Robot social navigation: a quantitative and qualitative benchmark of state-of-the-art algorithms in real-world

Social navigation has seen a considerable number of developments in the last few years, considering that the possible tasks that a robot can perform while being aware of the presence of other people are numerous. There are many aspects of socially aware navigation that contribute to the achievement of the current objective, but this thesis focuses on the local planner, also known as the local controller. The controller handles the information around the robot to send the velocity command that controls the rotational and translational speed of the robot. The main objective of this thesis is to compare the performance of two different algorithms, namely the DWA and the MPPI, in a set of laboratory experiments. Six total variations correspond to the base version of the algorithms and two enhanced versions with, respectively, a social costmap plugin and the Social Force Model. Considering that the Social Force Model was implemented solely on the DWA algorithm, the first part of this thesis focused on developing a method to integrate said model into the MPPI. This led to the development of a plugin that operates alongside the other critics and computes the social work for the randomly sampled trajectories. The second half of this work consisted of performing the tests themselves. The main aspect that differentiates the work developed in this thesis from the other papers present in the literature is the acquisition of both quantitative and qualitative metrics. The quantitative metrics are variables like the time of completion, the social work, and the minimum distance to the agents that were obtained through the acquisition of positions and velocities in the laboratory experiments. The qualitative metrics, on the contrary, include unobtrusiveness, friendliness, smoothness, and avoidance foresight and they were decided by the agents themselves after the end of each experiment, allowing a complete evaluation of each algorithm. The experiments were conducted in the PIC4SeR laboratory using the Jackal robot from ClearPath and a set of VICON cameras which were essential in tracking the position and the velocities of both the robot and the human agents, which were equipped with a custom-made accessory each to make them recognizable for the cameras. Eight different scenarios were developed to test the navigation prowess of the robot in different conditions. Said scenarios were characterized by a different disposition of the robot, the human agents, and the static obstacles. Each algorithm was tested five times, totaling 240 tests.