This paper describes a novel approach to learn an implicit, differentiable distance function for arbitrary configurations of a robotic manipulator used for reactive control. By exploiting GPU processing, we efficiently query the learned collision representation and obtain an implicit distance between the robot and the environment. The differentiable nature of the learned function allows for calculating valid gradients wrt. any robot configuration, providing a repulsive vector field in joint space that can be injected in various control methods to improve collision avoidance. We present preliminary results on solving collision avoidance for a 7DoF robot with a reactive inverse kinematics solution, as well as improving performance of a sampling-based model-predictive controller.