End-to-end deep learning for reverse driving trajectory of autonomous bulldozer

Abstract

A changeable and unstructured construction site presents challenges for the operating requirements of autonomous earthmoving machinery. We implement decision planning based on an end-to-end deep learning method, which fills the gap in the research related to the intelligent construction of autonomous bulldozers. Our proposed method can acquire relevant image features in both spatial attention and channel attention based on modified coordinate attention, and comparative analysis demonstrate advantages compared to traditional convolutional methods. We can obtain the output of turning angle and turning point by fusing multimodal data, including images and construction trajectories, and then calculate the reverse driving trajectory. The interpretability of the network is analyzed through visualization. Combined with the large-scale data of construction process collected from experienced operators, we extracted the data sets required for this research to train the model. Results show that our proposed method has anthropomorphic intelligence, which satisfies the decision-making and control process of experienced operators. It is effective in realizing an autonomous bulldozer in actual intelligence construction.