Slippage and Immobilization Detection for Planetary Exploration Rovers via Machine Learning and Proprioceptive Sensing

Abstract

This paper presents a new methodology where machine learning is used for detecting various levels of slip in the context of planetary exploration robotic missions. This methodology aims at employing proprioceptive rover sensor signals. Consequently, no operational complexity is added to the rover's commanding and it is independent of lighting conditions. Two supervised learning methods (Support Vector Machines and Artificial Neural Networks) are compared to two unsupervised learning approaches (K-means and Self-Organizing Maps (SOM)). Physical experiments using a single-wheel testbed equipped with an MSL spare wheel and a real planetary exploration rover validate the implemented methodology. Performance is evaluated in terms of well-known metrics both considering single data points and subsets of consecutive data points (moving median filter). Computation time and storage requirements are also examined. One of the SOM-based algorithms, semantic SOM method, demonstrates a proper balance between the benefits of supervised learning algorithms (high success rate, >96%) and the advantages of unsupervised learning methods (low storage requirements, 5 kb, and no need of manually-labeled training data). This paper also addresses the most convenient placement of IMU sensors on the rover chassis such that slippage detection is maximized.