Mutual Information Maps for Single and Multi-Target Ergodic Search

Abstract

Though robots have become more prevalent in search and rescue operations, they usually require human operators to direct their search. Automating the search process can allow larger teams of robots to be used in a wider variety of situations, such as when communication is limited. Furthermore, automated robots have the potential to outperform teleoperated ones, especially in cases where it is difficult for a human to interpret incoming sensor data in real time.

Recent works have used ergodic search methods to automatically generate search trajectories. Ergodic search scales to a larger number of agents compared to typical information-based algorithms while still allowing prior knowledge to be incorporated into the search procedure. The prior knowledge, whether about the locations of survivors or sensing capabilities of the searching agents, must be encoded in the form of an information map which specifies which areas agents should spend more time in when looking for survivors. In this work, we focus on the generation of mutual information-based maps for robots with binary detectors (i.e., that sense a 1 when a survivor is seen and a 0 otherwise), and demonstrate that, even with such limited sensing capabilities, the robots are able to hone in on the locations of multiple moving targets. We show that these information maps results in significantly lower mean absolute distance (MAD) than previously used maps through simulated search scenarios. Furthermore, we show that, using these maps, ergodic search can also outperform standard coverage-based methods for search.