Robust Algorithms for Mobile Robots to Learn Human Preferred Movement in a Hallway

Abstract:

We propose a dual expert algorithm (DEA) to assist a mobile robot in learning a person’s preference of moving direction when the human encounters the robot in a typical hallway. When in use in hospital and office environments, a mobile robot will routinely encounter people who are not its primary user. These people need to feel comfortable around the robot for its implementation to be a success. Precisely predicting the person’s choice in this setting is challenging because humans may not be consistent in choosing their direction to move to avoid the mobile robot. Hence it is impossible for any algorithm to be correct every time. An added difficulty is that a robot that rapidly changes it behavior to adjust to new information can be perceived as unpredictable, which would also make the person uncomfortable around the robot. The DEA is able to make consistent predictions on a person’s choice after sufficiently large number of encounters with the same person. It reaches a steady state error that is reflective of the number of ‘mis-steps’, or choices against the preferred direction that the human they are encountering makes. And we show that the DEA becomes more robust than the weighted majority algorithm as the number of encounters increase. If the person occasionally deviates from their preference, the DEA tends to ignore such mis-steps. The robust performance of the DEA is rigorously justified via theoretical analysis. The DEA is limited by the assumption that the human starts within the center of the hallway. While this assumption is supported by research into average pedestrian movement, it is not unreasonable that the robot could encounter a human walking much closer to one of the walls during a hallway encounter. Thus we developed an expanded version of the DEA, the EDEA (expanded dual expert algorithm). This algorithm takes into account the starting position of the human in the hallway and is able to find the position in the hallway where the human’s preference switches from one wall to the other. Similar to the DEA, the EDEA is able to make consistent predictions on a person’s choice after sufficiently large number of encounters with the same person. The EDEA also becomes more robust than a similarly expanded weighted majority algorithm as the number of encounters increase. The performance for DEA and EDEA have also been confirmed with both simulation results and experimental data collected through human robot interaction.