Estimating Wheel Slip of a Planetary Exploration Rover via Unsupervised Machine Learning

Planetary exploration rovers often encounter imperfect traction and wheel slip, which negatively impacts navigation and in the worst case can result in permanent immobilization. Recent studies have applied machine learning to estimate rover wheel slip, which this paper extends via the implementation of three unsupervised learning algorithms: self-organizing maps, k-means clustering, and autoencoding. Unsupervised learning is preferred since labelled training data may be risky or time-consuming to obtain on site; each algorithm classifies the rover's current slip state into one of several discrete categories. Proprioceptive sensors are used to avoid added complexity and prevent a reliance on visual odometry. The algorithms are validated using sensor data from a planetary rover driving on a sandy incline, and performance is evaluated for different velocities, sensor inputs, slip classes, algorithm parameters, and data filters. Self-organizing maps (SOM) demonstrate the best slip classification accuracy, achieving 97% immobilization detection in the ideal two-class case. At rover-like speeds of 0.10 m/s, 88% accuracy is demonstrated for three classes. For ten slip classes, 71% accuracy is obtainable. Compared to SOM, k-means loses 5-30% accuracy and autoencoders lose 2-10% accuracy. SOM is most computationally intensive while k-means is least. An analysis of significant parameters for algorithm tuning displays accuracy benefits of up to 25 %, and mis-classifications can be further reduced by modifying class boundaries. The algorithms are generic and can be trained for different terrain, environment or vehicle parameters, and although some labelled data is needed to directly associate unsupervised clusters with slip classes, it is significantly less than what a fully-supervised algorithm requires. Unsupervised learning is thus considered promising for robust real-time rover slip estimation.