Stochastic dynamic-programming applied to planning of robot grinding tasks

This paper proposes an intelligent manufacturing system that can make decisions about the process in light of the uncertain outcome of these decisions and attempts to minimize the expected economic penalty resulting from those decisions. It uses robot weld bead grinding as an example of a process with significant process variation. The need for multiple grinding passes, the poor predictability of those passes, the task requirements, and the process constraints conspire to make planning and controlling weld bead grinding a formidable problem. A three tier hierarchical control system is proposed to plan an optimal sequence of grinding passes, dynamically simulate each pass, execute the planned sequence of controlled grinding passes, and modify the pass sequence as grinding continues. The top tier, described in this paper, plans the grinding sequence for each weld bead, and is implemented using Stochastic Dynamic Programming, selecting the volumetric removal and feedspeed for each pass in order to optimize the satisfaction of the task requirements by the entire grinding sequence within the equipment, task, and process constraints. The resulting optimal policies have quite complex structures, showing foresight, anxiety, indifference, and aggressiveness, depending upon the situation.