DRL approaches to planning and navigation for social robotics
Path planning is a fundamental problem of mobile robotics. In order to properly navigate within the environment, the robot needs to build a trajectory-free trajectory towards the goal starting from previous information, like its pose and the position of objects around it. In this field, Deep Reinforcement Learning has received great attention due to its strong ability to generalize the same task on different environment domains. With DRL, the robot learns from its own gathered experiences in a trial-and-error fashion, receiving positive or negative rewards based on its actions.
Although traditional planning and navigation methods like A* and DWA remains very robust and efficient solutions and, in some cases, DRL-based agents achieve to increase their performances even more, some problems still arise when humans are inserted into the equation.
Social navigation is the robotics field that addresses the core challenge of integrating social norms during the definition of the motion plan. Robot navigation in crowded, public spaces is still a complex task that requires not only information from the environment but also profound knowledge about social norms and common sense. The final goal is to obtain a socially acceptable behavior that could allow autonomous robots to be efficiently deployed in public pedestrian environments.
Department: PhD in Eletrical, Eletronics and Communication Engineering
Supervisor: Marcello Chiaberge