Characterization of Machine Learning Algorithms for Slippage Estimation in Planetary Exploration Rovers

Abstract

This paper presents a comprehensive comparison of well-known machine learning algorithms for estimating discrete slip events associated with individual wheels in planetary exploration rovers. This analysis is performed with various tuning configurations for each algorithm (55 setups). This research also shows the key role that environment plays in the performance of the learning algorithms: rover speed (0.05–0.25 [m/s]), type of terrain (gravel vs. sand), and tire type (off-road tires vs. smooth tires). These contributions are validated by using a broad data set collected using a planetary rover equipped with proprioceptive sensing. This work not only identifies the best algorithm to be deployed for discrete slip estimation, but it also helps with the selection and the mounting position of the sensing systems to be employed in future robotic planetary missions.