S&AS: INT: COLLAB: Autonomy as a Service

点击翻译按钮获取中文摘要

How can one deploy teams of autonomous robots over long periods of time in such a way that they can be recruited and tasked by operators to perform a wide variety of tasks? Examples of such tasks include the environmental monitoring tasks encountered in biological conservation applications or in precision agriculture. This project will address this issue by letting the autonomous robots be available to the user in an on-demand manner through a novel 'Autonomy as a service' framework. To realize this idea, new tools will be developed for (i) describing the tasks in a way that can be understood by the robots, (ii) ensuring that the robots stay safe while executing the tasks, and (iii) methods for the robots to learn and improve over time in combination with the ability to assess their performance. The broader impact from the project will include implications for environmental monitoring, outreach programs for increasing STEM participation, and an integration of the research findings into the curriculum at the three participating institutions (Georgia Tech, BU, and MIT). In detail, the three main research themes are: (i) From Specification to Execution: The users must be able to recruit and task the robots with new missions, which calls for formally correct ways of going from high-level specifications, formulated as Linear Temporal Logic formulae, to coordinated control programs for the robots to execute. (ii) Resilient Autonomy: When delivering a system that can be commanded to perform tasks over long periods of time, the first concern must be to preserve the integrity of the system itself, i.e., basic functionality must be ensured even as the robot team is recruited to perform a particular set of tasks. This project will achieve this through the use of composable barrier certificates that ensure the forward invariance of the safe set, i.e., if the robots start safe, they will stay safe. (iii) Trajectory Based Learning from Massive Data Sets: The agent team must be able to assess the performance of whatever it is that they are monitoring. In this project, this will be achieved through models that can be effectively learned from massive data sets through novel tools for data compression and representation.