Learning from humans to improve Socially-Aware Motion Planning

Humans are more successful in planning collision-free trajectory with mutual avoidance manoeuvres in a populated environment than any motion planning algorithm so far. While humans easily can deal with predict the motion of surrounding people, robotics systems are still encountering problems. For this reason, we want to improve the robot navigation starting from the study of humans decisions. The majority of the approaches concentrate their attentions to predict the motion of agents individually without considering the interaction between humans. Thus, we model the motions of humans considering the interactivity of pedestrians in the presence of multiple agents using game-theoretic approach. The game theory has a lot of advantages over the reactive method, in fact can model the mutual anticipation of the influence of other agents and adapt the own decisions based on the possible actions of others. Non-cooperative game theory is applied to predict the decision of multiple humans that interact each other during navigation. Hence, the concept of Nash equilibrium in dynamic games is applied to solve the game human motion model. In the last 5 years, some scientists studied the humans motion from a game-theoretic point of view but the models were for 2 people and without taking into account the static obstacle and group of people. We have reformulated the problem considering a different cost function and extending the model to multiple people adding also the group recognition and the interaction between human and static object. The model has been validated with real-world surveillance videos. In the second part of the dissertation, we used that model to create a path planning for autonomous robots that navigate among multiple humans. In the end, we tried to validate the hypothesis that the final motion robot planning is socially acceptable for humans.