Mobile robots are becoming ubiquitous and an essential part of our everyday lives. They are increasingly taking their place in service-oriented applications including domestic and entertainment roles. They open up many potential opportunities, but they also come with challenges in terms of their limited sensing capability and accuracy and minimal on-board computing resources. In this dissertation, we address three fundamental problems in mobile robotics and demonstrate our approach to each of the problems with a mobile robot equipped with low-cost and low-end sensors. The problems we consider are those of mobile robot calibration, mobile robot localization, and simultaneous localization and mapping.
Motion and sensor models are crucial components in current algorithms for mobile robot localization and mapping. We demonstrate how the parameters of both the motion and sensor models can be automatically estimated during normal robot operations via machine learning methods thereby eliminating the necessity of manually tuning these models through a laborious calibration process. Mobile robot calibration is important especially for robots relying on cheap and less-accurate sensors. Results from real-world robotic experiments with a robot equipped with wheel encoders and sonar sensors are presented that show the effectiveness of the estimation approach.
Monocular vision has long been regarded as an attractive sensor for the localization of a mobile robot. In this dissertation, we present a particle filtering approach to real-time pose estimation for a small-scale indoor mobile robot equipped with wheel encoders for its odometry and aided by a standard perspective camera. Vision is used for detecting naturally occurring static three-dimensional point features or landmarks from the environment and utilizing the information for correcting the pose as suggested by the odometry. We validate the effectiveness of the particle filter approach extensively with both simulations as well as real-world data and compare its performance against that of the extended Kalman filter.
Simultaneous localization and mapping (SLAM) is a well-studied problem in mobile robotics and the majority of the existing techniques rely on the use of accurate and dense measurements provided by laser rangefinders to correctly localize the robot and produce accurate and detailed maps of complex environments. In this dissertation, we present our approach to SLAM with low-cost but noisy and sparse sonar sensors in large indoor environments involving large loops. Results from robotic experiments demonstrate that it is possible to produce good maps of large indoor environments with large loops despite the inherent limitations of sonar sensors.
