Deep Reinforcement Learning Methods for Navigational Aids

摘要

Navigation is one of the most complex daily activities we engage in. Partly due to its complexity, navigational abilities are vulnerable to many conditions including Topographical Agnosia, Alzheimer's Disease, and vision impairments. While navigation using solely vision remains a difficult problem in the field of assistive technology, emerging methods in Deep Reinforcement Learning and Computer Vision show promise in producing vision-based navigational aids for those with navigation impairments. To this effect, we introduce GraphMem, a Neural Computing approach to navigation tasks and compare it to several state of the art Neural Computing methods in a one-shot, 3D, first-person maze solving task. Comparing GraphMem to current methods in navigation tasks unveils insights into navigation and represents a first step towards employing these emerging techniques in navigational assistive technology.