p. 4
- Each candidate edge is tested by rendering depth and opacity along an orthographic view aligned with the connection and sized to the robot's margin-expanded cross-section.Unsupported pixels are unknown, and acceptance requires the rendered volume to be known and pass the depth test.Source: C. Graph Construction · p. 4
- An undirected edge must pass the render test in both directions, and a new milestone is committed only if it connects to the existing roadmap.Otherwise, the milestone is deferred until the coverage frontier reaches it; all RC-valid candidate edges are retained.Source: C. Graph Construction · p. 5
D. Completeness Guarantees
GS-PRM provides deterministic query completeness once certified coverage and connectivity hold, and probabilistic construction completeness as the number of independent construction trials grows. Physical collision freedom for returned paths is conditional on stated assumptions.
D. Completeness Guarantees
- For every pair of configurations in a connected component of the RC-free space, GS-PRM returns an RC-free path; under Assumptions 1 and 2, that path is collision-free.This is deterministic query completeness once the relevant component and roadmap conditions hold.Source: D. Completeness Guarantees · p. 5
- As the maximum number of independent construction trials tends to infinity, the probability that GS-PRM fails to connect the query configurations tends to zero.The guarantee is probabilistic completeness of roadmap construction.Source: D. Completeness Guarantees · p. 5
A. Start and Goal Localization
GS-PRM localizes image-defined endpoints against a map-registered 3DGS retrieval database and attaches them to a persistent roadmap. It validates searched and smoothed routes, then supplies rendered visual subgoals for short-horizon execution and pose correction.
A. Start and Goal Localization
- Image-defined start and goal images are localized through retrieved map views, feature matching, rendered-depth back-projection, and RANSAC-PnP; each endpoint must pass RC-free validation and attach through an RC-free connection.The estimated camera pose is converted to a robot configuration using camera-to-base calibration.Source: A. Start and Goal Localization · p. 5
B. Graph Search and Route Validation
- A* selects a milestone route using geometric length and normalized local coverage costs, while any shortcut or smoothing is accepted only after revalidation of the full modified trajectory.This guards against smoothing beyond volumes validated during roadmap construction.Source: B. Graph Search and Route Validation · p. 5
C. Visual Subgoal Execution and Online Robustness
- Rendered RGB and depth observations at sampled route poses become nearby visual subgoals for the goal-conditioned policy, extending its operating horizon; end-of-segment pose correction can mitigate accumulated localization drift.A deterministic straight-line planner is also described as an alternative to the learned policy.Source: B. Graph Search and Route Validation · p. 5, C. Visual Subgoal Execution and Online… · p. 5
VI. EXPERIMENTS AND RESULTS
In simulation, the RC-free criterion was evaluated against ground-truth geometry, and GS-PRM navigation was compared with SplatNav. All sampled configurations accepted as RC-free were confirmed collision-free, while some rejected configurations were also collision-free.
A. Verifying Physical Soundness
- The simulation experiments asked whether RC-free agrees with ground-truth collision checking and whether GS-PRM trajectories are reliable relative to a state-of-the-art approach.The study used Isaac Sim and framed these as two questions about physical soundness and navigation performance.Source: A. Verifying Physical Soundness · p. 5
1) Validation of RC-free Condition:
- 20.0% of uniformly sampled positions were classified as RC-free, and all were confirmed collision-free by simulator ground truth; the experiment reported no false-negative classifications.The validation used a 3DGS representation of the Isaac Sim environment.Source: 1) Validation of RC-free Condition · pp. 5–6
- Among the remaining 80% of rejected samples, 63.9% corresponded to actual collisions and 36.1% were collision-free, primarily reflecting conservative footprint and safety-clearance rejections.Fig. 2 presents validation of the RC-free criterion in the simulated environment.Source: 1) Validation of RC-free Condition · p. 6, Figure 2 · p. 6
2) Navigation Validation:
In simulation, GS-PRM+SLP was compared with SplatNav over randomly selected start-goal pairs. The real-world evaluation used two sites and multi-query routes composed of five subpaths, with rendered images serving as subgoals.
2) Navigation Validation:
- GS-PRM+SLP reduced planned trajectory length by 5.16%, traveled path length by 9.59%, and navigation time by 17.47% versus SplatNav over 20 simulated trials.Table I reports the average performance.Source: 2) Navigation Validation · p. 6
B. Real-World Experiment
- Real-world evaluation used success rate, SPL, penalized travel time, and global path segment coverage to assess multi-query navigation.Global path segment coverage is the ratio of covered path segments to total path segments over a multi-query task.Source: B. Real-World Experiment · p. 6
- The two real-world sites were a textured, consistently lit laboratory studio and a longer, texture-poor corridor with varying illumination; each evaluated route comprised five subpaths with rendered subgoal images.GS-PRM construction in the laboratory studio used 630K Gaussians and took 32 seconds.Source: B. Real-World Experiment · p. 6
1) Navigation Performance with Different Local Planning Policies:
In T1-5 and T2-5, GS-PRM-guided local planners were generally more reliable than standalone learned policies, while the SplatNav comparison favored GS-PRM+SLP. The conclusion states completeness conditions and boundaries tied to reconstruction quality and static scenes.
1) Navigation Performance with Different Local Planning Policies:
- In T1-5, standalone DevGRU and ViNT navigation was unreliable; GS-PRM-guided methods were generally more reliable, though GS-PRM+ViNT still struggled.SLP was relatively straightforward because it used rendered observations for frequent pose estimation and correction without a learned policy.Source: 1) Navigation Performance with… · p. 6
2) Comparison against SplatNav:
- In T1-5, SplatNav travel time was approximately 39% longer than GS-PRM+DevGRU and 134% longer than GS-PRM+SLP.The comparison reports averages over two runs per method and environment.Source: Table II · p. 7, 2) Comparison against SplatNav · p. 8
1) Navigation Performance with Different Local Planning Policies:
- In T2-5, all methods except SLP failed to complete the full task, while the failures of ViNT and DevGRU were likely associated with a domain gap.The authors state that performance can depend on 3DGS reconstruction quality and local-policy robustness to rendered observations.Source: 1) Navigation Performance with… · pp. 7–8
VII. CONCLUSION
- GS-PRM’s completeness guarantees depend on conservative rendering and robust local-connectability conditions, while poorly reconstructed regions, reconstruction quality, and dynamic scenes bound its scope.The current formulation assumes a static environment and does not explicitly account for dynamic obstacles or scene changes.Source: §VII · p. 8