Instructor: Aneesh
Date: June 5, 2025
📖 Overview
Today’s session introduced the foundations of Simultaneous Localization and Mapping (SLAM) — the core capability that allows robots to build a map of an unknown environment while simultaneously estimating their own pose within it.
We began by discussing the inherent circular dependency at the heart of SLAM: localization requires a map, and mapping requires localization. SLAM solves this interdependence by treating both as a joint optimization problem using sensor constraints from cameras, IMUs, LiDARs, and more.
We introduced the concept of pose graphs, where nodes represent poses and edges represent sensor-derived constraints (like odometry or loop closures). Optimization libraries such as G2O refine the entire graph to reduce drift and improve consistency.
A toy 1D SLAM example helped illustrate how landmarks and loop closures act as constraints. As drift accumulates over time, closing a loop enables the system to back-propagate corrections through the graph, significantly improving trajectory accuracy.
In the real world, SLAM powers everything from search and rescue operations and warehouse automation to consumer robots and AR/VR experiences, where robust and efficient localization is key.
We also reviewed various sensor configurations and SLAM systems: from monocular methods like ORB-SLAM2, RGB-D systems like ElasticFusion, LiDAR-based techniques like LOAM, to tightly-coupled visual-inertial odometry systems like VINS-Fusion and LIO-SAM.
Despite its promise, SLAM faces several challenges: sensor noise, drift accumulation, false loop closures, real-time constraints, and long-term consistency in dynamic environments.
Altogether, this session offered a foundational lens into how SLAM frameworks are built, what sensors and algorithms they depend on, and how they scale from academic demos to real-world autonomy.
🛠️ Hackable 3D Instance Mapping Setup:
-
We'll set up Concept Nodes, a lightweight re-implementation of ConceptGraphs by one of the original authors, Sacha Morin. Explore the repository by trying out different configurations and graph-building options.
-
Your goal is to understand how the framework attempts 3D instance mapping from RGB-D streams and investigate ways to improve robustness for indoor semantic mapping. Think about potential extensions, better descriptors, or improved spatial consistency strategies.
-
Understand various evaluation metrics (how are the AP, AP25, AP50 metrics calculated) used to assess mapping quality, such as semantic consistency and instance accuracy — refer to the ScanNet Benchmark and the ConceptGraphs evaluation discussion for details. Implement an evaluation script for ConceptNodes along these metrics.
| 📚 Topic | 🔗 Link |
|---|---|
| 📑 Lecture Slides – Visual SLAM |  |
| 🎥 Basics of SLAM – Cyril Stachniss |  |
| 📦 RTAB-Map |  |
| 🧠 Concept Graphs |  |
| ⚙️ g2o – Graph Optimization |  |
| ⚙️ Ceres Solver |  |
| 📖 The SLAM Handbook |  |
| 📖 14 Lectures on Visual SLAM by Gao Xiang (with C++ Impl.) |
 |
🚀 Recommended Reading: The SLAM Handbook
An open-access, very up-to-date textbook authored by leading experts in the SLAM community. Must read for anyone into perception for robotics!