The key, Levine explained, is the combination of low-level learned control approaches for moment-by-moment navigation, and higher-level planning, which is akin to RL. As Levine described it ...