The use of multiple cooperating robotic manipulators to assemble large aircraft structures entails the scheduling of many discrete tasks such as drilling holes and installing fasteners. Since the tasks have different tool requirements, it is desirable to minimize tool changes that incur significant time costs. We approach this problem as a multi-robot task allocation problem with precedence constraints, where the constraints are loosely enforced in terms of prioritizing the tasks to avoid unnecessary tool changes. To avoid the computational burden of searching over all possible task prioritization options, our main contribution is to develop a two-step, data-driven approach to automatically select suitable precedence relations. The first step is to adapt an iterative auction-based method to encode the precedence relations using scheduling heuristics. The second step is to develop a robust machine learning method to generate policies for automatically selecting efficient scheduling heuristics based on the problem characteristics. Experimental results show that the top performing heuristics yield schedules that are more efficient than those of a baseline partition-based scheduler by almost 17%–19%, depending on the robot failure profiles. The learned policies are also able to select heuristics that perform better than greedy selection without incurring additional computational costs.

使用多个协作机器人操纵器组装大型飞机结构需要安排许多离散任务，例如钻孔和安装紧固件。由于任务具有不同的工具要求，因此希望最大程度地减少需要大量时间成本的工具更换。我们将此问题作为具有优先权约束的多机器人任务分配问题进行处理，其中根据任务​​的优先级来避免不必要的工具更改，松散地实施了这些约束。为了避免搜索所有可能的任务优先级选项所带来的计算负担，我们的主要贡献是开发了一种两步式，数据驱动的方法来自动选择合适的优先级关系。第一步是调整基于迭代拍卖的方法，以使用调度启发法对优先级关系进行编码。第二步是开发一种强大的机器学习方法，以生成用于根据问题特征自动选择有效的调度启发式算法的策略。实验结果表明，根据机器人故障情况的不同，性能最高的启发式计划的效率要比基于基线分区的计划程序的效率高出近17％–19％。所学习的策略还能够选择性能比贪婪选择更好的启发式算法，而不会产生额外的计算成本。实验结果表明，根据机器人故障情况的不同，性能最高的启发式计划的效率要比基于基线分区的计划程序的效率高出近17％–19％。所学习的策略还能够选择性能比贪婪选择更好的启发式算法，而不会产生额外的计算成本。实验结果表明，根据机器人故障情况的不同，性能最高的启发式计划的效率要比基于基线分区的计划程序的效率高出近17％–19％。所学习的策略还能够选择性能比贪婪选择更好的启发式算法，而不会产生额外的计算成本。