Human aspects in collaboration of humans and robots, as common in warehousing, are considered increasingly important objectives in operations management. In this work, we let robots learn about human stress levels based on sensor data in collaborative order picking of robotic mobile fulfillment systems. To this end, we develop a multi-agent reinforcement (MARL) approach that considers human stress levels and recovery behavior next to traditional performance objectives in the reward function of robotic agents. We assume a human-oriented assignment problem in which the robotic agents assign orders and short breaks to human workers based on their stress/recovery states. We find that the proposed MARL policy reduces the human stress time by up 50% in comparison to the applied benchmark policies and maintains system efficiency at a comparable level. While the results may need to be confirmed in different settings considering different types of humans aspects and efficiency objectives, they also show a practicable pathway to control stress levels and recovery for related problems of human-robot collaboration, inside and outside of warehousing.