Challenges and Performance of SLAM Algorithms on Resource-constrained Devices

Evaluating the performance of Simultaneous Localization and Mapping (SLAM) algorithms is essential for the progress of robotic systems. However, conducting a comprehensive assessment of SLAM systems in the context of recent advancements is challenging due to the wide variety of hardware platforms, algorithm configurations, and datasets available. This study aims to test SLAM algorithms on resource-constrained devices such as the NVIDIA Jetson AGX Orin 64GB. Experiments are conducted with various visualbased localization algorithms that either leverage deep learning models for specific tasks within the SLAM process or are learned end-to-end to estimate camera pose. The evaluation focuses on the following systems: RDS-SLAM and VDO-SLAM, which utilize semantic information to achieve precise motion estimation; TSformer-VO, an end-to-end Transformer-based model designed for monocular visual odometry; and DeepVO, which based on recurrent neural networks. The systems are evaluated using several metrics, including ATE and RPE to assess pose accuracy and rotational drift, respectively, alongside runtime, energy consumption, and resource usage to gauge their efficiency and practicality for real-world applications.