High Score Hawkmoth video games reveal obstacle navigation policy


Meeting Abstract

73.5  Monday, Jan. 6 09:15  High Score! Hawkmoth video games reveal obstacle navigation policy MUNK, Y*; BRUNTON, B; HORVITZ, E; DANIEL, TL; University of Washington; University of Washington; Microsoft Research; University of Washington yomunk@uw.edu

Navigation within cluttered environments is a complex behavior, in which navigational policy may be influenced by numerous factors including visibility range, flight speed, and light level. We investigated obstacle navigation in the hawkmoth Manduca sexta using a closed-loop flight simulator in which subjects navigate within a three-dimensional virtual forest. Each experiment yields a set of trajectories through the virtual forest under varying conditions of flight speed, visibility range, and contrast between obstacles and background. Each trajectory potentially includes hundreds of navigational interactions with virtual obstacles, and we have collected 836 trials over 30 subjects. In analyzing these trajectories, we posit a scoring system that awards points based on how a moth controls its trajectory near obstacles. We then assess the likelihood that the proposed scoring system reflects the intent of the subject by comparing the score obtained by the real trajectory to scores obtained by a distribution of trajectories that are randomized via resampling, establishing a bootstrap statistical test. At low flight speed and moderate visibility range, moths were often attracted to both bright and dark obstacles, flying close to but not necessarily hitting them. At high speed and low visibility range, some moths exhibited avoidance behavior, tending to steer away from obstacles. We identify a policy selecting for different navigation control behaviors dependent on the availability of visual information in M. sexta that may represent general visual strategies in animals navigating within cluttered environments. More generally, we describe a methodology for coupling computer-generated virtual environments with parameter search to identify implicit, context-sensitive navigational policies.

the Society for
Integrative &
Comparative
Biology