A computational model of human trust in supervisory control of robotic swarms

摘要

Trust is an important factor in the interaction between humans and automation to mediate the reliance action of human operators. In this work, we study human factors in supervisory control of robotic swarms and develop a computational model of human trust on swarm systems with varied levels of autonomy (LOA). We extend the classic trust theory by adding an intermediate feedback loop to the trust model, which formulates the human trust evolution as a combination of both open-loop trust anticipation and closed-loop trust feedback. A Kalman filter model is implemented to apply the above structure. We conducted a human experiment to collect user data of supervisory control of robotic swarms. Participants were requested to direct the swarm in a simulated environment to finish a foraging task using control systems with varied LOA. We implement three LOAs: manual, mixed-initiative (MI), and fully autonomous LOA. In the manual and autonomous LOA, swarms are controlled by a human or a search algorithm exclusively, while in the MI LOA, the human operator and algorithm collaboratively control the swarm. We train a personalized model for each participant and evaluate the model performance on a separate data set. Evaluation results show that our Kalman model outperforms existing models including inverse reinforcement learning and dynamic Bayesian network methods. In summary, the proposed work is novel in the following aspects: 1) This Kalman estimator is the first to model the complete trust evolution process with both closed-loop feedback and open-loop trust anticipation. 2) The proposed model analyzes time-series data to reveal the influence of events that occur during the course of an interaction; namely, a user’s intervention and report of levels of trust. 3) The proposed model considers the operator’s cognitive time lag between perceiving and processing the system display. 4) The proposed model uses the Kalman filter structure to fuse information from different sources to estimate a human operator's mental states. 5) The proposed model provides a personalized model for each individual.