Splatblox: Traversability-Aware Gaussian Splatting for Outdoor Robot Navigation

TL;DR

Splatblox addresses robust outdoor navigation by constructing a traversability-aware ESDF through real-time Gaussian Splatting that fuses RGB-derived semantic costs with LiDAR geometry. The method embeds semantic traversability into a compact GSplat field, renders a frontal ESDF, and fuses it with a 360° LiDAR ESDF to support planning up to 100 meters, all on edge GPUs without offline training. It achieves significant gains over baselines, with up to 60% higher success rates, 40% fewer freezing incidents, and up to 13% faster time-to-goal, while enabling long-range missions. The work demonstrates practical outdoor applicability and lays groundwork for extending robustness with additional sensing modalities and faster update rates.

Abstract

We present Splatblox, a real-time system for autonomous navigation in outdoor environments with dense vegetation, irregular obstacles, and complex terrain. Our method fuses segmented RGB images and LiDAR point clouds using Gaussian Splatting to construct a traversability-aware Euclidean Signed Distance Field (ESDF) that jointly encodes geometry and semantics. Updated online, this field enables semantic reasoning to distinguish traversable vegetation (e.g., tall grass) from rigid obstacles (e.g., trees), while LiDAR ensures 360-degree geometric coverage for extended planning horizons. We validate Splatblox on a quadruped robot and demonstrate transfer to a wheeled platform. In field trials across vegetation-rich scenarios, it outperforms state-of-the-art methods with over 50% higher success rate, 40% fewer freezing incidents, 5% shorter paths, and up to 13% faster time to goal, while supporting long-range missions up to 100 meters. Experiment videos and more details can be found on our project page: https://splatblox.github.io

Figures (6)

• Figure 1: Navigation trajectory of our method Splatblox (red) compared to baselines (Nvblox millane2024nvblox, DWA fox2002dynamic, MIM sathyamoorthy2024mim, VERN sathyamoorthy2023vern, GA-Nav guan2022ga). Our method achieves the lowest normalized trajectory length (NTL) and highest success rate by about 3% and 60% respectively, compared to the second-best method, when navigating through the narrow corridor between bushes (see Table \ref{['table:table_1']} for quantitative results). The top-right inset shows the traversability cost volume produced by our GSplat module, with a colormap indicating traversability cost: yellow denotes higher cost (obstacles) and dark blue denotes lower cost (free space).
• Figure 2: Splatblox architecture: RGB frames are processed by a segmentation module to assign traversability costs to image pixels, which are then associated with LiDAR points projected into the camera frame. These costs and 3D points are used for spawning new Gaussian primitives, forming a traversability-aware volumetric field. The Gaussian Splatting module incrementally updates this field online, maintaining a compact representation on resource-constrained GPUs. From the volumetric field, we derive a GSplat-based Euclidean Signed Distance Field (ESDF) in the robot’s frontal region, which is fused with a LiDAR-based ESDF to ensure 360$^\circ$ geometric coverage. The fused traversability-aware ESDF is provided to the planner as a distance field for collision checking and trajectory generation, enabling the robot to compute safe and efficient paths in vegetation-rich outdoor environments.
• Figure 3: ESDF fusion strategy. Top: Robot’s camera view (left) and fusion strategy (right), where the frontal region is covered by GSplat ESDF (red) and 360° coverage is provided by LiDAR ESDF (green). Bottom: GSplat ESDF (left) encodes fine-grained traversability detail, LiDAR ESDF (middle) provides global geometric consistency, and their fusion (right) combines both to produce collision-free, traversability-aware trajectories. Warmer colors denote regions near obstacles (low distance), cooler colors represent free space (larger distance).
• Figure 4: Top: Navigation trajectories of our method Splatblox (red) compared with baselines (Nvblox millane2024nvblox, DWA fox2002dynamic, MIM sathyamoorthy2024mim, VERN sathyamoorthy2023vern, GA-Nav guan2022ga) across four outdoor scenarios with diverse terrains, including paved areas, vegetation, and uneven ground. In each case, Splatblox achieves successful traversal with the shortest normalized trajectory length (NTL) through narrow passages and cluttered regions, while several baselines either deviate or fail to reach the goal. Bottom: Traversability cost volume generated by our Gaussian Splatting module (represented as point cloud for visualization) per scene (on top), where yellow indicates high cost (obstacles) and dark blue indicates low cost (traversable). Splatblox (red) trajectories are overlaid. Unlike LiDAR-only baselines that treat all objects as obstacles, Splatblox leverages semantics to navigate safely through low-cost regions (e.g., grass, bushes), demonstrating the benefit of incorporating semantic cues into the fused ESDF. Overall, Splatblox achieves up to 60% higher success rate (SR), 40% lower freezing rate (FR), paths up to 7% shorter (NTL), and time to reach goal (TRG) improvements of up to 13% compared to baselines.
• Figure 5: Top: Splatblox (red) is the only method that successfully completes this scenario. The trajectory begins from the pavement into the grass (left image) and transitions to the vines and shrubs (right image). Our method is the only one able to navigate through the dense vegetation to reach the goal. Bottom: All methods succeed in a simpler scenario, where the trajectory begins on pavement (left image) and transitions into grass (right image). In this case, Splatblox ranks second in normalized trajectory length (NTL) but achieves the fastest time to goal (TRG), improving by 3% over the best baseline.