LRN2002 - Learning Methods for Robot Operation

This project addresses learning methods for mobile robots. Some example-applications of the mobile robots are: robots for service buildings such as hospitals and offices, human-oriented robots, factory mobile robots, automated car driving, autonomous transportation systems for cities, and planetary and underwater exploration. The available sensor technologies give the possibility to equip robots to have a better understanding of the surrounding world. However, robots must perceive useful information from raw sensor data, and learn models of the world for better deciding its actions according to its objectives. Learning techniques provide autonomy and flexibility on the creation of robot control competencies, and decreasing the need for a priori models of the robot, world, and robot world interactions. Additionally, learning only relevant aspects of the world can decrease computational overheads and complexity. Intelligent algorithms and paradigms, based on Fuzzy Logic, Neural Networks, supervised learning, self-organizing learning, reinforcement learning, memory-based learning, or other techniques, are being investigated and developed enabling robots to learn, navigate, plan and behave in the real world. Kalman estimators and energy minimization methods can be powerful tools for adjusting learned models. The main goal of this project is to expand the research from a navigation architecture previously developed by the team. For this purpose, at the world model level, it will be developed a multi-resolution memory-based method and a self-organizing neural network that will be used for implementing a map learning method for dynamic worlds. An important aspect for a mobile robot is the knowledge of its own location in the environment. This is crucial for learning the world model and for planning robot motions. An objective of this project is to perform research towards the integration into our navigation architecture of a method for mobile robot localization. Another goal of this project is to investigate how to integrate in our navigation architecture methods for learning robot behaviors with tightly-coupled sensing and action information. Applying reinforcement learning and supervised learning modules for learning in continuous spaces is an interesting and challenging research subject. With the work of this project we aim to investigate methods and algorithms towards the development of intelligent robots which presently are still very far from our needs, in spite of the big developments of science and engineering, Intelligent algorithms and paradigms, based on Fuzzy Logic, Neural Networks, supervised learning, self-organizing learning, reinforcement learning, memory-based learning, or other techniques, are being investigated and developed enabling robots to learn, navigate, plan and behave in the real world. Kalman estimators and energy minimization methods can be powerful tools for adjusting learned models. The main goal of this project is to expand the research from a navigation architecture previously developed by the team. For this purpose, at the world model level, it will be developed a multi-resolution memory-based method and a self-organizing neural network that will be used for implementing a map learning method for dynamic worlds. An important aspect for a mobile robot is the knowledge of its own location in the environment. This is crucial for learning the world model and for planning robot motions. An objective of this project is to perform research towards the integration into our navigation architecture of a method for mobile robot localization. Another goal of this project is to investigate how to integrate in our navigation architecture methods for learning robot behaviors with tightly-coupled sensing and action information. Applying reinforcement learning and supervised learning modules for learning in continuous spaces is an interesting and challenging research subject. With the work of this project we aim to investigate methods and algorithms towards the development of intelligent robots which presently are still very far from our needs, in spite of the big developments of science and engineering.