Achieving multi-robot cooperation in highly dynamic environments

Abstract

The research in this thesis addresses the problem of providing successful cooperation between multiple mobile robots in highly dynamic environments. Cooperation involves multiple robots working towards a common goal. It is difficult to implement meaningful robot operations in these environments due to the constantly changing environment features making action-selection and action-location highly transient. This problem becomes worse when the robots require coordination between them to carry out their tasks. Typical multi-robot control methods reduce the complexity of these problems by abstracting or subsampling the information detected from the environment and robot abilities. This reduction indicates the designer has encoded knowledge about the problem in the multi-robot architecture. To achieve this reliably, all environment and robot state transitions must be allowed for. Also, methods of handling any contingencies must be implemented. This approach has the effect of limiting the level of problem complexity that the robots can manage. The consequences of not restricting the environment and robot variables are the problem complexity increases, but the robot controller has more freedom to develop solutions. A goal of this thesis is to develop multi-robot control approaches that allow for this freedom and provide high levels of performance. Robot control is a function of the tasks the robots have to carry out. In the environments selected for this research, the tasks are based on small action sets that consist of directives. The directives consist of an action and a location for the action. The directives do not have to be obeyed since that would be to the detriment of carrying out tasks in an environment where the currently selected action for a robot needs to change with the environment dynamics. By using the directives this way, fast planning can result with the robots being directed but not bound to the commands. Two mechanisms are implemented that use this approach of multi-robot control based on using directives as outputs and less restrictive inputs. They are centralised systems and classified based on their intelligence source for developing the solutions. Centralised control allows for better coordination of robot actions since decisions are based on the same world model information. The first mechanism is a designer-based approach. It uses the extracted features of the environment and reconstructs them in virtual space where they are abstractly represented to align with the system goals. These features are modelled by potential fields. The fields are constructed under the control of high-level algorithms and the system is set up entirely under designer control. This produces a dynamic rule-based approach where the rules result from the combinations of potential fields.The second mechanism is designed from a machine intelligent perspective that takes advantage of the temporal aspects of the tasks the robots have to carry out in the selected environments. It is a temporal difference approach that uses an artificial neural network for the learning component. This allows for generalisation of solutions and less restriction on the input variables derived from the environment features. A problem with using temporal difference learning on complex problems is the time taken to train the system. A simulator strongly correlated to the real environment is used for this task since it allows for time to be compressed.The environments selected to develop the high-level controllers are the highly dynamic predator/prey and competitive robot soccer environments. Both benefit from planning and coordination even though this is difficult to achieve with the dynamics of the environment variables. Plans and coordination must be robust to the environment dynamics and opportunities to achieve either must be taken quickly.Once the approaches have been developed in their selected environments, they demonstrate high levels of performance for m