Batteries, camera, action! Learning a semantic control space for expressive robot cinematography