LiDAR SLAM from Scratch Scan Matching to Loop Closure

Build a 2D SLAM stack in Python and ROS 2, from ICP to pose-graph optimisation, measuring every claim as you go
What you'll learn

---

Implement ICP from scratch: correspondences, the SE(2) best-fit transform, convergence thresholds and what the residual really tells you
Write point-to-line ICP and surface normal estimation, and measure exactly when it beats point-to-point and when it does not
Build correlative scan matching with a coarse-to-fine search, and see why a search step finer than the grid is a lie
Construct an occupancy grid with log-odds updates and Bresenham ray tracing, and publish it as a real nav_msgs/OccupancyGrid
Detect loop closures end to end: scan descriptors, candidate retrieval, spatial filtering, geometric verification and false-positive rejection
Derive and code an SE(2) pose-graph optimiser with analytic Jacobians, gauge anchoring and robust weighting for bad loops
Diagnose a SLAM frontend: odometry prediction, scan correction, keyframe selection and where drift actually comes from
Measure your own system honestly with ATE, RPE and drift metrics, and compare it against an established ROS 2 SLAM package
Recognise when a benchmark is measuring your own bug rather than the algorithm, which happened repeatedly while building this course
Requirements

---

Comfortable Python: functions, classes, numpy arrays, reading a traceback
Basic linear algebra: vectors, matrices, and what a rotation matrix does. The SE(2) maths is derived from scratch
Helpful but not required: some ROS 2 exposure. The nodes are built step by step and the frames are explained
No LiDAR hardware needed. Everything runs in Gazebo, and the full simulated environment is provided
A Linux machine with ROS 2 Jazzy. Setup is covered in Section 1
Description

---

This course contains the use of artificial intelligence.
Most SLAM courses show you a map that closes beautifully and move on. This one builds the whole stack from an empty file and then tries to break every part of it, because the interesting engineering is in the failures.
You will write point-to-point ICP, point-to-line ICP, correlative scan matching, an occupancy grid, a loop-closure detector and an SE(2) pose-graph optimiser with analytic Jacobians. Everything runs against a simulated Clearpath Husky with a SICK LMS1xx in a Gazebo warehouse, so every number in the course came from a run you can reproduce.
What makes this different is that the measurements are allowed to disagree with the textbook. A few of the results you will derive
Point-to-line ICP doesnot rescue a degenerate wall. Its advantage is geometry-dependent, not universal.
"Log odds avoids numerical underflow" is false. The real reasons to use it are better than the one everyone repeats.
Submapslose on a six metre path, and understanding the regime where they lose is the actual lesson.
Subsampling a scan to 4 cm ismore accurate than keeping every point.
A standard outlier filter made matching 1.9x worse by deleting the far wall.
Residual and inlier fraction donot separate a true loop closure from a false one. The quantity that does is the one the aligner never saw.
Several of those started as bugs in my own harness. The course shows the broken version, the number that looked impressive, and how the mistake was caught, because recognising a measurement that is lying to you is the skill that transfers.
You finish with a working SLAM system, a test suite, and the habit of asking what a number actually measured.
Who this course is for

---

Robotics engineers who can run an existing SLAM package but could not write one, or debug it when the map folds
ML and software engineers moving into robotics who want the geometry and estimation rather than another library tour
Students and researchers who need to implement scan matching or pose-graph optimisation and want the derivations alongside working code
Anyone who has watched a SLAM demo close a loop perfectly and wondered what happens when it does not
Homepage

---