Enabling Verification of Language Instructions through Bidirectional Communication for Astrobee

In space, robots for exploration will be required to work alone and alongside humans to complete missions such as facility inspection, targeted exploration, and scientific sampling. Clarity of assumptions, requirements, and objectives are important for such human-robot teams because communication may be infrequent or intermittent and any action taken by the robot has some inherent risk. It is important that the robot resolves any uncertainty about the specification of their mission or assumptions about the environment model. Minimizing this uncertainty is critical for robust execution of the task and to minimize risk to the robot a<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$n$</tex>d/or members of the human-robot team. Robots must therefore be able to accept guidance from humans about the mission and the environment and be able to resolve uncertainty. Language interaction is a highly flexible mode of communication that provides a mechanism for reducing such uncertainty by accepting declared knowledge and asking questions to resolve ambiguities. Often, a human may assume that a robot does not know something about the environment, and therefore such information is provided in monologue in conjunction with an instruction. Alternatively, the human may incorrectly assume the robot knows something important for completing the task, which requires the robot to ask targeted questions to resolve this uncertainty. This paper introduces a method of correcting incomplete or incorrect world models for free-flying robots via natural language dialogue by using the output of a search-based motion planner. Questions are only posed by the robot when required to generate a satisfactory route to the goal with the current model of the environment. Responses to those questions confirm or deny the validity of such declarative knowledge statements. Further clarifications are requested when required to generate a route to the goal from the current state. Experiments utilizing the proposed method are demonstrated using a simulation of the International Space Station and the Astrobee robot i<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$n$</tex> the Granite Laboratory at NASA Ames Research Center.