Gaussian Splatting: Unlocking New Dimensions in 3D Reconstruction, Robotics, and Real-Time Media

Latest 41 papers on gaussian splatting: Sep. 19, 2026

Gaussian Splatting (3DGS) has rapidly emerged as a powerhouse in 3D computer vision, offering a compelling blend of photorealism, efficiency, and flexibility for representing complex scenes. Its explicit, point-based representation, unlike implicit neural fields (NeRFs), allows for real-time rendering while maintaining high fidelity. This collection of recent research highlights how 3DGS is not just evolving, but fundamentally transforming various domains, from dynamic scene understanding and robotic navigation to advanced rendering and even medical imaging. The breakthroughs showcased here address core challenges, pushing the boundaries of what’s possible with this exciting technology.

The Big Idea(s) & Core Innovations

At its heart, the recent advancements in 3DGS revolve around enhancing its robustness, efficiency, and applicability across diverse and challenging scenarios. A major theme is improving geometry and appearance quality under difficult conditions, such as sparse views, extreme lighting, or reflective surfaces. For instance, RGS: Reflection-aware Gaussian Splatting via Learning Geometry Continuity for Reflective Objects from the University of Technology Sydney and Adelaide University, tackles the pervasive “surface collapse” problem in reflective regions. Their novel use of a VGGT 3D foundation model for geometric priors and a reflection-guided densification strategy ensures smoother surfaces and better specular rendering. This is crucial as traditional 3DGS often struggles with such challenging materials.

Similarly, NOVA-GS: Noise-Aware View-Consistent Gaussian Splatting for Low-Light Novel View Synthesis by the Indian Institute of Technology Madras, tackles low-light conditions by unifying image enhancement, self-supervised denoising, and geometry optimization. Their noise-guided spherical harmonic regularization effectively suppresses view-dependent artifacts in noisy regions, leading to robust novel view synthesis where traditional methods fail.

Another critical area of innovation is making 3DGS more adaptable and efficient, particularly for real-time and resource-constrained applications. PhGS: Post-Hoc Pruning and Refinement of Single-View Feed-Forward 3D Gaussian Reconstructions from Keio University and the University of Stuttgart, introduces a backbone-agnostic post-hoc pruning and refinement method that significantly reduces spatial redundancy in single-view feed-forward 3DGS. This directly translates to lower memory consumption and faster rendering without sacrificing quality, making 3DGS viable for more applications.

EliGSiR: Continual RGB-D Mapping with Gaussian Splatting under Bounded Compute by Technical University of Munich, addresses the challenge of online 3D reconstruction from continuous RGB-D streams under strict computational budgets. Their adaptive allocation of optimization budget across view scheduling, supervision fidelity, and geometry growth results in higher-quality reconstructions with less compute. This is vital for real-time robotic applications.

RawSLAM: Online HDR Gaussian SLAM from Linear Radiance from Universidad Pablo de Olavide, introduces the first online Gaussian SLAM framework that directly processes 16-bit linear HDR imagery. By employing an MLP-free logarithmic parameterization for Gaussian color features and an HDR-aware photometric loss, RawSLAM significantly improves tracking accuracy under extreme lighting, a critical step for robust SLAM in diverse environments.

Beyond single scene reconstruction, researchers are pushing 3DGS towards dynamic, multi-agent, and specialized applications. SplashSplat: Reconstructing Splashing Liquids from Real-World Multi-View Videos by EPFL, ETH, Google, and Microsoft, offers a groundbreaking method for reconstructing dynamic liquid surfaces. By imposing physical structure only where observations constrain it and using Lagrangian carriers, SplashSplat outperforms state-of-the-art dynamic 3DGS methods, opening doors for realistic fluid simulation and rendering.

For collaborative robotics, CoRef-GS: Cooperative Referring Gaussian Splatting for Multi-Agent Scene Understanding from Hunan University and Karlsruhe Institute of Technology, enables multiple robots to build and fuse local semantic Gaussian maps for cooperative referring scene understanding. This allows robots to ground natural language commands over merged maps, significantly advancing embodied AI.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are often underpinned by novel technical components and rigorous evaluation on new or enhanced datasets:

• SplashSplat (https://niko-creater.github.io/splashsplat-web/) introduces the first synchronized multi-view video benchmark of splashing liquids, alongside their physics-informed method using Lagrangian carriers.
• PhGS leverages existing architectures like Flash3D, Niagara, and SHARP, and uses importance-score-based pruning with a lightweight recurrent refinement module, evaluated on datasets like RealEstate10K.
• RawSLAM (https://arxiv.org/pdf/2609.20589) contributes the novel RawSLAM dataset with 10 real-world indoor sequences featuring 16-bit RAW imagery, depth, IMU, and OptiTrack ground-truth poses.
• CoRef-GS (https://github.com/ruojiruoli17/CoRef-GS.git) provides the CoQuad-Ref benchmark for cooperative referring segmentation in dual-quadruped settings, with paired real-world and simulated scenes.
• EliGSiR (https://eligsir-project.github.io) evaluates on Replica, TUM RGB-D, and ScanNet++ datasets, integrating 3DGS-MCMC for Gaussian population refinement.
• GS-PI: An Optimization-Decoupled Appearance Decomposition Approach for Generating PBR Gaussian Assets (https://arxiv.org/pdf/2609.19907) uses a conditional point-cloud diffusion transformer for PBR material decomposition, tested on datasets like NeRF Synthetic, Objaverse, and Google Scanned Objects (GSO).
• VGGT-GS SLAM: Uncalibrated Monocular Gaussian Splatting SLAM with Feed-Forward Priors (https://arxiv.org/pdf/2609.19628) employs a generalizable backend improving multiple feed-forward geometry priors (VGGT, DA3-Large, Fast3R) on TUM-RGBD, 7-Scenes, ScanNet, and Replica.
• SLAMSqueezeBench: Comparing SLAM Systems under Resource Constraints (https://github.com/sfu-rsl/SLAMAdversarialLab) introduces a benchmarking framework for evaluating SLAM systems under CPU/memory constraints, finding that Gaussian Splatting systems generally fail under minimal embedded hardware constraints.
• ParticleSplat: Self-supervised Object-centric Latent Particle Splatting (https://lyuxinghe.github.io/ParticleSplat-website/) extends Deep Latent Particles (DLP) to 3D for self-supervised object-centric scene decomposition, demonstrating improved performance on robotic manipulation tasks.
• RGS (https://xiaobiaodu.github.io/reflectivegs/) uses VGGT 3D foundation model priors and evaluates on Shiny Blender, Glossy Synthetic, and Ref-NeRF Real datasets.
• ArtNVG: Content-Style Separated Artistic Neighboring-View Gaussian Stylization (https://arxiv.org/pdf/2412.18783) utilizes CSGO and Tile ControlNet with Stable Diffusion XL v1.0 for 3D stylization on Tanks and Temples and WikiArt datasets.
• Geometry beneath the Waves: Dense Priors for Sparse-View Underwater 3D Gaussian Splatting (https://arxiv.org/pdf/2609.18737) fine-tunes VGGT with LoRA adapters using synthetically degraded underwater images, improving RUSplatting on SeaThru-NeRF and Submerged3D.
• MoQSplat: Adaptive Progressive Streaming of 3D Gaussian Splatting via MoQ (https://github.com/emanuele-artioli/MoQSplat) is built on the moq-lite Pub/Sub framework and PyTorch-based gsplat renderer, enabling adaptive streaming of 3DGS content.
• CADSplat: Sparse-View 3D Gaussian Splatting Aided by CAD Models for Robust, Photorealistic Digital-Twin Reconstruction (https://arxiv.org/pdf/2609.18473) regularizes 3DGS with explicit CAD shape priors, evaluated with a new SCO-CAD dataset (to be released).
• Wind on Trees: Testing Physical Grounding in Dynamic 4D Gaussian Splatting (https://arxiv.org/pdf/2609.17810) presents a synthetic testbed with procedurally generated trees and a physically parameterized deformation prior using damped harmonic oscillators.
• SVRecon: Generalizable Neural Reconstruction of High-Fidelity Surfaces via Sparse Volumetric Representations (https://arxiv.org/abs/2507.05952) emphasizes the importance of 3D sampling resolution, achieving 512^3 reconstruction through learned occupancy-driven sparsity.
• Talking Head Synthesis with Facial Landmark Guidance via 3D Gaussian Splatting (https://arxiv.org/pdf/2609.17422) uses public video datasets (Lieu, Obama, May speakers) and pretrained audio-to-keypoint networks (Geneface, Sapiens).
• PanoGS-SLAM: Panoramic 3D Gaussian Splatting SLAM (https://github.com/guoxf304/SynPano-Dataset) introduces the SynPano synthetic panoramic dataset and evaluates on PALVIO for real panoramic aerial sequences.
• Bi-FlowGS: Bridging Generative View Completion and Gaussian Geometry through Bidirectional Flow Co-Refinement (https://arxiv.org/pdf/2609.17039) addresses “Geometry Cheating” in sparse-view 3DGS using optical flow, evaluated on DL3DV-10K, Mip-NeRF 360, Tanks and Temples, and CO3D.
• HLC-GS: Risk-Map-Guided Height-Layer Consistency Gaussian Splatting for DSM Reconstruction from Optical Satellite Imagery (https://arxiv.org/pdf/2609.16772) uses DFC2019 and IARPA2016 benchmarks for Digital Surface Model (DSM) reconstruction from satellite imagery.
• The Neverwhere Visual Parkour Benchmark Suite (https://ziyc.github.io/neverwhere-bench/) offers 60+ hyper-photorealistic 3D Gaussian environments for evaluating visual locomotion policies on legged robots.
• Racing in Volume with Flow Ensembles (https://humansensinglab.github.io/monaco4d) presents Monaco4D, a photorealistic Unreal Engine 5 benchmark for high-speed outdoor reconstruction, along with FastFlowGS for streaming 4D Gaussian reconstruction.
• SceneBench: A Hierarchical Benchmark for Vision-Language Understanding of 3D Scenes (https://github.com/kangxiaou/SceneSplat) uses 966 photorealistic 3DGS scenes (ScanNet++, InteriorGS) with hierarchical annotations for vision-language models.
• EditBench3D: What Makes a 3D Scene Editable? A Factorized Benchmark of Fidelity, Locality, Consistency, and Preservation (https://arxiv.org/pdf/2609.14899) introduces a representation-agnostic benchmark for neural 3D scene editing, evaluating eight representative editors across 240 scene-edit pairs.
• SCOUT-SLAM: Structurally-Coupled Dual Uncertainty-Aware 3DGS SLAM in the Wild (https://github.com/kumaran-3527/SCOUT-SLAM) employs a shared base network (MUNet) and Tracking Uncertainty Adapter (TUA) for dynamic benchmarks (TUM RGB-D, Bonn Dynamic, Wild-SLAM MoCap).
• Deformable 2D Gaussian Splatting for Efficient 4K Video Compression (https://arxiv.org/pdf/2609.14129) utilizes multi-scale 2DGS with a deformation network for 4K video compression, evaluated on the UVG dataset.
• SkyAnchor: Updating Metric-scale Aerial 3D Gaussian Scenes from Unposed Ground-View Sequences (https://arxiv.org/pdf/2609.13903) uses AerialMegaDepth, ISPRS, HorizonGS, and ALoG-Uni10k datasets for aerial-ground reconstruction.
• Is Semantic SLAM ready for embedded systems? A comparative survey (https://arxiv.org/pdf/2505.12384) benchmarks six representative SLAM systems on NVIDIA Jetson AGX Orin for embedded deployment.
• Is Gaussian Splatting Becoming Neural Again? A Taxonomy and Controlled Study of Learned Parameterization (https://arxiv.org/pdf/2609.12395) provides a five-axis taxonomy and controlled experiments on mip-NeRF 360 to analyze neural integration in 3DGS.
• VS-Splat: Voxel-Selective feed-forward Gaussian Splatting for end-to-end 3D object reconstruction from sparse-views (https://arxiv.org/pdf/2609.12343) introduces learnable voxel selection with Gumbel-Sigmoid, evaluated on GObjaverse, GSO, and CO3D datasets.
• ATLAS Navigator: Active Task-driven LAnguage-embedded Gaussian Splatting (https://ongdexter.github.io/atlasnav) presents a memory-efficient online language-embedded 3DGS framework for robot navigation, validated with real-world experiments.
• Hologram Representation via Quadratic Phase Gaussian Splatting (https://arxiv.org/pdf/2609.11434) introduces Complex-Valued Quadratic Phase Gaussian (CVQPG) for hologram representation, evaluated on DIV2K and Real Forward-Facing datasets.
• Tri-DehazeGS: Scene–Medium Decoupled Gaussian Splatting with Transmittance-Aware Optimization (https://github.com/aptx46/Tri-DehazeGS) uses a tri-plane fog field and Medium-Decoupled Transmittance Gradient Compensation for dehazing, evaluated on RealX3D, Mip-NeRF 360, and Fog-NeRF.
• RIDE: Relocalization-Informed Depth Estimation with 3D Gaussian Splatting (https://arxiv.org/pdf/2609.11079) derives sparse metric depth observations from PnP-RANSAC inlier correspondences during 3DGS relocalization.
• Shape-guided Gaussian Splatting for Sparse-View X-ray 3D Reconstruction (https://github.com/polyshape-lab/ShapeGuidedGaussian) leverages population shape and density priors from the NMDID CT scan database for X-ray 3D reconstruction.
• View-Structured Conformal Prediction for 3D Gaussian Splatting (https://arxiv.org/pdf/2609.10307) introduces VSCP for 3DGS to provide finite-sample valid RGB prediction boxes, evaluated on Tanks & Temples, Deep Blending, and Mip-NeRF 360.
• LinearMask-GS: Stable-Mask Importance Pruning for Compact 3D Gaussian Splatting (https://arxiv.org/pdf/2609.10095) proposes linear increment masking activation for reliable pruning, demonstrating generality across multiple Gaussian-based backbones.
• RouteBridge: Reliability-Routed Bidirectional Distillation Between Neural Radiance Fields and 3D Gaussian Splatting (https://arxiv.org/pdf/2609.09606) introduces a bidirectional distillation framework for NeRF and 3DGS using a reliability estimator, evaluated on mip-NeRF 360.

Impact & The Road Ahead

The collective impact of this research is profound. 3DGS is rapidly maturing from a novel rendering technique into a versatile tool for 3D AI. The ability to reconstruct dynamic scenes like splashing liquids with physical plausibility opens doors for advanced visual effects, robotics in dynamic environments, and even scientific visualization. The strides in sparse-view reconstruction, driven by methods like CADSplat and shape-guided approaches for X-ray imaging, demonstrate 3DGS’s potential in critical applications with limited data, from digital twins in manufacturing to medical diagnostics.

The focus on efficiency and scalability, as seen in EliGSiR and PhGS, is crucial for deploying 3DGS on edge devices and enabling real-time applications like AR/VR, autonomous navigation, and even 4K video compression (Deformable 2D Gaussian Splatting). However, the “Is Semantic SLAM ready for embedded systems?” survey by Université Paris-Saclay, CEA and U2IS, ENSTA, highlights that while promising, 3DGS-based SLAM still faces significant resource constraints for embedded deployment, pointing to a need for continued hardware-algorithm co-design.

The development of benchmarks like Neverwhere and SceneBench, along with specialized evaluation frameworks like EditBench3D, is essential for robust development, pushing models to handle complex semantic reasoning and real-world uncertainties. The exploration of “neuralization” in 3DGS, as taxonomized by YuanHang Wang et al. from the University of Technology Sydney, reveals a nuanced path forward: leveraging neural networks for shared attributes while preserving the explicit flexibility of Gaussians for local geometric control.

Looking ahead, we can anticipate further integration of physical priors into 3DGS for even more robust dynamic scene understanding, bridging the gap between perception and simulation. The synergy between 3DGS and large generative models, particularly in view completion and content creation (ArtNVG, Bi-FlowGS), will unlock unprecedented creative possibilities. Moreover, the focus on uncertainty quantification (VSCP) and robust, uncertainty-aware SLAM (SCOUT-SLAM) will be paramount for safety-critical applications in robotics and autonomous systems. Gaussian Splatting isn’t just a trend; it’s a foundational shift, continually redefining the landscape of 3D vision and its real-world applications.

Please leave this field empty

Hi there 👋

Get a roundup of the latest AI paper digests in a quick, clean weekly email.

Check your inbox or spam folder to confirm your subscription.