We present a mobile robot motion planning ap-proach under kinodynamic constraints that exploits learnedperception priors in the form of continuous Gaussian mixturefields. Our Gaussian mixture fields are statistical multi-modalmotion models of discrete objects or continuous media in theenvironment that encode e.g. the dynamics of air or pedestrianflows. We approach this task using a recently proposed circularlinear flow field map based on semi-wrapped GMMs whosemixture components guide sampling and rewiring in an RRT*algorithm using a steer function for non-holonomic mobilerobots. In our experiments with three alternative baselines,we show that this combination allows the planner to veryefficiently generate high-quality solutions in terms of pathsmoothness, path length as well as natural yet minimum controleffort motions through multi-modal representations of Gaussianmixture fields.