Technical Summary: A Sensor-Centric Survey of SLAM and Odometry for GPS-Denied Environments

Problem Definition

The paper addresses the challenge of Simultaneous Localization and Mapping (SLAM) and odometry in environments where Global Navigation Satellite System (GNSS) signals are unavailable, unreliable, or obstructed. These "GPS-denied" conditions include subterranean mines, building interiors, underwater settings, dense forests, and adversarial environments. In the absence of external positioning fixes, autonomous platforms must rely entirely on onboard sensing. The core problem is that localization performance in these degraded conditions is fundamentally constrained by the specific sensing modalities available; each sensor type possesses unique failure modes (e.g., cameras failing in low light or texture-poor scenes, LiDAR failing in geometrically repetitive structures, and inertial sensors drifting rapidly without correction).

Methodology

The authors adopt a sensor-centric taxonomy rather than an algorithmic one, organizing the literature based on the sensing configuration required by the system. The review covers six distinct categories:

1. Visual: Cameras only.
2. Visual-Inertial (VI): Cameras + Inertial Measurement Units (IMU).
3. LiDAR: Ranging laser scanners only.
4. LiDAR-Inertial (LI): LiDAR + IMU.
5. LiDAR-Visual-Inertial (LVI): All three sensors fused.
6. Residual Modalities: Sonar, radar, and thermal sensing.

To ensure a fair comparison between field-defining classics and very recent works, the authors employ a four-tier maturity framework within each modality:

• Foundational: Field-defining works established by community consensus.
• Established: Mature, widely adopted systems with stabilized citation curves.
• Recent: Methods with rising citation curves, selected via open-source adoption and benchmark performance.
• Frontier: Very recent works lacking citation data, selected based on publication venue, open-source activity, and benchmark standing.

Crucially, the temporal boundaries for these tiers are calibrated per modality rather than fixed to calendar years, acknowledging that subfields matured at different rates. The review synthesizes 135 selected papers, analyzing their front-end estimation, back-end optimization, loop closure mechanisms, and map representations.

Key Contributions

1. Hardware-Centric Taxonomy: The paper unifies SLAM and odometry literature under a GPS-denied operating assumption, categorizing methods by their required sensing configurations rather than algorithmic techniques.
2. Maturity-Aware Organization: It introduces a four-tier selection methodology that allows foundational and frontier works to be weighed on comparable terms, avoiding the bias of citation-count rankings which inherently favor older work.
3. Cross-Modal Analysis: The survey provides a comparative analysis of datasets, evaluation metrics, and the trade-offs between accuracy, robustness, and resource constraints (Size, Weight, and Power - SWaP) across all six configurations.
4. Identification of Field Trends: The authors synthesize three recurring movements across all configurations:
   • Progressive multi-sensor fusion to leverage complementary failure characteristics.
   • Integration of learning (neural networks) within geometric pipelines.
   • A migration toward neural and differentiable map representations (e.g., Neural Implicit Fields, 3D Gaussian Splatting).

Results and Findings

The survey concludes that no single sensing configuration is universally optimal for GPS-denied environments. Instead, robustness depends on the complementarity of sensor failure modes:

• Visual/VI: Lightweight and information-rich but vulnerable to low light, motion blur, and textureless environments.
• LiDAR/LI: Provides metric-scale geometry independent of illumination but struggles in self-similar or structure-poor environments (e.g., long tunnels) and is susceptible to motion distortion.
• LVI (LiDAR-Visual-Inertial): Represents the current state-of-the-art for robustness, as the sensors cover each other's blind spots (e.g., LiDAR provides geometry when vision fails; vision provides texture when LiDAR faces geometric degeneracy).
• Residual Modalities (Sonar/Radar/Thermal): Essential for specific regimes (underwater, smoke, dust, total darkness) where optical and laser sensors fail, though these modalities generally suffer from noisier, sparser data and less mature benchmarking.

The review highlights that while LiDAR-Inertial is the dominant configuration for ground and aerial robots in subterranean settings, the LiDAR-Visual-Inertial configuration offers the highest resilience, validated notably in the DARPA Subterranean Challenge. Furthermore, the field is shifting from sparse geometric maps to dense, differentiable, and photorealistic representations (such as 3D Gaussian Splatting) that are co-optimized with the trajectory.

Significance and Claims

The paper asserts that the primary obstacle to rigorous cross-modal evaluation and sustained progress is fragmented, degradation-agnostic benchmarking. Current benchmarks often lack controlled sensor degradation scenarios, making it difficult to assess robustness in the very conditions where GPS-denied localization is most critical.

The authors claim that the future of robust GPS-denied localization lies not in maximizing the nominal accuracy of a single modality, but in:

1. Fusing sensors with uncorrelated failure characteristics.
2. Developing estimators capable of quantifying and adapting to their own uncertainty (degeneracy awareness).
3. Moving toward lifelong, semantic, and collaborative operation.

The review serves as a comprehensive reference for understanding the trade-offs between sensing configurations and identifies the need for unified, multimodal benchmarks that systematically vary environmental degradations to properly evaluate the resilience of autonomous systems.