This dissertation addresses autonomous navigation of robots in a dynamic environment where the existence of moving and/or unknown objects leads to more serious challenges for the robots than those when operating in a traditional stationary environment. Therefore, the use of learning capabilities to facilitate proper robotic operation in a dynamic environment has become an important research area in the past decade. This dissertation proposes several novel learning-based methods to overcome the shortcomings in the existing approaches of autonomous navigation. Three aspects are addressed in the present work. First, a real-time path planning method is designed for autonomous navigation that can generate a path that avoids stationary and moving obstacles. To this end, learning ability is imparted to the robot. The present framework incorporates the statistical planning approach called probabilistic roadmap (PRM), Q-learning together with regime-switching Markov decision process (RSMDP) due to its beneficial characteristics, to form a robust Q-learning. Consequently, the initial path can be improved through robust Q-learning during interaction with a dynamic environment. Second, motion planning under constraints is investigated. Specifically, a closed-form piecewise affine control law, called piecewise affine-extended linear quadratic regulator (PWA-ELQR), for nonlinear-nonquadratic control problems with constraints is proposed. Through linearization and quadratization in the vicinity of the nominal trajectories, nonlinear-nonquadratic control problems can be approximated to linear-quadratic problems where the closed-form results can be derived relatively easily. Third, people detection is integrated into the autonomous navigation task. A classifier trained by a multiple kernel learning-support vector machine (MKL-SVM) is proposed to detect people in sequential images of a video stream. The classifier uses multiple features to describe a person, and learn its parameter values rapidly with the assistance of multiple kernels. In addition to the methodology development, the present research involves computer simulation and physical experimentation. Computer simulation is used to study the feasibility and effectiveness of the developed methodologies of path planning, motion planning and people detection. The experimentation involves autonomous navigation of a homecare robot system. The performance of the developed system is rigorously evaluated through physical experimentation and is improved by refining the developed methodologies. ; Applied Science, Faculty of ; Mechanical Engineering, Department of ; Graduate
Robot learning for autonomous navigation in a dynamic environment
2015-01-01
Theses
Electronic Resource
English
DDC: | 629 |
Autonomous Quantum Reinforcement Learning for Robot Navigation
Springer Verlag | 2016
|Autonomous Robot Navigation Function
Springer Verlag | 2023
|Autonomous Navigation for Mobile Robot
Springer Verlag | 2023
|Virtual Autonomous Navigation Environment
NTIS | 2008
|