The principles and development of mobile robots are active research topics with vast practical demand and future potential. Recent years have witnessed an increasing interest from industry in such applications as service robots (e.g. mail delivery in office buildings, food and medicine delivery in hospitals) and intervention robots (e.g. remote rescue, verification, or surveillance in hazardous environments).
Navigation is normally divided in different functional
blocks. At the highest level is the path planning block, determining the route(s) to be followed in order to achieve a specific objective. At a lower level are the emergency stop, obstacle detection and obstacle avoidance blocks. Localisation determines the robot's position and orientation in the environment coordinate frame. It provides periodic reference data to be used by the other navigation blocks. There is one block, for which there does not exist fully convincing solutions: the path follower. Its role is to map the high level goals, as determined by the path planner, to the actual situations the robot is facing in its motion. As an example, it would be up to the path follower to decide where to go whenever an obstacle is detected. This apparently simple function is not that easy to implement, the main reason being the fact that it entails some sort of mixed reasoning both at high level (the target coordinate framework) and low level (reasoning on what to do considering the sensorial information available).
Application of machine learning (ML) techniques to robotics is a natura idea that has been studied to a relative extent. ML allows easy incorporation of adaptivity, robustness, and other essential properties that are required in robots that will have to manage by themselves in unknown environments. Decision tree (DT) learning is among the most established ML techniques and it has had notable success in many practical applications both within and without robotics. There are many advantages to using DTs rather than alternative techniques.
The underlying purpose of this project is to explore the application potential of symbolic ML techniques in autonomous mobile robot navigation. The practical aim is to develop a path follower utilizing DT learning and, thus, create a full-scale navigation facility for an autonomous vehicle roaming around in a dynamic environment. The project continues work that has been going on in the Host Institution for five years now. The ground work has been laid in Santos' dissertation, where the basic obstacle avoidance component was developed, and in the work of Millan, which shows how goal-oriented strategies can be learned from experience. The obstacle avoidance module makes use of neural networks and a local (to the vehicle) perception map that is built on the basis of the data provided by the ultrasonic sensors of the robot. Millan's work utilizes reinforcement learning technique. These two projects have solved and implemented modules at the far-ends of a working navigation system. However, the obstacle avoidance system can only prevent the robot from bumbing into things by navigating the free space and contouring an obstacle that it happens to come across. No method for following a path given by, e.g. Millan's, path planner is provided. Hence, the two modules, as yet, cannot cooperate. The practical objective of this project is to bridge this gap and devel a DT-based system that can implement teleological (purposeful) local navigation instead of simply avoiding obstacles that happen to fall on the robot's path. More ground-laying objectives are to explore the utility of ML in a concrete, well-defined application and to investigate the structure of the navigation system of a mobile vehicle.