Imitation-based control of automated ore excavator: improvement of autonomous excavation database quality using clustering and association analysis processes

Abstract

To perform productive autonomous excavation of a fragmented rock pile, it is necessary to recognize the condition of the fragmented rock pile and to plan appropriate excavation motions depending on the fragmented rock pile condition. In our previous work, we have proposed an imitation-based motion planning method and developed a recognizer of the rock pile condition and a motion planner. Experimental results using a 1/10-scale excavation model have demonstrated the fundamental feasibility. They have also revealed that both the number and diversity of learning data must be considered to achieve high-productive excavation. In this paper, to use learning data more effectively and to diversify the autonomous excavation database, clustering and association analysis processes are applied to the learning data. We propose a method to improve the database quality using results of those processes. The method merges two bipolar compositions; one follows natural conditions and the other is modified intentionally based on the occurrence and usage frequency. The experiment verifies the effectiveness of the proposed method and clarifies that the database contents can be adjusted based on the occurrence frequency of rock pile conditions and usage frequency of the excavation motions. ARTICLE HISTORY

Deep Reinforcement Learning With Adversarial Training for Automated Excavation Using Depth Images

“…Automation of excavation has attracted significant attention because of its expected social impact [1]- [7]. Early studies on autonomous excavation, such as [2], [5], and [6], have focused on modeling soil behavior to design an efficient scooping motion.…”

“…Moreover, it is challenging to design an optimal strategy to achieve efficient excavation. A recent study by Fukui et al employed an approach based on imitation learning to automate excavation [7]. Imitation learning is an approach that obtains the optimal strategy by learning from human demonstrations [18], [19].…”

“…Imitation learning is an approach that obtains the optimal strategy by learning from human demonstrations [18], [19]. Fukui et al proposed classifying the excavation motions demonstrated by human experts and adapting them to achieve efficient excavation [7]. However, their method requires a database of excavation motions, and it is difficult to build such a database in practice.…”

“…Construction often involves tasks that pose danger to human operators, and the construction industry is recently facing a shortage of the workers. To address such issues, automation of excavation has been investigated for decades [1]- [7]. Recently, machine learning and robotics have demonstrated promising results in various applications [8], [9].…”

Development of Integrative Methodologies for Effective Excavation Progress Monitoring

“…Using the joint angle data obtained from this configuration and kinematic information, the dynamic status of an excavator was estimated, which can be further utilized for the estimation of a 3D profile of ground surface [ 11 , 12 ]. The stereo vision (depth camera)-based methodology has been utilized to identify the ground surface including rock piles [ 13 ] through stereo matching in the left and right images that allows obtaining the depth images of the surface. Although the stereo matching technique provides high-resolution depth information, its measurement accuracy is affected by the richness in the surface texture [ 14 , 15 ].…”