Crowdsourcing-based Virtual Commissioning of Dynamic Human-Robot Teams (CroViCo)

近年来对人机任务共享的研究主要集中在动态工作流上。各自的方法允许相互适应和持续协调行动，而不是遵循固定和僵化的共享计划。因此，它们根据工人偏好支持不同的产品组装序列，处理由于外部事件导致的错误或中断。因此，在寻求评估动态方法时，有必要全面考虑联合任务执行期间可能出现的各种工作流程。只有将基准测试应用到具体的应用场景中，才能对动态协作机器人系统的优点和局限性提供深刻的见解。在人机交互领域中常用的用户研究显然过于昂贵和耗时，因此，我们设想了一种新的基准测试方法，该方法从虚拟调试技术中获得灵感，并自动模拟动态执行过程。当试图生成这样的过程时，挑战在于协作处理过程中现实人类工作策略的可伸缩模拟。由于它们的不确定性，很难预测和建模。在这个研究项目中，我们解决了关于认知数字人类模型的这一科学差距：我们正在寻求调查在线众包可以收集到多少关于人类决策和行动的数据，以及如何通过工作策略的随机模型来推广各自的数据，这些数据可以应用于不同的基准任务，作为模拟人类的人工智能。从这一点来看，该项目的长期目标是建立一个虚拟调试框架，该框架具有一个界面，可以为许多动态人机团队方法提供自动、可重复的性能测量。

Recent research on human-robot task sharing has been putting emphasis on dynamic workflows. Respective methods allow for mutual adaptation and continuous coordination of actions instead of following a fixed and rigid shared plan. Hence, they support different product assembly sequences according to worker preferences, handling of errors or interruptions because of external events. It is therefore necessary to comprehensively consider various workflows that may emerge during joint task execution when seeking to assess dynamic approaches. Benchmarking can only then provide profound insights into the benefits and limitations of a dynamic cobot system when put into concrete application scenarios. User studies as commonly used in the field of human-robot interaction are clearly too costly and time-consuming for this purpose – we, consequently, envision a novel benchmarking approach that takes inspiration from virtual commissioning techniques and simulates dynamic execution processes automatically. The challenge when trying to generate such processes lies in the scalable simulation of realistic human working strategies in cooperative handling processes. These are hard to foresee and model due to their indeterminism. In this research project, we address this scientific gap regarding cognitive digital human models: We are seeking to investigate the extent to which data on human decisions and actions can be gathered by online crowdsourcing, and how respective data can be generalized by stochastic models of working strategies that can be applied to different benchmark tasks as an artificial intelligence of simulated humans. From this point, the long-term project goal is to establish a virtual commissioning framework with an interface that renders automatic, reproducible performance measurements for numerous dynamic human-robot teaming methods feasible.