Deep Learning Perception for Quarry-Bot: A Lunar Dragline Excavator

This work investigates the use of Quarry-Bot for automated excavation tasks relevant to lunar site preparation. Motivated by the NASA Artemis program's objective of deploying robotic systems ahead of human arrival, this platform leverages the high payload-to-mass ratio and reconfigurable nature of cable-driven systems to perform repetitive excavation operations autonomously. A physical prototype was implemented on a pergola-based sandbox test rig, configured as a four cablerobot equipped with load cells for tension sensing, and a RealSense D435 for workspace mapping. Control strategies such as Lyapunov-PD tracking with NNLS-based tension allocation and a feedback controller are implemented to maintain cable tensions above minimum allowable thresholds. Moreover, advancements in deep-learning computer vision enable robust and adaptive perception capabilities, given the low-lighting and high contrast in Lunar environments. Given its aptitude, YOLOv11, a popular object detection framework, was implemented and tested. Experimental trials validated the ability of QuarryBot to perform sustained excavation sequences by avoiding obstacles for soil removal. These findings establish the avoidance approach for autonomous digging tasks with deep-learning computer vision in inadequate lighting conditions in tandem with LiDAR. Future work will extend the system to mobile bases for repositioning and elevating anchors for improved excavation on uneven terrain.