Cognitive Mapping and Planning for Visual Navigation Supplementary Material

To demonstrate that our proposed mapper architecture works we test it in isolation on the task of free space prediction. We consider the scenario of an agent rotating about its current position, and the task is to predict free space in a 3.20 meter neighborhood of the agent. We only provide supervision for this experiment at end of the agents rotation. Figure A1 illustrates what the mapper learns. Observe that our mapper is able to make predictions where no observations are made. We also report the mean average precision for various versions of the mapper Table A1 on the test set (consisting of 2000 locations from the testing environment). We compare against an analytic mapping baseline which projects points observed in the depth image into the top view (by back projecting them into space and rotating them into the top-down view).