Information Gain Is Not All You Need

Supplementary Material

This document explains and comments the supplementary video material for the paper ``Information Gain Is Not All You Need’’.

The supplementary material consists of this document and nine video animations of the three methods (nearest frontier “NF”, distance advantage “DA”, and information gain maximization “IG”) on the three environments presented in the paper: office, cave, and maze.

Each frame of each video corresponds to one planning step, with the local window of the exploration state depicted with unknown in gray, free in white, and occupied in blue. The office also has exterior windows in cyan, these are not used by any method. The score map is overlaid in color with a color chart on the right. The score map is faded outside of the frontier cells; i.e., the target cell is the maximum scoring frontier cell, or the most red unfaded cell. The actual wall segments of the environment are drawn on top. The past trajectory is indicated by a blue line, and the planned trajectory in green with circles indicating where sensor scans are obtained.

When there are no more frontier cells, the robot takes a long path to the nearest frontier in the global map; however, since no planning happens during this long path, it looks as if thet robot makes a large jump. This is not the case.

The videos show how nearest frontier, being blind to prior knowledge of the environment, sometimes makes mistakes when there are multiple equally near frontier, leaving dead-ends of the environments unexplored. Information gain maximization commits similar mistakes, but more often and for a different reason. Its preference for gain drives it to abandon closets or dead-ends when there is little gain left. In contrast, distance advantage makes sure to finish these hard-to-reach regions before moving on, preventing the need to backtrack later on to finish them.