The marketability and adoption of robotic systems in agriculture is largely limited by economic and technology barriers that prevent highly efficient autonomous operations at a cost that justifies the generally low commodity values. From the technology perspective, autonomous systems exhibit brittleness in uncontrolled, unforeseen, and unlearned situations, prevalent in complex agricultural environments, which leads to inefficient operations and production losses. In this setting, human workers can complement autonomous systems through cooperation to not only minimize the total production cost by reducing the burden on autonomy, but also improve productivity by overcoming the shortfalls of autonomous systems. To this end, the focus of this work is on developing a systematic approach for human-robot collaboration in robotic fruit harvesting. Specifically, the main contribution of this work is in the development of stochastic optimization models for human-robot collaboration using a risk-averse framework to identify optimal harvester servicing policies that minimize the risk of economic losses all while guaranteeing a desired level of financial return. The solution of the models provide optimal policies for a human collaborator to service various sub-systems of a robotic harvester to improve performance in light of any operational inefficiencies in the sub-systems. The developed risk-averse optimization solution is validated in a simulated grove environment using data for various citrus varieties. Economic gains through human collaboration are confirmed through increase in the value of production and the net profit. The simulation results also provide insights into the temporal variation in the required human collaboration during the length of harvesting operation, which can benefit agricultural operations management in effective labor allocation. Furthermore, the intrinsic self-diagnosing ability of the developed model identifies component-level inefficiencies to provide an optimal servicing policy, thereby a human collaborator is not required to maintain situational awareness of the harvesting operation. Finally, it is demonstrated that the models can be solved to optimality in a relatively short time ($⩽2s)$ using standard commercial optimization packages.

机器人系统在农业中的适销性和采用在很大程度上受到经济和技术壁垒的限制，这些壁垒阻碍了高效的自主运行，而其成本却足以证明通常较低的商品价值。从技术角度来看，自治系统在不受控制，无法预料和未经学习的情况下表现出脆性，这在复杂的农业环境中普遍存在，这会导致低效的运营和生产损失。在这种情况下，人类工人可以通过合作对自治系统进行补充，不仅可以通过减轻自治负担来最大程度地降低总生产成本，还可以通过克服自治系统的不足来提高生产率。为此，这项工作的重点是开发一种系统的方法来进行机器人果实收获中的人机协作。具体来说，这项工作的主要贡献在于开发了一种人机协作的随机优化模型，该模型使用规避风险的框架来确定最佳的收割机服务政策，从而将所有经济损失的风险降至最低，同时又保证了所需的财务回报水平。模型的解决方案为人类协作者提供了最佳策略，以服务于机器人收割机的各个子系统，从而根据子系统中的任何操作效率低下来提高性能。所开发的规避风险的优化解决方案在模拟树林环境中使用各种柑橘品种的数据进行了验证。通过人为合作产生的经济收益可通过增加产品价值和净利润来确认。模拟结果还提供了对在收获过程中所需的人工协作的时间变化的见解，这可以在有效的劳动力分配中使农业运营管理受益。此外，所开发模型的内在自我诊断能力可以识别组件级别的效率低下，从而提供最佳的维修策略，因此不需要人类合作者来保持对收获操作的态势感知。最后，证明了可以在相对较短的时间内将模型求解为最优（所开发模型的内在自我诊断能力可以识别组件级别的效率低下，从而提供最佳的维修策略，因此不需要人类合作者来保持对收获操作的态势感知。最后，证明了可以在相对较短的时间内将模型求解为最优（所开发模型的内在自我诊断能力可以识别组件级别的效率低下，从而提供最佳的维修策略，因此不需要人类合作者来保持对收获操作的态势感知。最后，证明了可以在相对较短的时间内将模型求解为最优（$⩽\mathrm{2个}s）$ 使用标准的商业优化包。