Physics-inspired motion planning for information-theoretic target detection using multiple aerial robots

Abstract

This paper presents a motion-planning strategy for multiple, mobile sensor platforms using visual sensors with a finite field of view. Visual sensors are used to collect position measurements of potential targets within the search domain. Measurements are assimilated into a multi-target Bayesian likelihood ratio tracker that recursively produces a probability density function over the possible target positions. Vehicles are dynamically routed using a controller based on a concept from artificial physics, where vehicle motion depends on the target probability at their location as well as the distance to nearby agents. In this paradigm, the inverse log-likelihood ratio represents temperature, i.e., high likelihood corresponds to cold temperature and low likelihood corresponds to high temperature. Vehicles move at a temperature-dependent speed along the negative gradient of the temperature surface while interacting locally with other agents via a Lennard-Jones potential in order to emergently transition between the three states of matter--solid, liquid, and gas. We show that the gradient-following behavior corresponds to locally maximizing the mutual information between the measurements and the target state. The performance of the algorithm is experimentally demonstrated for visual measurements in a motion capture facility using quadrotor sensor platforms equipped with downward facing cameras.