Collaborative execution of exploration and tracking using move value estimation for robot teams (mvert)

• Authors: Tucker Balch;Ashley W. Stroupe
• Venue: Collaborative execution of exploration and tracking using move value estimation for robot teams (mvert)
• Year: 2003

This work presents Move Value Estimation for Robot Teams (MVERT), an architecture specifically designed for selecting low-level actions for multi-agent teams. The design goal for MVERT is to produce reasonable performance that takes advantage of a heterogeneous team while maintaining computational efficiency. MVERT is fully distributed—each agent selects actions based on its knowledge and knowledge provided it by teammates. Each robot approximates the expected next-step teammate contributions and, given these predictions, each robot can select its action to maximize the team's progress. MVERT represents progress with mathematical value functions that map state and robot task performance models to a numerical value representing mission utility. Many action selection approaches (optimal trajectory planning, for example) in large state-spaces may be computationally prohibitive, particularly for online mission replanning. However, taking advantage of a team's multi-agent nature to provide efficiency requires consideration of teammate contributions. Thus, in selecting an action with MVERT, each robot approximates the next-step contributions of teammates by applying their sensing models, task capabilities, and current poses in the value functions. The robot then evaluates its candidate actions by applying the value functions and its own sensor models. The action resulting in the overall highest-valued pose is selected and executed. Performance in each task is described by an individual value function. State includes current locations of teammates and objects in the environment. Performance models include task capabilities and sensor models. Value functions may be any mathematical representation of task performance. To determine an actions' overall value, independent task values are combined by weighted average. Weighting each task's value allows prioritizing tasks in accordance with desired performance. As progress reduces potential for improving value on some tasks, the weights automatically shift focus to the other tasks. Weights can be dynamically adapted as mission needs change. MVERT has been applied in simulation and on physical robots for mapping, dynamic target tracking, and complex multi-task missions (planetary exploration). MVERT improves team mission performance time and completeness compared to individual action selection and greatly improves computation time compared to a one-step optimal. MVERT produces contextually appropriate actions for successfully performing complex multi-task, multi-robot missions.