Mattia Manfroni
In this thesis we propose an automatic methodology to synthesise robotic agent programs. This kind of technique is characterised by two main components: the agent program model and the optimisation algorithm. In our methodology, the agent program model is given by a Boolean Network (BN) and the optimisation algorithm consists of a metaheuristic technique. We model the BN design as a constrained combinatorial optimisation problem by properly defining the set of decision variables, constraints and the objective function. More precisely, the Boolean functions of the nodes represent the decision variables and the optimisation algorithm searches for their optimal values according to the objective function.
The methodology is first validated by experiments on abstract case studies, such as reachability problems in BNs. After the validation, we apply the method to robotics case studies. In a first test case, we design a BN-robot able to attain a path following task. The second, more difficult, test case concerns phototaxis/antiphototaxis behaviour, in which the robot is required to keep a sort of internal memory to achieve a given goal. The robot must seek the light and go towards it (phototaxis); subsequently, when it perceives a sharp sound, it must change its behaviour and move away from the light (antiphototaxis). In addition, the robot must be robust, i.e., able to correct its trajectory in case of noise and external perturbations. At the end of the automatic design process, the obtained agent programs are ported into a real robotic platform. The obtained results show that BN dynamics is suitable to realise complex behaviours notwithstanding the simplicity of the model.
