Investigating Human-Robot Teams for Learning-Based Semi-autonomous Control in Urban Search and Rescue Environments

Teams of semi-autonomous robots can provide valuable assistance in Urban Search and Rescue (USAR) by efficiently exploring cluttered environments and searching for potential victims. Their advantage over solely teleoperated robots is that they can address the task handling and situation awareness limitations of human operators by providing some level of autonomy to the multi-robot team. Our research focuses on developing learning-based semi-autonomous controllers for rescue robot teams. In this paper, we specifically investigate the influence of the operator-to-robot ratio on the performance of our proposed MAXQ hierarchical reinforcement learning based semi-autonomous controller for USAR missions. In particular, we propose a unique learning-based system architecture that allows operator control of larger numbers of rescue robots in a team as well as effective sharing of information between these robots. A rigorous comparative study of our learning-based semi-autonomous controller versus a fully teleoperation-based approach was conducted in a 3D simulation environment. The results, as expected, show that, for both semi-autonomous and teleoperation modes, the total scene exploration time increases as the number of robots utilized increases. However, when using the proposed learning-based semi-autonomous controller, the rate of exploration-time increase and operator-interaction effort are significantly lower, while task performance is significantly higher. Furthermore, an additional case study showed that our learning-based approach can provide more scene coverage during robot exploration when compared to a non-learning based method.