Artificial intelligence tools for path generation and optimisation for mobile robots
The ultimate goal in robotic systems is to develop machines that learn for themselves based on experience. In order to achieve on-line learning some software tools are needed to allow the robots to continually adapt their behaviour in order to constantly optimise their performance. This thesis presents research work focused on path planning for mobile robots with the objective of generating optimal paths for any type of mobile robot in an environment containing any number of static obstacles of any shape. The research specifically recognises that an optimal path can be defined according to several criteria including distance, time, energy consumption and risk. The easiest and most commonly used measure is to minimise distance, but this does not by itself optimise task performance, and the other criteria are generally far more important. Distance is used mainly because there is no direct method to optimise time, energy and risk as they depend on the characteristics of the robot and the environment. This is solved in this research by using a set of Artificial Intelligence tools working together to perform an optimisation process strictly on the criteria selected. The path planning system developed consists of an original and novel two-stage 4 process comprising generation followed by optimisation. Path generation is achieved using cellular automata whose behaviour has been determined by a genetic algorithm. A program called Rutar has been written in which the best behaviour found by the genetic algorithm is encoded, and it has been tested and shown to infallibly generate all the non-redundant paths between any two points around any obstacles. An interesting and valuable feature of Rutar is that the time taken to generate paths depends only on the amount of free space available in which the robot can move and therefore the more obstacles there are present, and hence the more complex the layout, the faster the execution time. The paths generated are sub-optimal solutions, which are then optimised according to the user's selection of a combination of Time, Energy, Distance and Risk criteria. The optimisation process is performed by another genetic algorithm. The original scheme used in this work allows any combination of all the desired criteria in a single optimisation process, allowing it to handle very complex non-linear problems. All of the optimisation criteria can be used in situations where the environment and the robot are considered to be unchanged during the interval in which the robot moves. This optimisation can be performed either off-line or on-line. However, the ability of the developed system to generate and optimise the paths very fast provide an opportunity for dynamic path optimisatiorý which ultimately can lead to on-line learning. This potential of the tools developed for the path planning system is explored and recommendations for further exploitation are made.