Title: Mutual Information Maps for Single and Multi-Target Ergodic Search

Abstract:

This thesis addresses use of multi-agent systems to perform autonomous search for moving targets. Target search has many applications, including search and rescue and surveillance, but most robotic systems used in these situations require human operators. Recent works have used ergodic search methods to automatically generate multi-agent search trajectories. In ergodic search, agents are controlled so the time spent in any area of the search space is proportional to the amount of information in that area, according to a given distribution of information called the “information map”. Though trajectory optimization for ergodic search has been well studied, relatively little research has gone into the generation of the information maps themselves.

This work proposes mutual–information-based maps for robots with binary sensors (i.e., that sense a 1 when a target is seen and a 0 otherwise). Mutual information measures the expected decrease in the uncertainty of targets’ locations as a result of a sensor observation. We derive efficient-to-compute equations for the mutual information for both the single- and multi-target cases. For the multi-target case, we first derive filtering equations to track an unknown number of targets based on the cardinalized probability hypothesis density (CPHD) filter. We demonstrate that, using ergodic search with our proposed maps, agents are able to hone in on the locations of multiple moving targets despite the limited capabilities of their binary sensors. The mutual–information-based maps result in significantly lower mean absolute distance (MAD) than previously used maps for ergodic search in simulated search scenarios. Furthermore, using these maps, ergodic search also outperforms standard coverage-based methods for search in our experiments.

Committee:

Howie Choset (Advisor)

David Wettergreen

Jaskaran Singh Grover

Zoom link: https://cmu.zoom.us/j/98927499824?pwd=QWU1VjVmNXM0RndGTnJrYlJFL01UZz09

Meeting ID: 989 2749 9824

Passcode: 560698