Goal Recognition Using Deep Learning in a Planetary Exploration Rover Developed for a Contest

Abstract

We participated in the “A Rocket Launch for International Student Satellites (ARLISS)” competition in which original design planetary exploration rovers competed to reach close to the target using autonomous control. In this competition, the rovers of various teams approached the target position using the global positioning system (GPS). However, they could only approach to within a few meters of the target due to the GPS positioning error. Our rover recognized the red traffic cone, placed at the goal point, by its color and in the Tanegashima Rocket Contest 2018, the rover was controlled to the point where the distance to the goal was 0 m. However, image recognition of goal objects by their colors suffers from the problem of unstable recognition due to changes in ambient lighting, which occurs due to, for example, weather changes. We therefore attempted to resolve this problem by employing deep learning. However, a considerable amount of calculation time is taken by a general deep learning model to run on a small planetary exploration rover computer and thus cannot be applied as it is. Therefore, we proposed a deep learning model with a short calculation time and high recognition accuracy. Using the proposed method, a recognition rate of over 99% could be achieved in a few seconds. Furthermore, we won the contest by demonstrating the effectiveness of the rover using the proposed method and thus proved the effectiveness of this method.