SLAM (Simultaneous Localization and Mapping) is the technology behind estimating "where am I" and "what does my surroundings look like" at the same time, using nothing but sensors. This article surveys the broader technical currents across SLAM as a whole; the specifics of individual fields (Visual-SLAM, LiDAR-SLAM) are covered in their own articles.

ROS 2ROS 2

Image: Robot Operating System logo, Wikimedia Commons (CC BY-SA 4.0)

The First Current: Starting from Filter-Based Methods

Early SLAM was dominated by sequential estimation via the Extended Kalman Filter (EKF) or particle filters — updating state every time a new sensor observation arrived. Computationally cheap, but with a weakness: error accumulates easily, and the approach tends to break down in large-scale maps.

The Second Current: The Shift to Graph Optimization

What became dominant after that was pose-graph optimization (graph SLAM). It builds a graph where each node is the robot's pose at some point along its trajectory, and each edge is a relative relationship derived from sensor observations, then optimizes the whole thing at once as a nonlinear least-squares problem. Combined with loop closure (recognizing a previously visited place and retroactively correcting accumulated error), this made it possible to build maps that hold up even in large-scale environments. slam_toolbox and FAST-LIO, both used in newbot, belong to this lineage.

What's Inside Graph Optimization: Formulated as Nonlinear Least Squares

Looking a bit more concretely at the problem graph SLAM is actually solving internally: it builds a graph where each node is the robot's pose at a point in time along its trajectory, and each edge is the relative positional relationship derived from sensor observations, and it's formulated as a nonlinear least-squares problem that adjusts every node's placement to be maximally consistent with every edge's constraint. Because sensor observations carry noise, the constraints between edges are bound to conflict with each other to some degree, and the role of this optimization is to distribute that conflict in the "least unreasonable" way possible.

Formally, for the full set of poses to be solved for, \mathbf{x} = (\mathbf{x}_1, \dots, \mathbf{x}_n), the problem is to minimize the following objective.

\mathbf{x}^{*} = \arg\min_{\mathbf{x}} \sum_{(i,j) \in \mathcal{C}} \left\| \mathbf{e}_{ij}(\mathbf{x}_i, \mathbf{x}_j) \right\|^2_{\Sigma_{ij}^{-1}}

Here \mathcal{C} is the set of node pairs (edges) constrained by a sensor observation, \mathbf{e}_{ij} is the error function expressing the mismatch between node i and j's estimated poses and the sensor observation, and \Sigma_{ij}^{-1} is the weighting from that observation's confidence (the inverse covariance).

Iterative methods like Gauss-Newton or Levenberg-Marquardt are used to solve it. The error function is linearized around the current estimate (\mathbf{e}_{ij}(\mathbf{x} + \Delta\mathbf{x}) \approx \mathbf{e}_{ij}(\mathbf{x}) + \mathbf{J}_{ij}\Delta\mathbf{x}), and the normal equation \mathbf{J}^\top \mathbf{J} \, \Delta\mathbf{x} = -\mathbf{J}^\top \mathbf{e} is solved for the update \Delta\mathbf{x}, repeated until convergence. When the 3D positions of observed feature points (landmarks) are included as optimization targets alongside camera poses, this process specifically goes by the name Bundle Adjustment. SLAM problems have a large number of variables (poses, landmarks), but each individual observation only touches a small handful of them, so the resulting matrices (Jacobian, Hessian) come out sparse. Solving efficiently by exploiting that sparsity is the practical key here, and general-purpose optimization frameworks like Ceres Solver, g2o, and GTSAM are widely used as libraries for exactly this kind of sparse nonlinear least-squares problem.

The Third Current: Learning-Based, AI-Native SLAM

The current in progress right now is folding deep learning into the SLAM pipeline. Replacing explicit feature extraction and matching with deep learning is an active area of research aimed at improving robustness to lighting changes and texture-poor environments. More recently, a direction called "AI-native" or "semantic SLAM" has also emerged — integrating graph neural networks, or even large language models (LLMs) and Vision-Language-Action (VLA) models, into the SLAM pipeline. The goal isn't just "where am I" anymore, but building a semantic map that simultaneously understands "what's there."

A Shift in the Map Representation Itself

The most conspicuous recent change is a reinvention of the map's own "representation format." Beyond the traditional point cloud and occupancy grid, map representations built on Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) are rapidly gaining ground. These are increasingly outperforming traditional methods on mapping accuracy, rendering quality, and computational efficiency, and new 3DGS/NeRF-based methods — MHED-SLAM, GTS-SLAM, MTE-SLAM, HL-SLAM, PMET-SLAM, and more — have continued to appear at major conferences and journals well into 2026. See the separate article "Technology Trends in Visual-SLAM" for the details.

Late-2026 Snapshot: 3DGS-SLAM Diversifying by Domain

Looking concretely at the 3DGS/NeRF-based methods mentioned above through August 2026 publications, what stands out is how much the target domains have diversified. RoSe-SLAM builds semantically-tagged 3D Gaussian maps from monocular video in dynamic scenes (accepted to IROS 2026); Stipple, from Daniel Cremers' lab, incrementally builds a map in real time while tightly coupling visual and inertial data; GLAM-SLAM scales to city-block-sized environments through flow densification and spatial decomposition (also accepted to IROS 2026); and Real-Time LiDAR Gaussian Splatting SLAM combines LiDAR point clouds directly with 3D Gaussians. Specialization is proceeding along distinct axes — indoor vs. outdoor, dynamic scenes, large-scale environments. The approach is also spreading beyond robotics into medicine: EndoMD-SLAM targets endoscopic SLAM, and Track2Map (accepted to MICCAI 2026) targets laparoscopic surgery.

Integrating Semantic Understanding: Open-Vocabulary and Foundation-Model-Coupled SLAM

Digging further into the "learning-based, AI-native SLAM" current: 2026 has been the year multiple methods emerged that wire vision foundation model features directly into the SLAM pipeline. RADIO-ViPE (April 2026) tightly couples multi-modal embeddings from an agglomerative vision foundation model online, achieving open-vocabulary semantic SLAM that can localize arbitrary text-specified categories on the map even in dynamic environments. FeatureSLAM (January 2026) rasterizes features aligned with a vision foundation model onto each Gaussian in a 3D Gaussian Splatting map, building a semantically searchable map while preserving real-time performance. This direction is laid out systematically in the October 2025 survey "Semantic Visual SLAM: A Survey on State of the Art, Challenges, and Future Directions," which extends its scope all the way to large language model integration — underscoring how SLAM is shifting from pure geometric estimation toward systems that also carry semantic understanding.

Broadening Sensor Modalities: 4D Radar as an Option

Alongside LiDAR and cameras, research using 4D radar (millimeter-wave radar that provides range, azimuth, elevation, and Doppler velocity) as a third sensor modality — prized for its resilience to bad weather — continues to be actively updated. 4D Radar-Inertial Odometry, which combines 3D Gaussian modeling with multi-hypothesis scan matching, targets improved odometry accuracy in fog, rain, and dust conditions where cameras and LiDAR both struggle; first published in late 2024, it received a revised version as recently as March 2026. Benchmark datasets are also maturing toward more realistic operating conditions, as with the Hilti-Trimble-Oxford dataset (July 2026), which uses a 360-degree camera plus IMU and incorporates floor-plan priors.

Regional Patterns in Industry and Research

Analysis of SLAM-related patents and papers has pointed to China carrying a disproportionate weight in LiDAR-centric 2D SLAM, ROS-based implementations, graph optimization, and multi-robot/edge-computing research. Demand from practical applications — drones, service robots — appears to be strongly steering the direction of research.

Summary: The Three Currents Aren't Mutually Exclusive

The three currents — filter-based, graph optimization, learning-based — aren't a case of older methods being replaced by newer ones over time; today they coexist and blend together. FAST-LIO, used in newbot for instance, belongs to the lineage of direct methods, distinct from graph optimization, while at the same time drawing on learning-based feature extraction and 3DGS-based map representation happening around it. Which approach to choose depends on available compute, sensor configuration, and what the map is actually needed for — that's the practical conclusion at this point in time.

References

#SLAM #Deep Learning