We study the problem of optimally coordinating multiple fixed-wing UAVs to perform vision-based target tracking, which entails that the UAVs are tasked with gathering the best joint vision-based measurements of an unpredictable ground target. We utilize an analytic expression for the error covariance associated with the fused measurements of the target’s position, and we employ stochastic fourth-order models for all vehicles, thereby incorporating a high degree of realism into the problem formulation. While dynamic programming can generate an optimal control policy that minimizes the expected value of the fused geolocation error covariance over time, it is accompanied by significant computational challenges due to the curse of dimensionality. In order to circumvent this challenge, we present a novel policy generation technique that combines simulation-based policy iteration with a robust regression scheme. The resulting control policy offers a significant advantage over alternative approaches and shows that the optimal control strategy involves coordinating the UAVs’ distances to the target rather than their viewing angles, which had been a common practice in target tracking.