Human-Centered Design of Robot Explanations

Abstract

As robots perform tasks in human-occupied environments and especially when those tasks are collaborative, people are increasingly interested in understanding the robots’ behaviors. One approach to improving understandability is to enable robots to directly explain their behaviors, either proactively, or in response to a query. Give the many possible ways that explanations could be generated, our goal is to understand how people explain actions to other people. We achieve this through a crowdsourcing approach that captures human explanations and then ranks those explanations using measures of clarity and generalizability. We then use those findings to propose human-centered design principles for robots’ explanations to follow the human patterns that were ranked highest. Our first set of studies focus on the natural language humans use when referring to blocks on a tabletop. We draw parallels to findings from psychology literature and present statistics about what composes a clear natural language reference. Our second set of studies focus on finding characteristics of robot trajectory navigation demonstrations that convey the most information about a robot’s underlying objective function. We develop a theory of critical points along a trajectory and summarize how including these points in demonstrations affects human understanding of a robot’s behaviors. Given our human-centered principles for explanation, we propose both perception and natural language algorithms to allow real robots to generate these explanations automatically.