Socially cognizant robotic navigation: Design methods and deployment challenges

Mobile robotic systems are increasingly being deployed in environments shared by everyday people. They can already be found cleaning floors in homes and shopping centers, delivering parcels on college campuses and sidewalks, and serving drinks at restaurants around the world. While they are equipped with safety systems that prevent collisions with obstacles (including people) in their environments, the motions of these robots as they navigate in spaces can be disruptive, displeasing, and distressing for coexisting people. As robots become more common in social settings, it becomes more urgent to carefully consider the social impacts they on the people around them. Within the field of human-robot interaction (HRI), researchers have investigated many factors of robot design that affect human responses; however, this area of research is relatively young, and even the published findings have gained little traction in the systems that are being sold and deployed at scale. This dissertation focuses on the spatial interactions that robots have with people in real-world deployments through three connected research efforts. First, the impacts of robot motions on the affective states of coexisting people are characterized through a series of human-subject experiments. This necessary step ensures that novel designs of robot path planners are informed, constrained, and driven by empirical data about how people respond to robot motions. The second research focus is on the application of these experimental findings to robot path planners. Improvements to traditional optimization-based methods are developed and tested in simulation environments and on a real-world platform. The final element of this dissertation is a qualitative analysis of the mobile robotics industry based on interviews with experts currently working in that space. This analysis is critical for two reasons: understanding the current practices and standards in industry, and identifying barriers to adopting new techniques, such as the novel path planners developed in this dissertation. One of the central contributions of this research is to empirically characterize the ways that varied robot trajectories impact the affective states of coexisting people. This is an important first step in defining the requirements of socially cognizant robot path planners. A virtual reality-based study of human subjects demonstrates that robot path curvature does have an impact on human affective states. This finding motivates a second experiment that assesses the effects of robot trajectory on nearby people in a more realistic setting. The second experiment does not find a significant effect of path curvature on human affective state, but does find effects due to robot speed and variance from a nominal path. To leverage the findings from these human subject experiments, the results must be applied to robot path planners to achieve socially acceptable robot motions. Two approaches are investigated. First, a local planner, or controller, is augmented with a trajectory critic that penalizes high-curvature motions. Second, a global path planner is augmented with a novel social cost layer and implemented on a commercially available mobile robot. This implementation penalizes the robot for planning paths that intersect with a circular sector in front of moving people, with the radius and angle of this sector dependent on the velocity of the detected person. Both approaches are found to effectively reduce the sharpness of curvature of a robot's motion. In order for socially cognizant robotic path planners to be widely adopted by industry roboticists, it is critical to understand the current state of socially aware path planning in industry, as well as the barriers to deploying socially cognizant path planning systems on commercial robots. Interviews are conducted with sixteen mobile robotics industry professionals and thematically analyzed to better understand the current state of social navigation on commercial platforms and the challenges for continued improvement. One important outcome from this effort is detailed insight into the person detection methods used in commercially available mobile robots, which differ from robot to robot, and vary from low-level obstacle avoidance to sophisticated semantic representations of different classes of people. Additionally, this analysis reveals that three key feedback mechanisms are just as important to the development of their systems as the initial design process. Moreover, these mechanisms interact with each other in constructive, but sometimes challenging ways. The outcomes of these research efforts individually advance the state of the art in the areas of empirical research for HRI, socially constrained path planning, and bridging gaps between academic and industry research efforts. Collectively, this will lead to robotics research that is more informed by the social dimension of human interactions, and robotic products that are more deliberately designed to be acceptable to the people that encounter them. This dissertation lays the groundwork for continued research into human-robot interaction that builds on the interdisciplinary approaches that have proven to be fruitful.