Semantic Visual Navigation by Watching Y ouTube Videos - Supplementary Materials

Our hierarchical policy is motivated by Chaplot et al. [2], and consists Current location is indicated by the hollow circle. High-level policy outputs the most promising direction to pursue as the short-term goal. This process is repeated, i.e., the high-level policy takes feedback from the low-level We describe these two policies in more detail below. Low-level policy continues to re-plan when the occupancy map updates. We assume access to depth images, and adapt code from the map and plan implementation from [4], to implement the low-level policy.