Introduction
LiDAR-based global localization is a cornerstone of autonomous navigation, yet it faces significant challenges in environments characterized by "semantic ambiguity." In settings like massive warehouses, repetitive parking structures, or dense forests, standard LiDAR sensors often perceive geometrically identical features. This leads to the "mirror image" problem, where a robot cannot distinguish between two locations because their spatial layouts—and even the centroids of detected objects—are indistinguishable.
Existing state-of-the-art methods utilizing 3D scene graphs typically group points into object-level nodes. However, by focusing primarily on object centers, these methods discard the subtle environmental context that could break symmetry. To overcome this, we present Outram-GSF, a localization system that leverages "Gaussian Semantic Fields" to capture the unique geo-semantic signature of an area.
The Gaussian Semantic Field Approach
The fundamental shift in this work is moving from discrete semantic labels to continuous semantic distributions. Instead of simply noting that a point belongs to a "tree" or a "building," the system models how these semantics are distributed in the space surrounding an object.
By using Gaussian Processes (GPs), the system transforms sparse, noisy semantic data into a continuous, differentiable function. This function acts as a local "fingerprint." Even if two pillars in a garage are spaced identically, the specific distribution of background vegetation or ground geometry behind them is rarely perfectly symmetric. The GSF leverages this "background noise" as a robust feature for localization.
Methodology
The Outram-GSF pipeline is built upon a tri-layered scene graph architecture, integrating metric, semantic, and field-based information.
- Tri-layered Graph Generation: The architecture consists of a Metric Semantic Layer (raw points), a Gaussian Semantic Field Layer (continuous distributions), and an Object Layer (high-level entities).
- Metric Semantic Feature Generation: To handle the computational intensity of GPs on dense LiDAR data, the authors developed a semantic-based sparsification algorithm. This method intelligently downsamples points based on their class, ensuring the GP learns the most representative features while maintaining efficiency.
- Substructure Matching via Wasserstein Metric: To compare the semantic field of a current scan against the map, the system employs the 2-Wasserstein metric. This mathematical framework is ideal for measuring the similarity between probability distributions, allowing for precise matching even under slight viewpoint variations.
- Semantic Stability Mask: Not all landmarks are equally reliable. The system applies a weighting mask that prioritizes stable infrastructure (e.g., buildings) over natural elements (e.g., bushes) and transient objects (e.g., vehicles).
Experimental Evaluation
The researchers evaluated Outram-GSF on the MulRan and MCD NTU datasets, with a specific focus on highly repetitive sequences.
- Robustness in Repetitive Environments: In the MCD NTU "repetitive" test, Outram-GSF achieved an 87.36% success rate, a dramatic improvement over the previous best method, SGTD, which reached only 55.26%.
- High Precision: On standard urban benchmarks, the system achieved a 99.32% success rate. When paired with Generalized Iterative Closest Point (GICP) refinement, the Average Translation Error (ATE) was reduced to just 0.41 meters.
- Generalization: The method demonstrated consistent performance across different sensor setups and urban environments without requiring extensive per-scene tuning.
Limitations and Future Work
While highly accurate, the system's complexity results in a computational latency of approximately 563ms per frame. This is suitable for "one-shot" global relocalization (e.g., when a robot first wakes up) but may be too slow for real-time high-frequency tracking.
Future research directions include optimizing the Gaussian Process inference for real-time performance and developing "localizability certificates." Such certificates would allow the system to autonomously identify areas that are too symmetric for reliable localization, prompting the robot to seek more distinctive landmarks before confirming its position.
Summary updated