Machine Learning Challenges in Mars Rover Traverse Science

The successful implementation of machine learning in autonomous rover traverse science requires addressing challenges that range from the analytical technical realm, to the fuzzy, philosophical domain of entrenched belief systems within scientists and mission managers. These challenges are many. They include helping scientists understand the benefits and risks of using machine learning onboard and guiding them to distill and translate science goals into clear tasks that can be accomplished by algorithms. The technical challenges include, among other things, developing robust verification and validation plans. The ultimate test of onboard machine learning acceptance is if it flies and is executed onboard and delivers successful results during the mission. This last hurdle can be overcome by developing an implementation plan that poses an acceptable risk vs. reward scenario for mission managers. We are working on rover traverse science applications that address each of these issues. In this paper we will describe what we are doing and how we approach these challenges.