Towards Next-Generation SLAM: A Survey on 3DGS-SLAM Focusing on Performance, Robustness, and Future Directions

Abstract: Traditional Simultaneous Localization and Mapping (SLAM) systems often face limitations including coarse rendering quality, insufficient recovery of scene details, and poor robustness in dynamic environments. 3D Gaussian Splatting (3DGS), with its efficient explicit representation and high-quality rendering capabilities, offers a new reconstruction paradigm for SLAM. This survey comprehensively reviews key technical approaches for integrating 3DGS with SLAM. We analyze performance optimization of representative methods across four critical dimensions: rendering quality, tracking accuracy, reconstruction speed, and memory consumption, delving into their design principles and breakthroughs. Furthermore, we examine methods for enhancing the robustness of 3DGS-SLAM in complex environments such as motion blur and dynamic environments. Finally, we discuss future challenges and development trends in this area. This survey aims to provide a technical reference for researchers and foster the development of next-generation SLAM systems characterized by high fidelity, efficiency, and robustness.

• The paper introduces 3DGS-SLAM, a novel pipeline that leverages explicit 3D Gaussian representations for photorealistic real-time mapping and robust tracking.
• The paper details a comprehensive optimization framework using differentiable projection, parallel rasterization, and adaptive loss minimization to enhance performance and memory efficiency.
• The paper benchmarks state-of-the-art techniques achieving high PSNR, sub-centimeter tracking accuracy, and scalable deployment in dynamic, challenging environments.

Surveying the Progress and Future Directions of 3DGS-SLAM

Introduction: Motivations for 3DGS Integration in SLAM

Simultaneous Localization and Mapping (SLAM) is foundational for robotics, AR/VR, and autonomous vehicles, requiring precise pose estimation and map construction within unknown environments. Classical SLAM systems—using geometric representations such as sparse point clouds, voxels, and meshes—achieve real-time operation but remain limited in their rendering realism, inability to recover photorealistic views, and vulnerability to environmental complexity. The emergence of neural scene representations, notably Neural Radiance Fields (NeRF), introduced high-fidelity mapping at substantial computational cost and strict requirements on dense sampling.

3D Gaussian Splatting (3DGS) offers a paradigm shift: it provides an explicit representation that delivers photorealistic rendering, fast volume rasterization, and differentiable optimization suitable for real-time applications. Recent efforts have targeted the synthesis of robust, efficient SLAM pipelines leveraging 3DGS as the underlying scene model, leading to the new subfield of “3DGS-SLAM” highlighted in this survey (2602.04251). The review provides a technical taxonomy of methods, benchmarks, and future challenges, with exhaustive coverage of optimization strategies in rendering quality, tracking accuracy, speed, memory, and robustness.

Figure 1: Evolution of SLAM: from early geometric models and probabilistic filtering to deep neural rendering and the 3DGS methodology.

Foundations of 3D Gaussian Splatting and Its SLAM Pipeline

3DGS operates through four algorithmic modules: point/primitive initialization, differentiable projection, parallel rasterization, and iterative scene optimization. Scene coverage is seeded from multi-view images and structure-from-motion-derived point clouds; each primitive is parameterized as a 3D anisotropic Gaussian with explicit position, opacity, spatial covariance, and spectral color. Differentiable projection transforms Gaussians from 3D world coordinates to 2D image space via camera extrinsics using frustum culling, affine transformation, and Jacobian mapping.

Rendering is accomplished by depth-sorted tile-wise rasterization backed by CUDA parallelization, where pixel color composition is governed by ordered alpha blending of projected Gaussians. The core optimization iteratively updates Gaussian parameters to minimize a photometric loss blending $L_1$ and multi-scale D-SSIM, with adaptive splitting and merging to maintain efficient density and fidelity.

Figure 2: 3DGS pipeline: initialization, projection, tile-wise rasterization, and differentiable optimization enabling highly photorealistic rendering from sparse input.

In SLAM integration, typical pipelines contain initialization, frame-to-model tracking, Gaussian mapping, and global loop closure. Upon camera entry, pose priors and dense per-pixel initialization build the starting Gaussian set. Pose estimation exploits color and depth residuals only in reliable regions, adding frames passing visibility checks to the keyframe queue. Gaussian mapping then densifies the representation using mask-based heuristics and multi-source depth, while joint local loss minimization refines geometry and appearance. Loop closure triggers global pose-graph or bundle adjustment on overlapping keyframes, with re-optimization of affected Gaussians for drift correction.

Figure 3: End-to-end 3DGS-SLAM pipeline: tracking, mapping, keyframe selection, and loop closure for global consistency and fidelity.

Performance Optimization: Strategies and Benchmark Results

Rendering Quality

Rendering fidelity in 3DGS-SLAM is constrained by sparse views, missing depth, and scale ambiguity. Five major categories address these issues:

• Hybrid Explicit-Implicit Representations: Combining explicit Gaussian primitives with implicit neural fields or volumetric priors ensures robust initialization and high-frequency detail restoration for sparsely observed/partially reconstructed regions.
• Vision-Guided Perception: Densification and placement guided by visual residuals, frequency analysis, and structural priors (e.g., Manhattan world, surfel anchors) optimize fine detail recovery.
• Depth-Guided Optimization: Multi-source depth fusion—with uncertainty weighting and MVS network integration—regulates geometry regularity, addresses textureless regions, and curtails artifacts caused by external depth noise.
• Progressive Training: Hierarchical coarse-to-fine pyramid-based optimization enables global convergence before local refinement, avoiding premature overfitting.
• Multi-Agent Collaboration: Distributed agent submaps are fused using optimized loop closure, visibility masks, and loss weighting, providing scalability for large-scale exploration tasks.

  Figure 4: Five rendering quality optimization strategies for 3DGS-SLAM systems targeting diverse environments and fidelity requirements.

Top-performing configurations (e.g., Gaussian-SLAM, DROID-Splat, VTGaussian-SLAM) report Replica dataset PSNR up to 43.34, SSIM reaching 0.996, and minimum LPIPS values of 0.012, starkly outperforming NeRF-based SLAM and prior explicit geometry methods on photorealistic benchmarks.

Tracking Accuracy

Pose estimation and map consistency are optimized via:

• Local Optimization: Joint optimization in constrained temporal or spatial windows leverages co-visibility, adaptive Gaussian management, and geometric priors.
• Global Pose-Graph Optimization: Loop closure detection (via ORB, NetVLAD, CLIP/DINOv2) with pose graph solvers enforces global trajectory consistency—robust under viewpoint shifts.
• Global Bundle Adjustment: Factor graph formalizations enable full photometric-geometric joint parameter updates, integrating scale and depth constraints and history optimization.

Across Replica, TUM, and ScanNet, leading approaches achieve sub-centimeter ATE RMSE, e.g., SplatMAP and GauS-SLAM at 0.18 and 0.06 cm, respectively.

Reconstruction Speed

Accelerated mapping is realized through hierarchical Gaussian initialization (using dense point cloud priors or Patch-Grid sampling), density management (dynamic pruning, hierarchical update scheduling), and parallel processing/hardware support (multi-thread, CUDA, custom accelerators). GPS-SLAM achieves 252.64 FPS on Replica with 37.24 dB PSNR, demonstrating the feasibility of deploying high-fidelity 3DGS-SLAM on mobile and embedded systems.

Memory Consumption

Memory overhead is mitigated by:

• Generation Control and Sparsification: Occupancy-based primitive spawning, geometric/image gradient placement, and loss-driven pruning.
• Hierarchical Map Decomposition: Scene partitioning into submaps with on-demand scheduling and aggressive fusion/compression.
• Compact Encoding: Dimensionality reduction (R-VQ, voxel anchoring, surfel abstraction) and parameter compression drive efficient storage.

  Figure 5: Memory optimization modules in 3DGS-SLAM: controlled generation, hierarchical decomposition, and compact encoding.

  

  Figure 6: Model size comparisons, Replica dataset—demonstrating paramount memory and storage efficiency.

MGSO compresses model size to 4.3 MB and MemGS achieves memory usage of 1.95 GB, proving scalable deployment even for city or building-scale mapping.

Robustness: Motion Blur and Dynamic Scene Adaptation

Robust operation under motion blur and dynamic scenes is critical for real-world deployment.

Motion Blur

Motion blur disrupts tracking and mapping, leading to floater artifacts and degraded accuracy. State-of-the-art solutions include robust system coupling (fusion bridges, residual masking) and explicit physical modeling (trajectory-aware rendering and multi-frame integration). MBA-SLAM models camera exposure trajectories for blur compensation, and Deblur-SLAM decomposes blurred frames into virtual subframes for recovery, enabling reliable tracking under adverse conditions.

Figure 7: Motion blur optimization via tight system integration and explicit blur modeling for stability in high-speed/low-light environments.

Dynamic Scenes

Handling moving objects is addressed via:

• Semantic Priors: VLMs and LLMs generate masks for dynamic object suppression; spatio-temporal refinements and geometric alignment further enhance mask accuracy.
• Geometric Consistency: Motion-based penalization and probabilistic modeling (e.g., CRFs, GGMs) enable dynamic detection without relying on object labels.
• Explicit Dynamic Modeling: Foreground-background decoupling and motion prediction modules allow full dynamic reconstruction rather than simple masking.
• Uncertainty Modeling: Feature-driven adaptive weighting suppresses unreliable regions during loss computation.

State-of-the-art algorithms (GARAD-SLAM, ADD-SLAM, WildGS-SLAM) attain ATE < 3 cm and STD < 1.05 cm on dynamic benchmarks (Bonn, TUM), with full dynamic scene rendering, challenging the conventional SLAM paradigm.

Implications and Prospective Developments

The survey details how 3DGS-SLAM advances the capabilities of SLAM in representation, optimization, robustness, and computational efficiency, removing fundamental trade-offs that restricted earlier approaches. This foundation has direct implications for robotics, AR/VR, autonomous systems, and digital twins. Future research directions include:

• Event-Camera Fusion: Incorporating microsecond-resolution event sensor data to overcome limitations in blur and dynamic scenes.
• Extreme Environment Adaptation: Multi-modal mapping and robust priors for hostile scenarios (low texture, occlusions, environmental interference).
• Physics-Aware Modeling: Integration of elasticity, friction, and non-rigid dynamics into Gaussian representations.
• Large Vision Model Integration: End-to-end learning of geometry and appearance via transformers and vision-language foundation models for self-supervised mapping generalizable to myriad environments.

Conclusion

This survey provides an exhaustive technical review of 3DGS-SLAM, benchmarking innovations in rendering, tracking, speed, memory, and robustness, and synthesizing their practical and theoretical implications. The seamless fusion of explicit 3DGS representation with advanced SLAM optimization strategies delivers high-fidelity photorealistic mapping in real-time, robust to blur and scene dynamics. Emerging directions in event-based sensing, physics integration, and transformer-guided mapping are poised to redefine SLAM capabilities for complex operational domains (2602.04251).