Modelling and simulation of mobile robot motion and its interaction with humans

In recent years, technological advances have allowed robots to become a part of our everyday lives. The integration of such robots in human-populated environments presents new challenges, particularly in the context of navigation. To operate efficiently in these scenarios, robots must move safely and in a socially acceptable manner, considering the physical and psychological safety of pedestrians. To achieve these goals, it is essential to develop socially aware navigation algorithms able to perceive humans as social entities, predict their movements, and perform mutual avoidance maneuvers. Although existing approaches ensure safe robot navigation even in crowded environments, many of them exhibit reactive behavior and treat pedestrians as mere dynamic objects without predicting their future movements. Recently, various machine learning techniques have been used to predict human motion with optimal results. However, most of them focus on predicting human movements individually, neglecting potential interactions between pedestrians during navigation. Therefore, modeling social behaviors is crucial for designing socially-aware robots that can predict future pedestrians' motions and adjust their decisions accordingly. To face these challenges, this thesis proposes a navigation algorithm that combines game theory with the well-known Social Force Model (SFM). Unlike previous approaches, game theory allows to explicitly model the decision-making process typical of human beings, considering pedestrians and the robot as rational agents capable of influencing each other's decisions. Here, navigation is modeled as a non-cooperative game. Each agent has a set of possible actions represented by trajectories generated from different sets of Social Force Model parameters. Each agent aims to find optimal trajectories considering potential interactions with other players. The solution of the game is established by reaching Nash equilibrium. To ensure a higher level of naturalness and comfort, it is necessary for game theory to solve the game by considering trajectories that are as real as possible. However, the manual definition of the SFM parameters can be very complex due to the high sensitivity of the model and the significant variability of the behaviors they determine. Therefore, a differential evolution algorithm has been applied to a real trajectory dataset ("Thör") to extract the best SFM parameters approximating each trajectory. However, the excessive time required for estimation makes DE difficult to use in real time and unable to ensure variability among the parameter sets needed for game theory. To overcome this problem, two neural networks have been employed: one for the robot and the other for pedestrians. In this context, Differential Evolution has been used to generate the corresponding training datasets, in which the estimated best parameters have been used to label specific features of the analyzed real trajectory.