Robot-Proxy Grounding

Abstract

Exploration robotics has traditionally utilized an encoder-decoder model of communication between users and a robot. This means that users construct a plan (sequence of actions) to be sent to the robot; the robot executes the plan and returns data to the users, who then construct another plan. The problem with this interaction paradigm is twofold: (1) users must develop a complex mental model of the robotic system in order to create intricate plans, yet the data returned to them is not necessarily sufficient to help them develop such a model, and (2) the robot does not have the users’ specialized domain knowledge, so the robot does not have any way to ensure that how it handles unexpected events in the field is consistent with the users’ goals (what the users were trying to accomplish through the plan). In order to address these problems, this thesis introduces Robot-Proxy Grounding, a novel interaction model for exploration robotics. Robot-Proxy Grounding is derived from common ground theory, a model of human-human communication introduced and experimentally validated by Herbert Clark and his colleagues. Robot-Proxy Grounding is also based on detailed observations and analysis of the Life in the Atacama exploration robotics project, which indicated that a majority of the errors and miscommunications which occurred during the project resulted from a lack of common ground between participants even as the robot became more autonomous. Because the cost of communication with the remote robot is extremely high, this work introduces the concept of a “robot proxy,” a software system which models both the robot’s capabilities and the user’s goals. Robot-Proxy Grounding occurs as the proxy interacts with the user in real-time in place of the robot so as to promote common ground between the two. A proof of concept study was conducted which compared the effects of an encoder-decoder planning system and a prototype robot proxy; the study suggested that the use of a robot proxy was effective in improving task efficiency and fostering feelings of collaboration. A full implementation of a robot proxy-based planning system was constructed and evaluated. A user study demonstrated that participants who used the robot proxy were more efficient at the task, collected higher-quality data, and possessed more accurate information about the robot’s internal state and its context than participants without a robot proxy. The results suggest that the implementation was successful at promoting common ground with the user, resulting in improved task performance.