> For the complete documentation index, see [llms.txt](https://panav.gitbook.io/robotics-handbook/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://panav.gitbook.io/robotics-handbook/slam-and-state-estimation/slam-and-state-estimation.md).

# Overview

SLAM and state estimation - 2D/3D visual/LiDAR SLAM, learned SLAM, sensor fusion, filter and graph backends, benchmarking on real robots.

## What this section is

This is the authoritative SLAM reference in the handbook. The short page in `mobile-robotics/slam-and-navigation.md` is a primer - fine for a 10-minute overview. This section is what I wish I'd had when I started building SLAM systems: opinionated, math-grounded, and biased toward what actually works on real robots in 2026.

I built [GO-SLAM](/robotics-handbook/authors-projects/go-slam.md) from scratch in two months - GICP front-end, custom Levenberg-Marquardt solver, pose-graph back-end, loop closure - without leaning on Ceres, g2o, or GTSAM. I work on GPU-accelerated SLAM and multimodal EKF state estimation in production (2026). Most of the opinions here come from things that broke at 3 AM in a warehouse.

If you only read one thing: **SLAM is a state estimation problem with a map as a side effect.** Everything else - features, ICP, factor graphs, neural fields - is implementation detail layered on top of probabilistic state estimation.

### In this section

<table data-view="cards"><thead><tr><th></th><th data-type="content-ref"></th><th data-hidden data-card-cover data-type="files"></th></tr></thead><tbody><tr><td>Filter-Based SLAM</td><td><a href="/robotics-handbook/slam-and-state-estimation/filter-slam.md">Filter-Based SLAM</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2FzsgJ72dhFvp8P2xG7UZb%2FKalman_filter_animation%2C_1d.gif?alt=media&amp;token=6ee3a93b-605d-428c-95f5-113684ca7050">Kalman_filter_animation,_1d.gif</a></td></tr><tr><td>Graph SLAM</td><td><a href="/robotics-handbook/slam-and-state-estimation/graph-slam.md">Graph SLAM</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2F3U3lTxy2zGUIfX8IlhLs%2Fgraph-slam-pose-pose.gif?alt=media&amp;token=9b2df12d-255b-4b4c-ad33-ed0f516bc8e5">graph-slam-pose-pose.gif</a></td></tr><tr><td>Visual SLAM</td><td><a href="/robotics-handbook/slam-and-state-estimation/visual-slam.md">Visual SLAM</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2FTwM5Ny0yAbz1vtQEjdmo%2Fball-tracking-animated-02.gif?alt=media&amp;token=a8621238-89b8-4303-b14c-3908f88114bf">ball-tracking-animated-02.gif</a></td></tr><tr><td>LiDAR SLAM</td><td><a href="/robotics-handbook/slam-and-state-estimation/lidar-slam.md">LiDAR SLAM</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2FPEUId0zevO29SaqrcyZb%2FLIDARanim.gif?alt=media&amp;token=2664e350-9911-4bbc-936a-cf0ad3b13b31">LIDARanim.gif</a></td></tr><tr><td>Learned &#x26; Neural SLAM</td><td><a href="/robotics-handbook/slam-and-state-estimation/learned-slam.md">Learned &amp; Neural SLAM</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-5041dd346ec6e6878a0028d1f7327a9111e386e5%2Fneural-slam.gif?alt=media">neural-slam.gif</a></td></tr><tr><td>Sensor Fusion</td><td><a href="/robotics-handbook/slam-and-state-estimation/sensor-fusion.md">Sensor Fusion</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-05dbf7ed0f6b65ec210272845729773c12b32546%2Fsensor-fusion-ekf.gif?alt=media">sensor-fusion-ekf.gif</a></td></tr><tr><td>SLAM Evaluation &#x26; Benchmarks</td><td><a href="/robotics-handbook/slam-and-state-estimation/slam-evaluation.md">SLAM Evaluation &amp; Benchmarks</a></td><td><a href="https://923066007-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FJR2yBCOFnSaxfostbWYx%2Fuploads%2Fgit-blob-f6cf3f248c77ae09773bac2572daeedd564cb514%2Fkitti-dataset-viz.gif?alt=media">kitti-dataset-viz.gif</a></td></tr></tbody></table>

***

## SLAM taxonomy

SLAM literature is a swamp of acronyms. There are four orthogonal axes you can slice it on. Pick a point in this 4D space and you've picked an algorithm class.

### Axis 1: dimensionality - 2D vs 3D

| Dimension | Sensors                            | Typical algorithms                              | Use case                            |
| --------- | ---------------------------------- | ----------------------------------------------- | ----------------------------------- |
| 2D        | 2D LiDAR, wheel odometry           | gmapping, Cartographer (2D mode), slam\_toolbox | Indoor AMRs on flat floors          |
| 3D        | 3D LiDAR, stereo/RGB-D camera, IMU | LIO-SAM, FAST-LIO2, ORB-SLAM3                   | Drones, outdoor robots, multi-floor |

**Field notes:** 2D SLAM is *not* obsolete. If your robot lives on a flat warehouse floor, a Hokuyo URG-04LX and slam\_toolbox will outperform a 64-beam Ouster running FAST-LIO2 every time on compute-per-dollar. Reach for 3D only when 2D fails you.

### Axis 2: estimator - filter vs graph

| Type       | What it is                                                                           | Strength                           | Weakness                                     |
| ---------- | ------------------------------------------------------------------------------------ | ---------------------------------- | -------------------------------------------- |
| **Filter** | Recursive Bayes filter (EKF, UKF, particle filter). Online, marginalizes past poses. | Constant per-step cost, online     | No revisiting; linearization errors compound |
| **Graph**  | Build a factor graph of poses + landmarks, batch-optimize when needed.               | Loop closure trivially fixes drift | Need to re-solve when graph changes          |

Modern SLAM is mostly graph-based. Filters survive in two niches: AMCL for pure localization on a known map, and tight IMU integration where the IMU is the highest-frequency signal (you'll see "IEKF" or "ESKF" in VIO papers).

Details: [filter-slam.md](/robotics-handbook/slam-and-state-estimation/filter-slam.md), [graph-slam.md](/robotics-handbook/slam-and-state-estimation/graph-slam.md).

### Axis 3: representation - dense vs sparse

| Type           | Map representation                     | Algorithms                                    | Trade-off                     |
| -------------- | -------------------------------------- | --------------------------------------------- | ----------------------------- |
| **Sparse**     | Keyframes + 3D point landmarks         | ORB-SLAM3, VINS-Fusion, FAST-LIO2 (point map) | Fast, scales well, no surface |
| **Semi-dense** | Pixels with strong gradient + depth    | LSD-SLAM, DSO                                 | Middle ground                 |
| **Dense**      | Every pixel / every voxel has geometry | KinectFusion, NICE-SLAM, Gaussian-Splat SLAM  | Beautiful maps, GPU-bound     |

Sparse wins for navigation. Dense wins for AR, mapping-as-product, or anything that needs photoreal reconstruction.

### Axis 4: cadence - online vs batch

* **Online:** every-frame estimation with bounded latency. The only mode that matters for a robot deciding where to go right now.
* **Batch:** post-process the whole log. Useful for ground-truth-grade maps, calibration, dataset construction. SfM (Structure from Motion) is batch SLAM.

***

## Which SLAM do I use? - decision tree

```
Is your robot indoors on a flat floor?
├── Yes → 2D LiDAR + slam_toolbox. Stop reading.
└── No  → Continue.

Do you have GPS that works (outdoor, sky-visible)?
├── Yes → GPS + IMU + wheel odom in an EKF (robot_localization). You don't need SLAM, you need state estimation.
└── No  → Continue.

What's your dominant sensor?
├── 3D LiDAR + IMU
│   └── Real-time, CPU-bound? → FAST-LIO2 (ikd-tree, IEKF, no loop closure)
│       Robust loop closure mandatory? → LIO-SAM (factor-graph back-end)
│       Simple, no IMU? → KISS-ICP
├── Camera only (monocular/stereo/RGB-D)
│   └── Need scale (mono)? → Add IMU → VINS-Mono / VINS-Fusion
│       Have stereo or RGB-D? → ORB-SLAM3
│       Indoor reconstruction priority? → SplaTAM / Gaussian-Splat SLAM
└── Camera + IMU + LiDAR (multi-modal)
    └── LIO-SAM + visual loop closure, or roll your own graph (this is where research lives).
```

I don't include "EKF-SLAM" in the tree because in 2026 you should not use EKF-SLAM for a new system. It's a teaching tool. See [filter-slam.md](/robotics-handbook/slam-and-state-estimation/filter-slam.md) for why.

***

## The subpages

| File                                                                                  | What's in it                                                                                                             |
| ------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------ |
| [filter-slam.md](/robotics-handbook/slam-and-state-estimation/filter-slam.md)         | EKF-SLAM, particle filter / FastSLAM, AMCL. Bayes filter intuition, when particles collapse.                             |
| [graph-slam.md](/robotics-handbook/slam-and-state-estimation/graph-slam.md)           | Pose graph optimization, factor graphs, g2o, GTSAM, Ceres. Information matrix, Schur complement, marginalization.        |
| [visual-slam.md](/robotics-handbook/slam-and-state-estimation/visual-slam.md)         | ORB-SLAM3, VINS-Mono/Fusion, DSO/LDSO. Monocular vs stereo vs RGB-D. Feature-based vs direct.                            |
| [lidar-slam.md](/robotics-handbook/slam-and-state-estimation/lidar-slam.md)           | LOAM → LIO-SAM → FAST-LIO2 → KISS-ICP. Deskewing, ICP variants, IMU preintegration.                                      |
| [learned-slam.md](/robotics-handbook/slam-and-state-estimation/learned-slam.md)       | NICE-SLAM, the ICCV 2023 GO-SLAM, GS-SLAM, SplaTAM, Gaussian-Splat SLAM. (Not Pan's GO-SLAM - see disambig note inside.) |
| [sensor-fusion.md](/robotics-handbook/slam-and-state-estimation/sensor-fusion.md)     | EKF/UKF, ICP variants, IMU preintegration, robot\_localization. When to fuse what.                                       |
| [slam-evaluation.md](/robotics-handbook/slam-and-state-estimation/slam-evaluation.md) | KITTI, TUM RGB-D, EuRoC, Newer College, Hilti. ATE / RPE / evo. Benchmarking pitfalls.                                   |

***

## What the textbooks don't tell you

A few opinions I've earned the hard way:

1. **The hardest part of SLAM is not SLAM.** It's TF trees, clock synchronization, sensor calibration, deskewing, and outlier handling. Get those right and a mediocre back-end will look great. Get them wrong and the best back-end will diverge.
2. **Loop closure is mostly about descriptors, not optimization.** g2o vs GTSAM vs Ceres for the back-end barely matters - they all converge to the same minimum if your data association is correct. Place recognition (BoW, NetVLAD, Scan Context) is where most loop closure failures live.
3. **IMU preintegration (Forster et al. 2015/2017) changed everything.** Before it, tight visual-inertial coupling was a research curiosity. After it, every modern VIO and LIO system uses it. If you don't know what it is, read [sensor-fusion.md](/robotics-handbook/slam-and-state-estimation/sensor-fusion.md) before anything else.
4. **GPU SLAM is the present, not the future.** Gaussian splatting, neural fields, GPU ikd-trees - by 2026 the question is no longer "should I use GPU?" but "which parts run where?" This is what my current production work focuses on.
5. **Benchmark on your own data.** KITTI is solved. EuRoC is solved. Your warehouse with reflective floors, dynamic forklifts, and a tilted dock leveler is not. See [slam-evaluation.md](/robotics-handbook/slam-and-state-estimation/slam-evaluation.md).

***

## Canonical references

These are the books and surveys I keep open:

* **Probabilistic Robotics** - Thrun, Burgard, Fox (2005). Still the foundation. Chapters 9-15 cover EKF, particle filter, and GraphSLAM.
* **Factor Graphs for Robot Perception** - Dellaert & Kaess (2017). [Free PDF from author](https://www.cs.cmu.edu/~kaess/pub/Dellaert17fnt.pdf). The mental model for modern SLAM back-ends.
* **State Estimation for Robotics** - Tim Barfoot (2nd ed, 2024). Lie groups, on-manifold optimization, the math under GTSAM.
* **Past, Present, and Future of SLAM - Cadena et al. (2016)** - [arxiv.org/abs/1606.05830](https://arxiv.org/abs/1606.05830). The "where is SLAM going" survey everyone cites.
* **Visual SLAM survey** - Macario Barros et al. (2022). [A Comprehensive Survey of Visual SLAM Algorithms](https://www.mdpi.com/2218-6581/11/1/24).

***

## Cross-references in this handbook

* [Polka](/robotics-handbook/authors-projects/polka.md) - my multi-LiDAR fusion + deskewing node. Deskewing is covered in depth in [lidar-slam.md](/robotics-handbook/slam-and-state-estimation/lidar-slam.md).
* [GO-SLAM (Pan's)](/robotics-handbook/authors-projects/go-slam.md) - my from-scratch GICP + pose-graph SLAM. Linked from [lidar-slam.md](/robotics-handbook/slam-and-state-estimation/lidar-slam.md) and called out in [learned-slam.md](/robotics-handbook/slam-and-state-estimation/learned-slam.md) to disambiguate from the ICCV 2023 paper of the same name.
* [Cameras, Depth Sensors and LiDAR](/robotics-handbook/perception-and-computer-vision/cameras-depth-sensors-and-lidar.md) - sensor-level details that SLAM builds on.
* [Linear Algebra for Robotics](/robotics-handbook/foundations/linear-algebra-for-robotics.md) - SE(3), rotations, the linear algebra under every estimator on this page.