Is Semantic SLAM Ready for Embedded Systems ? A Comparative Survey

TL;DR

Is Semantic SLAM ready for embedded systems? This paper surveys three AI-enhanced SLAM classes—Geometric-based, NeRF-based, and Gaussian Splatting—evaluating their viability on resource-constrained hardware, specifically the NVIDIA Jetson AGX Orin. It conducts a comparative analysis of accuracy, segmentation quality, memory usage, and energy consumption, revealing that semantic Geometric SLAM currently offers the best embedded practicality, while NeRF and Gaussian Splatting deliver richer semantic detail at substantial cost. The study advocates hardware-aware co-design and lightweight representations as essential directions to enable real-time, energy-efficient semantic SLAM in dynamic real-world environments.

Abstract

In embedded systems, robots must perceive and interpret their environment efficiently to operate reliably in real-world conditions. Visual Semantic SLAM (Simultaneous Localization and Mapping) enhances standard SLAM by incorporating semantic information into the map, enabling more informed decision-making. However, implementing such systems on resource-limited hardware involves trade-offs between accuracy, computing efficiency, and power usage. This paper provides a comparative review of recent Semantic Visual SLAM methods with a focus on their applicability to embedded platforms. We analyze three main types of architectures - Geometric SLAM, Neural Radiance Fields (NeRF), and 3D Gaussian Splatting - and evaluate their performance on constrained hardware, specifically the NVIDIA Jetson AGX Orin. We compare their accuracy, segmentation quality, memory usage, and energy consumption. Our results show that methods based on NeRF and Gaussian Splatting achieve high semantic detail but demand substantial computing resources, limiting their use on embedded devices. In contrast, Semantic Geometric SLAM offers a more practical balance between computational cost and accuracy. The review highlights a need for SLAM algorithms that are better adapted to embedded environments, and it discusses key directions for improving their efficiency through algorithm-hardware co-design.

Figures (9)

• Figure 1: Timeline of key advancements in Semantic SLAM research from 2018 to 2024. The progression is categorized into three primary approaches: Semantic Geometric SLAM (Turquoise), Semantic NeRF SLAM (Coral), and Semantic GS SLAM (Yellow), with the size of each circle representing the relative volume of publications in that year. Survey papers on SLAM systems are represented in Purple. Among them, there is currently no paper specifically dedicated to Semantic SLAM, regardless of whether it is in an embedded context or not.
• Figure 2: Typical pipeline of a geometric SLAM system, showing the fundamental components: tracking for pose estimation, mapping for environment reconstruction, and optimization through bundle adjustment.
• Figure 3: Citation network and relationships in Semantic SLAM research. The visualization represents five distinct categories: Geometric SLAM approaches (Turquoise), NeRF-based SLAM methods (Coral), GS-based SLAM techniques (Yellow), Framework implementations (Green), and Segmentation/Detection Models (Purple). Node sizes correspond to citation impact, with larger nodes indicating higher citation counts. The network topology reveals two distinct clusters: (1) A left-oriented cluster dominated by NeRF and Gaussian Splatting SLAM approaches, sharing common theoretical foundations and citation patterns and (2) A right-oriented cluster primarily composed of geometric approaches, centered around foundational works like ORB-SLAM.
• Figure 4: Architecture of RDS-SLAM 9318990 based on ORB-SLAM3. The traditional geometric SLAM components are shown in blue, while the purple elements highlight the additional semantic threads. The Semantic Thread processes dynamic points detection, and the Semantic-based Optimization Thread enhances both mapping accuracy and localization through semantic information.
• Figure 5: Overview of VDO-SLAM architecture zhang2020vdoslam. The system extends traditional geometric SLAM (in blue) with additional components (in purple) to handle dynamic environments. The pipeline introduces a preprocessing stage with depth estimation, instance segmentation, and optical flow computation. A parallel dynamic object tracking stream complements the standard geometric tracking, enabling the system to simultaneously handle both static and dynamic elements in the mapping process.