Iteratively refined feasibility checks in robotic assembly sequence planning

Due to shorter product life cycles and increasing customization, production lines must be able to quickly adapt to novel product variants. This requires the automatic generation of assembly sequence plans from product specifications, as manual engineering of plans is slow and labor intensive. The main challenge in assembly planning is that the search for a valid plan must take the capabilities of the robotic system that will execute the plan into account. But checking the feasibility of executing a plan requires a simulation of the system, which slows down the search for a valid and executable plan. We therefore propose and implement two ideas to reduce search times. First, we iteratively refine the feasibility check of an assembly plan from levels taking only parts into account (which is fast) to high-fidelity levels including motion planning and full robotic simulations (which is high fidelity, but slow). Slower levels are only checked if faster levels succeed. The second is that errors in these levels are propagated upwards as symbolic rules that prune the search tree. We demonstrate how our contributions reduce the need for high-fidelity simulations on a two-armed robotic system that assembles product variants out of aluminum profiles.