Viewpoint-Aware Task Planning and Model Predictive Control for Applications in Videography and Multi-Target Tracking

Abstract

We seek to combine high level planning with low level reactive control to solve a variety of viewpoint-constrained target following tasks. In the scenarios we consider, a team of tracking agents is desired to gain some sort of visual information about one or more target agents. A high level planning algorithm accounts for coarse, global decisions, such as “Which targets should each tracker be responsible for?”, or “When should a tracker visit each target?” This level of planning is combinatorial in nature and requires coordination between the tracking agents. We combine this process with a lower-level reactive control accounts for stochastic target motion. By making this controller aware of a viewpoint cost function, the behavior of the tracking agents can be both more performant and easier to deploy on real robots.