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

arXiv:2610.11752 · cs.RO · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map".

Dev: The gist—2DGS-Planner proposes 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.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at 2DGS-Planner today, which is this path planner for ground robots that reads geometry directly from a 2D Gaussian splatting map using rasterization instead of treating every single Gaussian primitive as an obstacle <ref:2610.11752#pg1,path planner for ground robots that reads>.

Dev: That's right, Rosa. It’s about bypassing the need for that intermediate geometric representation entirely and querying the map through rasterization to get planning-relevant geometry, which is a big shift.

Taro: So instead of building a mesh or some other explicit structure first, you're going straight from the raw 2DGS data to making connections on the roadmap, which sounds much more efficient for real-time scenarios <ref:2610.11752#pg1>.

Rosa: Exactly. And during that offline roadmap construction phase, they introduce this multi-view attribution technique that converts what’s rendered as normal dispersion into structural scores for non-ground disks supported by those reconstruction views.

Dev: That’s how they get these structural scores which then guide the adaptive node sampling on the ground, meaning they place more nodes near corners and cluttered areas and fewer nodes in open regions.

Taro: I mean, if you can intelligently decide where to put your nodes based on how complex the surface looks from different angles, that should help the planner focus its computational effort much better than just throwing points everywhere.

Rosa: That's the point of this multi-view attribution; it uses that rendered normal dispersion—which is calculated as one minus the square of a normalized normal—to identify areas where multiple surface orientations contribute to a single pixel, like near corners or thin structures.

Dev: And they then aggregate that per-disk attribution across all reconstruction views to get these structural scores Σi for each non-ground disk, and they normalize it into these structural scores in the range of zero to one.

Taro: So they’re essentially quantifying the local complexity of the scene based on how much different parts of the map contribute to a single point, excluding ground disks from that ranking so their low dispersion values don't skew the results upward.

Rosa: Right, and then those structural scores guide where they sample their nodes during roadmap construction, and then for path validation, they use path-aligned orthographic queries to screen candidate edges.

Dev: And after that screening, they use cylindrical queries to estimate local clearance fields which are cached right onto the edges themselves.

Taro: So if an edge passes the orthographic screen, the cylindrical query gives you a field describing how much space is around that edge in a local cylindrical coordinate system, and those fields get saved for later refinement.

Title and authors: Rosa: That cached clearance field is what they reuse during online planning when the graph search finds an initial route, allowing them to refine it while accounting for the robot’s own dimensions and ground constraints.

Dev: They use a path-aligned orthographic query to validate those edges, and then the cylindrical queries estimate local clearance for the remaining candidate edges, which they keep cached for path refinement.

Taro: When you're actually refining that route online, they use that cached field to figure out the clearance c(qn) as the distance to the nearest surface sample associated with that specific route edge.

Rosa: And then they update the B-spline control points iteratively by minimizing this energy function which includes a term for required clearance, specifically where rn is calculated as min b plus one half a squared kappa n, max a, b.

Dev: That last part about accounting for the robot's dimensions during the refinement step is crucial because it increases the required clearance as you go around turns based on local path curvature.

Taro: So what this means practically is that when planning a turn, the system knows it needs more space because of how curved that path segment is, which makes sense for a physical robot to navigate.

Rosa: It shows they’ve incorporated robot-specific constraints directly into the optimization process, rather than just planning in an idealized environment.

Dev: And on the larger scale, this framework lets them perform real-world deployment because they use the reconstructed map registered with LiDAR data and execute that refined path using a pure-pursuit controller.

Taro: So you can take this from a lab setup, through mapping and planning, right into actual physical navigation in the field.

Rosa: It’s been quite an effort to move from just having a map to actually having a reliable, geometrically informed planner that handles both the structural understanding and the dynamic constraints of the robot.

Dev: And while path refinement adds computation for online planning, they report that this full planner is still faster than RRT* on both scenes, though it does show higher latency than Disk PRM and ESDF because of that extra refinement work.

Taro: That’s a fair trade-off, I guess; you get better path quality and clearance estimation for the cost of a bit more computation when you need to adjust mid-flight.

Rosa: Overall, this paper, 2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map, really demonstrates that treating rasterization as the primary geometric query interface is an effective way to plan directly on these Gaussian maps without needing a separate intermediate geometric representation <ref:2610.11752#pg1,2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map>.

Title and authors: Dev: And it’s not just about the map reading; the structural scores derived from multi-view attribution and the cached clearance fields for refinement show a solid approach to handling complex, non-ground surfaces.

Taro: It's interesting how they use that structural information to guide node placement, specifically placing more nodes near corners and cluttered areas, which is a smart way to prioritize sampling effort.

Rosa: So what this implies for the broader field is that you can leverage the explicit and efficiently rasterizable nature of Gaussian splatting for robot navigation planning directly.

Dev: It moves away from treating individual Gaussians as simple obstacles and instead uses the structure learned from those Gaussians to intelligently guide the entire path planning process.

Taro: For autonomy researchers, it suggests that building structural scores based on rendered normal dispersion is a useful way to inject scene complexity information into sampling strategies for navigation.

Rosa: And for deployment, the fact that it works with real-world LiDAR registration and pure-pursuit control shows it’s not just theoretical work stuck in simulation.

Dev: The latency reports confirm that while refinement adds computation, the full planner remains faster than RRT* on both scenes, which is good news for real-time constraints.

Taro: I wonder what happens when the world misbehaves unexpectedly? Does this structural scoring system still give you enough guidance when things deviate from the learned scene structure?

Rosa: That's a valid question. The paper notes that this formulation approximates the robot body with an ellipsoid, which limits how detailed and articulated those robot geometries can be represented right now.

Dev: They also admit that they don't explicitly account for robot dynamics, meaning future work will need to incorporate kinematic and dynamic feasibility into the planning process.

Taro: So this is a solid foundation for using 2DGS maps for planning, but the next step needs to be tackling those more detailed robot models and ensuring dynamic feasibility when things get tricky <ref:2610.11752#pg1>.

Rosa: Exactly. The 2DGS-Planner paper shows how powerful this specific pipeline is, especially with its novel approach to querying geometry through rasterization and using structural attributes to guide both node placement and path refinement <ref:2610.11752#pg1>.

Dev: And that's the gist of 2DGS-Planner: it’s a ground robot path planner that queries 2DGS maps through rasterization without an intermediate geometric representation, introducing multi-view attribution for adaptive sampling and cached cylindrical queries for clearance estimation <ref:2610.11752#pg1>.

Taro: It’s a practical application of using scene structure to make the planning process more intelligent than just blind obstacle avoidance.

Rosa: Well, that’s our take on 2DGS-Planner today <ref:2610.11752#pg1>. We'll be taking a quick break, and when we come back we're going to look at how other systems are handling real-world learning problems.

The paper's summary: Taro: So it skips that whole intermediate step where you have to build some mesh or explicit geometric representation first?

Dev: Exactly. They're treating the rasterized data as the geometric query interface itself, so you get planning-relevant geometry right from the 2DGS map without needing a separate structure <ref:2610.11752#pg1>.

Rosa: And they’re making sure that this works for ground robots in real-world scenarios, not just some pristine lab environment where everything is perfectly aligned.

Taro: It's interesting how they tie in these structural scores derived from multi-view attribution to guide the node placement on the roadmap.

Dev: That’s right, those scores tell them where to put more nodes—near corners and cluttered spots—and fewer nodes in open areas, which is a smarter way to sample the space.

Rosa: And then for refining the path once they have an initial route, they use these cached clearance fields from cylindrical queries to figure out exactly how close the robot is to surfaces along that edge.

Taro: That cached information is what allows them to update those B-spline control points iteratively, accounting for the robot's actual size and how much extra space you need when you’re turning.

Dev: They use an energy function for that optimization that explicitly includes a term for required clearance based on path curvature, which makes sure the robot has enough room to maneuver during those turns.

Rosa: It shows they’ve put robot-specific constraints right into the refinement loop, instead of just planning in some idealized vacuum.

Taro: So what this means practically is that you can take a 2DGS map and feed it straight into a planner that intelligently uses the scene's surface complexity to guide both how it builds its roadmap and how it refines its final path <ref:2610.11752#pg1>.

Dev: The latency reports show that while this refinement adds computation for online planning, the full planner still keeps up with things like RRT* on both scenes, even though it’s slower than some of the simpler methods they compared.

Rosa: It’s about getting a better quality path and clearance estimate for that extra computational cost when you need to adjust mid-flight.

Taro: So this moves away from just blind obstacle avoidance and uses the learned structure from the 2DGS map to make a much more intelligent decision about where to sample and how to follow the path <ref:2610.11752#pg1>.

Dev: The authors do mention that they approximate the robot body with an ellipsoid, which means it doesn't yet handle really detailed or articulated robot geometries perfectly.

Rosa: And they also admit that they haven't explicitly accounted for robot dynamics yet, so future work will need to factor in things like kinematic and dynamic feasibility.

Taro: That makes sense. It’s a solid foundation for using 2DGS maps for planning, but the next step is tackling those more detailed models and making sure the plan actually works when you hit a tricky situation <ref:2610.11752#pg1>.

Dev: So this paper shows how powerful this specific pipeline is with its novel approach to querying geometry through rasterization and using structural attributes to guide node placement and path refinement.

Taro: And it’s interesting how they use that structural information to guide node placement, specifically placing more nodes near corners and cluttered areas, which is a smart way to prioritize sampling effort.

Dev: It’s about getting a better quality path and clearance estimate for that extra computational cost when you need to adjust mid-flight.

Rosa: So what this implies for the broader field is that you can leverage the explicit nature of Gaussian splatting for robot navigation planning directly, bypassing those intermediate geometric steps.

Taro: For autonomy researchers, it suggests building structural scores based on rendered normal dispersion is a useful way to inject scene complexity information into sampling strategies.

The paper's improvements: Taro: So we're looking at what they suggest for making this planner even better, beyond just getting the basic path found.

Dev: They focus on how to make those clearance estimations more robust when things get messy in the real world, which is where most of the gains are.

Rosa: They’re talking about incorporating robot-specific constraints more deeply into that refinement loop than they did in the initial version.

Taro: Right, they want to move beyond just a rough ellipsoid approximation and actually model how a robot's turning radius affects the required clearance field during path optimization.

Dev: So instead of just using a simple formula for curvature, they suggest making that term more detailed so it better reflects the actual space needed for an elongated body to turn safely.

Rosa: That means when you're refining the route online, the system should be smarter about how much clearance it needs based on where you are turning in relation to your robot's physical dimensions.

Taro: It’s about making sure that the path refinement isn't just mathematically sound but also physically feasible for a real machine to execute without scraping walls.

Dev: I see what they mean; they want that clearance term, rn, to be more accurately calculated based on the local path curvature kappa n and those robot dimensions.

Rosa: So this pushes them toward a more detailed model of the robot body itself rather than just treating it as a simple shape for collision checking.

Taro: And they also touch on incorporating dynamic feasibility, which is something they admit they haven't fully done yet, but it’s the next logical step for making this planner truly autonomous.

Dev: It's a bit of a limitation, though; they’re still approximating the robot body with an ellipsoid and haven't added full dynamic constraints to the planning process yet.

Rosa: So while they get great at geometric planning and structural understanding from the 2DGS map right now, future work has to focus on making it handle more complex robot shapes and actual physics <ref:2610.11752#pg1>.

Taro: That’s a fair point; we’re still building on this foundation, moving from structural awareness to full physical feasibility.

Dev: It means that while the current planner is fast and good at finding a path, deploying it reliably in complex dynamic environments will need that extra work on the robot model side.

Rosa: So what this suggests is that we can get really good geometric planning using 2DGS now, but we need to pair it with more sophisticated robot modeling for true real-world autonomy <ref:2610.11752#pg1>.

Conclusion: Rosa: So to wrap up, this paper on "2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map" shows how you can use rasterization directly for planning on these 2DGS maps instead of needing some other intermediate structure <ref:2610.11752#pg1,2DGS-Planner: Rasterization-based Path Planning in 2D Gaussian Splatting Map>.

Dev: It’s a solid approach because it avoids that extra geometric conversion step, which is pretty efficient for the loop rate we need.

Taro: For autonomy researchers, this is significant because it shows how scene structure information can guide adaptive sampling on the roadmap based on structural scores derived from multi-view attribution.

Rosa: Yeah, and they proved that you can get better clearance estimation by caching those cylindrical queries along the edges of your path.

Dev: That caching part is key for online planning because it means you reuse that calculated data instead of re-running expensive queries every single time you slightly adjust the route.

Taro: And what this changes for someone listening is that it demonstrates a way to inject scene complexity directly into the sampling strategy, which helps the planner focus its effort intelligently.

Rosa: It's about getting better quality path and clearance estimates for that extra computational cost when you need to adjust mid-flight.

Dev: The caveat is that they still approximate the robot body with an ellipsoid, so it’s not yet perfect for really complex or articulated robot shapes.

Taro: And they also don't fully account for dynamics yet, so the next big challenge is moving from geometric planning to something that considers physics in real-time.

Rosa: Exactly. So we see a strong framework here leveraging the structure of 2DGS maps for navigation, but there's still work to do on modeling the robot and its movement properly <ref:2610.11752#pg1>.

Dev: I agree, it’s a great starting point for getting a fast, good path in complex scenes with this specific map representation.

Taro: It really highlights that leveraging explicit geometry like rasterized 2DGS data can be a very effective interface for planning directly on these maps <ref:2610.11752#pg1>.

Jiwon Park, Dong-Uk Seo, Hyun Myung

cs.RO

Submitted: 2026-10-08

Updated: 2026-10-08

Project page: https://2dgsplanner.github.io

The gist: The gist—2DGS-Planner proposes a path planner for ground robots that reads planning-relevant geometry from a 2D Gaussian splatting (2DGS) map through rasterization, rather than treating individual

Key concepts

Rasterization: This process converts the 3D geometry stored in the 2DGS map into a 2D surface structure. Instead of analyzing each individual Gaussian point, rasterization creates a continuous representation that allows the planner to read surface features and local clearance directly, which is more efficient for path planning.
Multi-view Attribution: This technique extracts structural scores from the rendered normal dispersion across multiple reconstruction views. It analyzes how surface orientations contribute to a single pixel's appearance, identifying areas like corners or clutter where complex geometry exists, and uses these scores to prioritize where the planner should place nodes.
Local Clearance Field (LCC): This field estimates the surrounding space available for a robot along a proposed path edge. It is cached on the edges during planning and allows for corridor-constrained refinement by quickly calculating the distance to nearby surfaces, which helps in optimizing the path to avoid obstacles.
Structural Score ($\Sigma_i$): This score quantifies the local geometric complexity of a non-ground disk based on its rendered normal dispersion. High scores indicate areas with many surface orientations contributing to a single pixel, such as corners or thin structures, guiding adaptive node placement in these challenging regions.

Terminology

Gaussian Splatting — 25× in this paper · explained in 4 episodes
Fidelity — 7× in this paper · explained in 2 episodes
AUC — 6× in this paper · explained in 2 episodes
Adaptive Sampling — 5× in this paper · explained in 1 episode
alpha — 4× in this paper · explained in 1 episode
Consistency — 2× in this paper · explained in 3 episodes
3D Gaussian Splatting — 1× in this paper · explained in 2 episodes
3D Gaussian Splatting (3DGS) — 1× in this paper · explained in 4 episodes
Ablation — 1× in this paper · explained in 1 episode
Accuracy — 1× in this paper · explained in 1 episode
Aggregation — 1× in this paper · explained in 1 episode
Alignment — 1× in this paper · explained in 1 episode
Core — 1× in this paper · explained in 2 episodes
Gaussian Splatting (GS) — 1× in this paper · explained in 3 episodes
Regularization — 1× in this paper · explained in 3 episodes
Transformer — 1× in this paper · explained in 2 episodes

Summary

The gist—2DGS-Planner proposes 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.

How it works

The framework constructs a roadmap directly from rasterized geometric queries without converting the scene into an intermediate geometric representation From Rasterization to Path Planning, Fig 1, 2DGS-Planner uses rasterization to read surface structure and local clearance directly from a 2D Gaussian splatting (2DGS) map >
During offline roadmap construction, multi-view attribution converts rendered normal dispersion into structural scores for non-ground disks supported by the reconstruction views Multi-view attribution derives structural scores from rendered normal dispersion >
These scores guide adaptive node sampling on the ground, placing more nodes near corners and cluttered areas and fewer in open regions The scores guide adaptive node placement >
Path-aligned orthographic queries validate candidate edges, while cylindrical queries estimate local clearance fields that are cached on the edges Orthographic queries screen candidate edges, and cylindrical queries estimate surrounding clearance for path refinement >

Map Preprocessing and Structural Scoring

The map preprocessing stage involves assigning semantic labels to the Gaussian disks and aligning the reconstruction with gravity Ground-labeled disks satisfying the orientation and height criteria form Ggrd ⊂ G, which defines the support surface for roadmap sampling and path refinement Multi-view attribution estimates local structural complexity from the variation of rendered normals and uses it to guide roadmap sampling The per-pixel normal dispersion is computed as Vk(p) = X i w hat ki(p)ni - nbar k(p) squared = 1 - nbar k(p) squared This dispersion rises where multiple surface orientations contribute to a single pixel, as occurs near corners, thin structures, and cluttered regions >

The per-disk attribution aggregates the rendered dispersion across reconstruction views using each disk’s alphacompositing contribution The structural score Σi is converted into a normalized structural score Σi ∈ [0, 1] using its empirical percentile among eligible non-ground disks Ground disks are excluded from the ranking to prevent their numerous lowdispersion values from biasing non-ground scores upward >

Roadmap Construction and Query Formulation

Nodes are sampled from ground-disk centers and lifted to the prescribed ride height, with each candidate q assigned a local complexity score u(q) The edges are proposed using a Gabriel graph [23], and each candidate edge is first screened by an orthographic query, after which a cylindrical query estimates local clearance for the remaining edges For edges passing the orthographic screen, the cylindrical query records surrounding geometry in a local cylindrical coordinate (LCC) system The clearance field ρe(s, θ) is defined only for observed rays with A(s, θ) ≥ τA

Path Optimization and Refinement

The initial path is connected using the same path-aligned queries, and A∗ provides a graph route for B-spline initialization The corridor-constrained refinement reuses the cached local cylindrical coordinate (LCC) field to evaluate clearance c(qn) as the distance to the nearest surface sample associated with the corresponding route edge The spline control points are then updated iteratively by minimizing E = wlenX n qn+1 − qn squared + wclrX n [rn + δ − c(qn)] squared This iterative process accounts for the additional space required by an elongated body during turning, increasing the required clearance with local path curvature κn as rn = min b + 1/2 a 2κn, max(a, b) >

Evaluation and Results

Experiments demonstrate improved roadmap connectivity, more accurate clearance estimation, and higher planning success compared to the tested baselines 2DGS-Planner achieves the highest SPL on both scenes Furthermore, LCC fields yield lower clearance error and fewer false-blocked configurations than the tested ESDFs The planner's performance is evaluated across roadmap construction, geometric query fidelity, and end-to-end planning In real-world deployment, the robot successfully tracks the planned path using a pure-pursuit controller The results support rasterization as an effective geometric query interface for planning directly on Gaussian maps >

Robot-Size-Aware Planning

The planner accounts for robot dimensions by revalidating the path using the corresponding robot-specific orthographic footprint query A robot with a lateral half-width of 0.3 is routed through the wider passage, while reducing the half-width to 0.1 allows the planner to use the narrower passage and obtain a shorter path This shows that robot dimensions affect roadmap connectivity and the resulting path through robot-specific orthographic footprint queries >

Online Planning Latency

Path refinement increases online computation, but the full planner remains faster than RRT∗ on both scenes The latency reports show that while refinement adds computation, the full planner is faster than RRT∗ on both scenes Online planning latency is higher than Disk PRM and ESDF, reflecting the additional computation required for refinement alongside the improved SPL and reference collision-pass performance >

Conclusion

The current formulation approximates the robot body with an ellipsoid, which limits the representation of detailed and articulated robot geometry, and does not explicitly account for robot dynamics Future work will consider more detailed robot-body models and incorporate kinematic and dynamic feasibility into the planning process >

The gist

2DGS-Planner proposes 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.