Control and Learning for Robotic Excavation

Abstract

Despite a large body of research and engineering work in robotic excavation, spanning decades, commercially viable systems for fully autonomous excavation are still in their infancy. The challenge is that performance is strongly influenced by the conditions of interaction between the excavator and the material, which are typically unknown and changing throughout an excavation process. For example, material such as fragmented rock, is not homogeneous in size and composition, can be wet and sticky, or dry, depending on local factors, and it inherently moves and exposes new hidden material that is not visible prior to executing the excavation process. Thus, a fully autonomous solution for robotic excavation requires adaptation to these changing conditions in order to achieve consistent bucket filling performance.

The robotic excavation research presented in this thesis focuses on the development and study of control and learning approaches for autonomous loading of fragmented rock using robotic load-haul-dump (LHD) machines, which are utilized in underground mining; although, the developed approaches are applicable to surface loaders as well. Using an admittance-based dig control strategy for autonomous loading---which uses force feedback to regulate motion commands to a loader's bucket actuator for autonomous digging and loading---as a base control framework, preliminary field experiments are conducted with a 14-tonne capacity robotic LHD machine to gain insights into the science of the autonomous loading process. These insights are used to improve the overall dig controller design, which increases the controller's robustness and facilitates consistent bucket filling performance. This robust and consistent base control framework enables the use of simple learning algorithms to adapt the controller to changing material characteristics. An iterative learning algorithm is developed and validated, which adapts control parameters to track a desired bucket fill weight at each excavation pass. A material classification methodology is also developed and validated, which uses information in the controller's force feedback signal to classify excavation materials to further improve learning and adaptation. Compared to AI-based approaches that require many training samples and advanced computing, the developed control and learning approaches have practical significance as demonstrated through the full-scale field experiments.