Disclosed are techniques for radar-aided single-image three-dimensional (3D) depth reconstruction. In an aspect, at least one processor of an on-board computer of an ego vehicle receives, from a radar sensor of the ego vehicle, at least one radar image of an environment of the ego vehicle, receives, from a camera sensor of the ego vehicle, at least one camera image of the environment of the ego vehicle, and generates, using a convolutional neural network (CNN), a depth image of the environment of the ego vehicle based on the at least one radar image and the at least one camera image.