GAN-SLAM: Real-Time GAN Aided Floor Plan Creation Through SLAM

TL;DR

GAN-SLAM addresses the challenge of noisy 2D occupancy grids produced by SLAM by integrating a GAN-based cleaning module into the SLAM pipeline. It adapts 3D LiDAR odometry to 2D OGMs and uses an image-to-image translation approach with a PatchGAN discriminator and patchwise contrastive loss to produce pixel-perfect maps in real time. The approach is validated on real-world Haslegrave-building data and on simulated Radish/PseudoSLAM data, showing improved map fidelity and the ability to complete partial observations for floor plan creation, albeit with increased computational cost and limitations in outdoor or highly varied architectural contexts. This work suggests a practical path to large-scale floor plan creation with a single LiDAR, while highlighting the need to manage potential hallucinations and domain-specific generalization challenges.

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 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.

Figures (7)

• Figure 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.
• Figure 2: GAN-SLAM System Architecture Diagram for real-time occupancy grid mapping. Main stages: 1) Take the input of consecutive LiDAR pointclouds to calculate pose through generalised-ICP and b17. 2) Combine calculated pose with 2D conversion of incoming scan to build an estimated $OGM_e$. 3) Preprocess $OGM_e$ through filtering operation and pixelisation. 4) Parse preprocessed $OGM_e$ to GAN model to produce clean variant. 5) Filter 'clean' OGM back into OGM format from $OGM_e$ and publish it as a map to the robot.
• Figure 3: Training procedure for 2D occupancy grid cleaning through an I2I GAN, $I_x$ and $G(I_x)$ are generated by our modified PseudoSLAM simulator and are in image format, we use a Query-selected attention module b24 selection of $K^-$, $K^+$ and anchor points $q$ to compute a Multi-layer Patchwise Contrastive Loss from b23.
• Figure 4: Samples from the dataset used to train the GAN model. Top row: samples of erroneous occupancy grid maps, bottom row: samples of 'clean' occupancy grid maps.
• Figure 5: A real-time prediction made by GAN-SLAM of the Haslegrave building at REDACTED University. (a) Estimated OGM, used as input to GAN model, (b) Clean OGM variant prediction produced by GAN-SLAM.