We propose a simple robot learning recipe leveraging privileged information to train active perception policies on real robots.