Impact of Data Quantity and Composition on Bucket Filling Performance for Wheel Loaders

This paper investigates the impact of training data with different quantities and compositions on the performance and robustness of a Neural Network (NN) controller for the wheel loader bucket filling task. Collecting training data for machine learning methods with a real-world Heavy Duty Mobile Machine (HDMM) is expensive, and therefore knowing how to collect the data and in what quantities will significantly reduce the data collection effort. We collected 2000 bucket fillings of non-homogeneous material, more specifically, a blasted rock pile with a kernel size of 0-400 mm. No previous study has reported such a challenging material composition. The collected data was divided into 6 datasets with sizes of 10, 20, 50, 100, 500, and 2000 bucket fillings. We use the Dynamic Time Warp (DTW) distance, k-medoids clustering, and the silhouette score, to create diverse and dissimilar datasets. Furthermore, one additional dataset was created with 10 bucket fillings, which are as similar as possible, resulting in 7 datasets in total. The datasets were used to synthesize 7 controllers that were then evaluated with a set of experiments to compare their performance to one another and the human operator. The results showed that the controller trained on similar bucket fillings was not robust and had poor performance, as expected. The experiment also showed that all the controllers trained on diverse data were robust enough to load the blasted rock material. However, the loaded material weight was less than the human operator, where the best controller loaded 9% less material weight, but 11% faster than the human operator.