GAN-SLAM: Real-Time GAN Aided Floor Plan Creation Through SLAM
Leon Davies, Baihua Li, Simon Sølvsten, Qinggang Meng · 2025
Read GAN-SLAM: Real-Time GAN Aided Floor Plan Creation Through SLAM by Leon Davies, Baihua Li, Simon Sølvsten, Qinggang Meng (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-19653.txt.
Introduction
Leon Davies, Baihua Li, Simon Sølvsten, and Qinggang Meng introduced GAN-SLAM in 2025, a mapping method for robots that cleans maps as they are built. Robots use such maps to understand where they are, but motion can add noise and gaps, which spoils tasks like drawing floor plans.
Their system uses a pair of rival neural networks to clean and complete the grid map during the run. They also adapted pose tracking methods from three-dimensional mapping for flat maps. On real-world data, run in real time, the maps came out much cleaner with few errors. Better maps open up direct uses such as drafting floor plans from a robot's survey.
Catalog record
| CREATOR | Leon Davies, Baihua Li, Simon Sølvsten, Qinggang Meng |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-19653
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GAN-SLAM: Real-Time GAN Aided Floor Plan Creation Through SLAM Leon Davies, Baihua Li, Simon Sølvsten, Qinggang Meng Abstract SLAM is a fundamental component of modern autonomous systems, providing robots and their operators with a deeper understanding of their environment. SLAM systems often encounter challenges due to the dynamic nature of robotic motion, leading to inaccuracies in mapping quality, particularly in 2D representations such as Occupancy Grid Maps. These errors can significantly degrade map quality, hindering the effectiveness of specific downstream tasks such as floor plan creation. To address this challenge, we introduce our novel ‘GAN-SLAM‘ a new SLAM approach that leverages Generative Adversarial Networks to clean and complete occupancy grids during the SLAM process, reducing the impact of noise and inaccuracies introduced on the output map. We adapt and integrate accurate pose estimation techniques typically used for 3D SLAM into a 2D form. This enables the quality improvement 3D LiDAR-odometry has seen in recent years to be effective for 2D representations. Our results demonstrate substantial improvements in map fidelity and quality, with minimal noise and errors, affirming the effectiveness of GAN-SLAM in for real-world mapping applications within large scale complex environments. We validate our approach on real-world data operating in real-time, and on famous examples of 2D maps. The improved quality of the output map enables new downstream tasks, such as floor plan drafting, further enhancing the capabilities of autonomous systems. Our novel approach to SLAM offers a significant step forward in the field, improving the usability for SLAM in mapping based tasks, and offers insight into the usage of GANs for OGM error correction. † † impactstatement: Occcupancy grid maps in their current form can be interpreted as unappealing to human users as their intended use is primarily for navigation tasks by robots. We propose a GAN model embedded into the SLAM process trained to convert the noisy maps produced by SLAM into clean variants by identifying errors and noise in the SLAM process. Reliable reduction of errors and artefacts in maps produced by SLAM has the potential to improve the usability of occupancy grid mapping for new domains such as facilities management, or the insurance industry that could benefit from quickly produced accurate maps. Our GAN-SLAM also offers insight into reliable error reduction to common SLAM errors such as sensor noise, linear and angular offsets through GANs. I Introduction Occupancy grid mapping (OGM) is a 2D representation of SLAM in which the world is represented as a grid map. Each grid cell represents a state of occupation for its position in the map as occupied, free or unexplored. By combining data collected from various sensors such as LiDARs or Inertial Measurement Unit (IMU) sensors, OGM enables robots to navigate and interact autonomously and dynamically with their surroundings. This enables OGMs to be a reliable format for further tasks such as exploration, obstacle avoidance and path planning. OGMs built through SLAM are prone to noise and artefacts due to the fundamental underlying processes of estimation which enable SLAM. This can be attributed to the sensors employed, such as high dimensional LiDAR scanners, environmental factors, such as the occurrence of transparent or reflective materials, and the dynamic motion of the robot itself. Additionally, errors in pose estimation and drifts in odometry contribute to inaccuracies in the generated maps such as linear or angular offsets. These errors may be more pronounced when utilising lower cost sensors [ 1 ] . However, higher power sensors have the potential to introduce different types of erroneous data such as LiDAR capturing data through open doorways or glass. Furthermore, Partially made observations or blocked segments of rooms due to occlusions leave unfinished sections of maps. In the past these factors have made OGM not as reliable for tasks such as automated floor plan creation. Fig. 1: GAN-SLAM: Data pipeline, producing a high quality defect free OGM from the singular input of 3D LiDAR. (A) 360-degree LiDAR scanner. (B) 3D PointCloud captured by LiDAR scanner. (C) 2D LiDAR data + Odometry calculated from LiDAR data. (D) 2D/3D SLAM produced by LiDAR & pose estimate. (E) Estimated 2D OGM. (F) GAN-Cleaned OGM. Traditional methods of SLAM optimisation and error correction include the usage of bundle adjustment [ 2 ] , loop closure [ 3 ] and probabilistic methods [ 4 , 5 ] . These are fundamental processes in modern day SLAM systems. These methods are not only susceptible to irregularities and artefacts but in many cases are the root cause for these errors. Many irregularities have become accepted as normal within modern day 2D OGM. This degradation in map quality can make OGM unappealing for usage as floor plan diagrams. Errors such as linear/angular offsets, accidentally captured observations, and partial observation completion can be difficult to resolve with traditional programming practices but are easy for humans to identify. This inspires our proposed GAN-SLAM method which uses a deep learning approach to resolve these issues. We present a novel concept within SLAM which utilises deep neural networks trained to identify various SLAM errors, predict a completed room from a partial observation, and apply these changes and resolve them in real-time. We implement state of the art LiDAR-odometry traditionally used in 3D SLAM to 2D OGM. This massively improves localisation accuracy within large scale complex environments, with no further sensors required. This also allows us to benefit from the increase in range that more powerful 3D scanners offer. Our results demonstrate robust and reliable mapping quality and offer a novel concept of floor plan creation through error correction and observation completion in OGM through GANs. Through our SLAM, complex environments are able to be successfully mapped, our method out performs all other pure LiDAR OGM SLAM tested in terms of mapping quality and accuracy. Mature 2D laser scan matching methods with 2D LiDARs are commonly used for OGM tasks. Our specific context involves mapping several thousand large factories across the UK with autonomous systems. In this scenario the increased range 3D LiDARs offer and more precise pose estimate is necessary. With our method we are able to produce reliable results from partially complete scans that are completed by the GAN. This is a never before seen feature within a SLAM and can amount to much better efficiency of mapping. SLAM algorithms often show success in small scale controlled environments but often fail to produce reasonable results in larger more unpredictable environments [ 6 ] . With our method we prove and contribute a new SLAM that can be relied upon for large scale mapping tasks. II Related Work II-A SLAM Optimisation Traditional methods of particle filter based SLAM enable real-time simultaneous pose estimation and map construction. Methods such as Monte Carlo Localisation [ 4 ] and Rao-Blackwellized filtering [ 5 ] utilise a probabilistic interpretation where pose is represented by a set of weighted particles, enabling robust SLAM in small-scale dynamic and unknown environments. Extended Karlman Filtering [ 7 ] incorporates elements of Karlman filters for state estimation typically in conjunction with feature-based SLAM. These SLAM optimisation techniques may introduce errors in complex or uncertain environments and cannot discern between intended mapping and accidental mapping such as observations captured through open doorways or transparent materials. Feature matching techniques