- 57 views
- 48 downloads
Towards a Robustness/Resilience-Aware Simultaneous Localization and Mapping (SLAM)
-
- Author / Creator
- Ali, Islam AM
-
In this thesis, we investigate the problem of robustness and resilience in simultaneous localization and mapping systems (SLAM). With the vast adoption of robotics in many industries and disciplines, robustness and resilience are becoming of immense importance for the reliable and safe deployment of robotics in real-world settings. The current literature in SLAM treats accuracy as a synonym for both robustness and resilience. In this thesis, we start by providing a rigorous and formal definition for robustness and resilience as two major objectives in modern robotics, which reveals the following requirements for achieving them: (1) quantitative characterization of operating conditions, (2) objective evaluation of SLAM, (3) predictability of performance, and (4) real-time fault detection and performance monitoring. Thus, the thesis is divided into four parts addressing each of the aforementioned requirements.
Robustness and resilience are tightly coupled to measurable operating conditions in which a robot is deployed. Due to the difficulty of SLAM evaluation in the real-world, researchers utilize datasets for that purpose. These datasets implicitly include the operating condition at which the reported performance is guaranteed. Thus, our first proposal addresses this problem by introducing a generic and extensible framework for the quantitative characterization of SLAM datasets and benchmarks. The proposed system automatically characterizes datasets based on defined metrics that collectively represent the operating conditions imposed by an environment or robot motion. Additionally, the proposed system automatically and systematically provides analysis of datasets on the measurement, sequence, and dataset level, which can be used to push the boundaries of SLAM evaluation. Moreover, it analyzes the correlation between metrics and SLAM performance, revealing the sensitivity of a given algorithm to many environmental conditions.
Current evaluation methodologies in SLAM are very redundant in nature and do not take into consideration the characteristics of the deployment environment. To that end, our second proposal tackles this point by proposing a tabulation-based dynamic programming algorithm that utilizes the quantitative characterization of SLAM datasets to achieve two goals, which are: elimination of redundancy in the evaluation pool of data sequences, and objective selection of the most proper subset of sequences to ensure coverage of the deployment conditions.
Moreover, we address the problem of performance predictability in SLAM in our third proposal by employing a supervised ensemble learning regression model to predict SLAM errors on the pose level and the trajectory level. We utilize the concept of sub-trajectories for diversifying the number of training samples and apply a 1-D global average pooling function for dimensionality reduction. Additionally, we ensure the efficacy of the method with limited training data and analyzes out-of-distribution predictions.
Fault detection is critical for resilience in SLAM since it highlights the need to engage recovery mechanisms to converge when divergence is detected. Therefore, our fourth proposal uses a decoupled calibrated IMU-based kinematics model as a high-accuracy supervisory monitoring signal for SLAM. The method relies on a modified version of DUET, a deep learning IMU calibration method, to provide reliable IMU measurements. Then, consistency is measured between IMU-based and SLAM pose streams and is then used as an indicator of faults. Moreover, the proposed method is non-invasive and algorithm agnostic since it has no assumptions on the SLAM system itself.
-
- Subjects / Keywords
-
- Graduation date
- Fall 2024
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.