Modelling and Validation of Socially-Aware Navigation Algorithms for Mobile Robots in Populated Environments

In our current era, robots have transcended industrial confines and are being deployed across multiple sectors, necessitating a heightened degree of autonomy. Anticipating the imminent future, it is plausible that human-robot interaction will be part of our daily activities, and as a consequence, robotic entities will inevitably manage their navigation within environments shared with humans. In such contexts, the agents' actions must be similar to behaviors that humans would organically manifest under analogous circumstances, and, in order to be socially acceptable, the motion needs to be safe but must also respect social rules and conventions. Although existing algorithms for socially-aware robot navigation ensure safety, many of them generate unnatural trajectories and have limitations in predicting pedestrians' future movements. In this context, this final thesis project aims to perform an in-depth study of the state of the art in navigating autonomous robots in human-populated environments, identifying the best algorithms currently used and proposing an innovative approach based on game theory. Unlike the majority of the solutions in the literature, game theory models the decision-making process of humans, treating them and the robot as rational agents that interact with each other. Each agent aims to find an optimal sequence of actions to generate an optimal trajectory to follow. The solution of the game is considered to be the attainment of the well-known Nash equilibrium. Simulations have shown good performance, but the time needed for computations makes the overall algorithm impractical for real-time implementation on a real robot. To overcome this issue, the algorithm is used to create a training dataset for a fully-connected neural network, which is used in practice to replace the computation of the action decided by the game theory planner (GTP). The proposed approach has been preliminarily validated through a Monte Carlo numerical simulation performed in the software Gazebo, reproducing a virtual representation of operating conditions. Specifically, the GTP is compared with two other state-of-the-art algorithms: the Social Force Model (SFM) and Optimal Reciprocal Collision Avoidance (ORCA). The comparison has been performed using state-of-the-art metrics and indicates that GTP outperforms the other approaches. In order to also reach a qualitative and social evaluation of the algorithm and conduct an in-depth study about the impact that a real robot would have on the emotional state of pedestrians when facing it, an experimental test has been designed and performed. Each algorithm is inserted as a local planner into the navigation stack of a real robot (Locobot wx250s of Trossen Robotics), and several point-goal trials are performed in an environment shared with volunteer participants. The social acceptability of the trajectories generated by these algorithms has been assessed through a state-of-the-art questionnaire submitted to the participants, where the measured metrics are the human-likeness of the trajectories and the discomfort experienced during the tests.