Value-based action selection for observation with robot teams using probabilistic techniques

We present an approach for directing next-step movements of robot teams engaged in mapping objects in their environment: Move Value Estimation for Robot Teams (MVERT). Resulting robot paths tend to optimize vantage points for all robots on the team by maximizing information gain. At each step, each robot selects a movement to maximize the utility (in this case, reduction in uncertainty) of its next observation. Trajectories are not guaranteed to be optimal, but team behavior serves to maximize the team's knowledge since each robot considers the observational contributions of team mates. MVERT is evaluated in simulation by measuring the resulting uncertainty about target locations compared to that obtained by robots acting without regard to team mate locations and to that of global optimization over all robots for each single step. Additionally, MVERT is demonstrated on physical teams of robots. The qualitative behavior of the team is appropriate and close to the single-step optimal set of trajectories.