Obstacle avoidance for robotic excavators using a recurrent neural network

In this paper, we present a recurrent neural network to resolve the obstacle avoidance problem of excavators. The conventional pseudo-inverse formulation requires excessive computation time for on-line or real time application. To effectively accomplish following goals: excavation task execution, joint limit control, and obstacle avoidance at the same time, conventional Newton-iteration scheme was replaced by a recurrent neural network algorithm in this study. The recurrent neural network was implemented for better kinematics control of the excavator with obstacle avoidance capability. In automated excavation environments, potential dangers exist if a worker is within the workspace of the excavator. When an obstacle is detected by a sensor, accidents can be easily prevented by halting the excavation process using a simple fail-safe algorithm. However, it would be more desirable to handle the unforeseen obstacles intelligently on-line while continuing the excavation task instead of stopping. For excavators, an obstacle can be classified into two categories. The first category includes obstacles on the ground such as trees, workers, and buildings. The second category of obstacles includes underground obstructions such as tree roots, boulders and etc. This paper focuses on the first category of these obstacles and was written to meet the emphasis requirements of avoiding obstacles on the ground for the excavator.