We must first define what we mean by improve. For the purposes of the work reported here, there are three ways to improve an interaction: (1) to increase the efficiency with which the robot performs the task in which the interaction happens; (2) causing the human’s subjective perceptions of the robot to become more positive; and (3) to increase the efficiency with which the human performs whatever task they are engaged in. This third aspect is one that is often neglected, but is one of the main reasons that we want people to interact with robots; to make the humans more efficient at what they do.
In this work, we consider a robot navigating in a hallway with a person as seen in figure 5.1. As a robot and a person approach each other in a hallway, it is unclear what the robot should do in order to most efficiently pass the person. When two people pass each other, they have a shared body of implicit knowledge about social situations, and swap a multitude of subtle
social cues in order to manage the interaction. Both of these are typically missing in the human-robot setting.
If the robot used its standard navigation algorithm, the results are socially suboptimal. The robot is programmed to take the most efficient path, which often results in a path that drives down the center of the hallway until a collision with the person seems imminent. This behavior treats the person as it would any other obstacle, ignoring the context of the obstacle as a moving, decision-making entity, that will not only move in a particular way, but also react to the robot’s movements. In essence, the robot has no effective theory of mind for the people it encounters. This is problematic from the perspective of the person because they will have no way to predict which side of the hallway the robot will pass on, leading to higher uncertainty and less effective task behavior.
The use of navigation behavior that is not contextually aware has been shown to be quite problematic. In a long-term ethnographic study by Mutlu and Forlizzi [70], an autonomous delivery robot in certain hospital environments was found to not only be ineffective in its tasks (bumping into people and obstacles) but also made the people around it feel “disrespected” and anxious. The study found that these problems persisted after up to three years of interaction with the robot, indicating that the disconnect with how they expect it to act is a lasting problem.
The path that the robot takes is only one portion of the problem. Even if the robot did move to one side of the hallway or another, the person has no way to know why the robot performed such an action. It could have been because the person was there, or it could have been something else. Without implicit social cues, it is hard for the person to tell the difference.
Our hypothesis is that if the robot modifies its behavior to use a predictive model for the person, in which the person’s social behavior is taken into account, the person will recognize the social behavior and be able to create a theory-of-mind for the robot that is more similar to people. We can examine how long it takes for the robot and the human to reach their goals, as well as how the person’s opinion of the robot differs as a result of the interaction. As a result or our additional social navigation, the person will be able to complete their task more efficiently, since they are then able to use all of their (often implicit) prior knowledge of social interactions to predict how the robot will behave.
Figure 5.1: An Example of Standard vs. Social Navigation - These diagrams show results from an experiment as a robot (blue square) and a human participant (yellow circle) pass each other in a hallway. In the top example, using the standard navigation, the person is forced to slow down drastically while the robot passes (c). With the social navigation in the bottom example, the person is able to pass the robot passes the robot with much greater ease and ends up completing the task more quickly as a result.
In this work, we address these two problems, inappropriate paths and poor signaling of intent, with two techniques. Similarly to the work discussed in section 4.3, we modify the robot’s costmaps to reflect the social behaviors we want to show in the planned paths. To better communicate the robot’s intent to the human, we use a gaze behavior that directs the robot’s head at either the human or at the hallway ahead. Our goal in implementing these behaviors is to make the interaction more natural, and hence efficient, for the human. In a pilot study with two human actors in a motion-capture environment [59], we explicitly studied the scenario of two people passing each other in a hallway. We observed that both gaze and the relative position of each person played a role in the passing behavior.