Human-robot collaboration applications require safe and reactive planning. Euclidean distance fields (EDF) are a promising representation of such dynamic scenes due to their ability to reason about free space and the readily available distance to collision costs. A key challenge for the commonly used discrete EDF representations, however, is the need for differentiable distance fields to produce smooth collision costs and efficient updates of dynamic objects. In this paper, we propose to use a Gaussian Process (GP) distance field-based framework that enables both, differentiable distance fields and fast dynamic scene updates. Moreover, we combine this framework with the Riemannian Motion Policies as a local reactive planner to enable safe human-robot interactions. We design a collision avoidance policy that models the repulsive motion using the distance and gradient fields from our GP. We show our reactive planner in an experiment with a UR5e interacting safely and smoothly with a human.
